aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp2098
1 files changed, 2098 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp
new file mode 100644
index 0000000000..883bd5afdd
--- /dev/null
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qa_mmla_4x16/generic.cpp
@@ -0,0 +1,2098 @@
+/*
+ * Copyright (c) 2021, 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifdef __aarch64__
+
+#include "arm_gemm.hpp"
+#include "../../utils.hpp"
+
+#include <cassert>
+#include <cstdint>
+
+namespace arm_gemm {
+
+void a64_hybrid_s8qa_mmla_4x16 (
+ unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<int8_t> A_arg,
+ size_t M, size_t N, const int8_t *B_ptr, IndirectOutputArg<int8_t> output_arg,
+ const Requantize32 *qp, const int32_t *col_bias, unsigned int
+)
+{
+ struct KernelArgs {
+ unsigned int num_strings = {};
+ const unsigned int *string_lengths = {};
+ size_t N = {};
+ const int8_t *B_ptr = {};
+ size_t output_offset = {};
+ size_t input_initial_col = {};
+ size_t input_offset = {};
+ } ka;
+
+ unsigned long flags=0;
+ void *output_ptr;
+ void *input_ptr;
+
+ if (output_arg.is_indirect) {
+ output_ptr=(void *)(output_arg.indirect.ptr);
+ ka.output_offset=output_arg.indirect.offset;
+ flags |= 0x4;
+ } else {
+ output_ptr=(void *)(output_arg.direct.base);
+ ka.output_offset=output_arg.direct.stride;
+ }
+
+ if (A_arg.is_indirect) {
+ input_ptr=(void *)(A_arg.indirect.ptr);
+ ka.input_offset=A_arg.indirect.start_row;
+ ka.input_initial_col=A_arg.indirect.start_col;
+ flags |= 0x8;
+ } else {
+ assert(num_strings==1);
+ input_ptr=(void *)(A_arg.direct.base);
+ ka.input_offset=A_arg.direct.stride;
+ }
+ ka.num_strings = num_strings;
+ ka.string_lengths = string_lengths;
+ ka.N = N;
+ ka.B_ptr = B_ptr;
+ if (qp->c_offset > qp->minval) {
+ flags |= 0x20;
+ }
+ __asm__ __volatile__(
+ "1:" // Row loop
+ "cmp %x[M], #0x4\n"
+ "bge 97f\n"
+ "cmp %x[M], #0x2\n"
+ "bgt 65f\n"
+ "beq 33f\n"
+ "mov x10, %x[col_bias]\n"
+ "movi v11.4s, #0x0\n"
+ "movi v15.16b, #0x1\n"
+ "bic %x[flags], %x[flags], #0x80000000\n"
+ "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
+ "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+ "mov x27, %x[output_ptr]\n"
+ "2:" // Height 1: Column loop
+ "movi v16.4s, #0x0\n"
+ "movi v17.4s, #0x0\n"
+ "movi v18.4s, #0x0\n"
+ "movi v19.4s, #0x0\n"
+ "movi v20.4s, #0x0\n"
+ "movi v21.4s, #0x0\n"
+ "movi v22.4s, #0x0\n"
+ "movi v23.4s, #0x0\n"
+ "3:" // Height 1: setup done
+ "mov x26, #0x0\n"
+ "4:" // Height 1: String loop
+ "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+ "ldr w25, [x20, x26, LSL #0x2]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 5f\n"
+ "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x24, [x20, #0x0]\n"
+ "cbnz x26, 6f\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+ "add x24, x24, x20\n"
+ "b 6f\n"
+ "5:" // Height 1: setup direct input
+ "mov x24, %x[input_ptr]\n"
+ "6:" // Height 1: input setup done
+ "cmp x25, #0x10\n"
+ "blt 11f\n"
+ "ldr q1, [x24, #0x0]\n"
+ "ldr q5, [x28, #0x0]\n"
+ "cmp x25, #0x20\n"
+ "ldr q6, [x28, #0x10]\n"
+ "ldr q7, [x28, #0x20]\n"
+ "ldr q8, [x28, #0x30]\n"
+ "ldr q9, [x28, #0x40]\n"
+ "ldr q10, [x28, #0x50]\n"
+ "ldr q4, [x28, #0x60]\n"
+ "blt 9f\n"
+ "7:" // Height 1: Multiply loop: Main loop head
+ "trn1 v0.2d, v1.2d, v27.2d\n"
+ ".inst 0x4e85a410 // smmla v16.4s, v0.16b, v5.16b\n"
+ "ldr q25, [x28, #0x70]\n"
+ "trn2 v1.2d, v1.2d, v27.2d\n"
+ ".inst 0x4e86a414 // smmla v20.4s, v0.16b, v6.16b\n"
+ "ldr q24, [x28, #0x80]\n"
+ ".inst 0x4e87a411 // smmla v17.4s, v0.16b, v7.16b\n"
+ "ldr q30, [x28, #0x90]\n"
+ ".inst 0x4e88a415 // smmla v21.4s, v0.16b, v8.16b\n"
+ "ldr q29, [x28, #0xa0]\n"
+ ".inst 0x4e89a412 // smmla v18.4s, v0.16b, v9.16b\n"
+ "ldr q28, [x28, #0xb0]\n"
+ ".inst 0x4e8aa416 // smmla v22.4s, v0.16b, v10.16b\n"
+ "ldr q27, [x28, #0xc0]\n"
+ ".inst 0x4e84a413 // smmla v19.4s, v0.16b, v4.16b\n"
+ "ldr q26, [x28, #0xd0]\n"
+ ".inst 0x4e99a417 // smmla v23.4s, v0.16b, v25.16b\n"
+ "ldr q25, [x28, #0xe0]\n"
+ ".inst 0x4e98a430 // smmla v16.4s, v1.16b, v24.16b\n"
+ "ldr q24, [x28, #0xf0]\n"
+ ".inst 0x4e9ea434 // smmla v20.4s, v1.16b, v30.16b\n"
+ "add x24, x24, #0x10\n"
+ ".inst 0x4e9da431 // smmla v17.4s, v1.16b, v29.16b\n"
+ "add x28, x28, #0x100\n"
+ ".inst 0x4e9ca435 // smmla v21.4s, v1.16b, v28.16b\n"
+ ".inst 0x4e9ba432 // smmla v18.4s, v1.16b, v27.16b\n"
+ ".inst 0x4e9aa436 // smmla v22.4s, v1.16b, v26.16b\n"
+ ".inst 0x4e99a433 // smmla v19.4s, v1.16b, v25.16b\n"
+ ".inst 0x4e98a437 // smmla v23.4s, v1.16b, v24.16b\n"
+ "tbnz %x[flags], #31, 8f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f942b // sdot v11.4s, v1.16b, v15.16b\n"
+ "8:" // Height 1: Multiply loop: unique 1: skip row sum
+ "ldr q1, [x24, #0x0]\n"
+ "ldr q5, [x28, #0x0]\n"
+ "sub x25, x25, #0x10\n"
+ "cmp x25, #0x20\n"
+ "ldr q6, [x28, #0x10]\n"
+ "ldr q7, [x28, #0x20]\n"
+ "ldr q8, [x28, #0x30]\n"
+ "ldr q9, [x28, #0x40]\n"
+ "ldr q10, [x28, #0x50]\n"
+ "ldr q4, [x28, #0x60]\n"
+ "prfm pldl1keep, [x24, #0x80]\n"
+ "bge 7b\n"
+ "9:" // Height 1: Multiply loop: Single iteration only
+ "trn1 v0.2d, v1.2d, v24.2d\n"
+ ".inst 0x4e85a410 // smmla v16.4s, v0.16b, v5.16b\n"
+ "ldr q25, [x28, #0x70]\n"
+ "trn2 v1.2d, v1.2d, v24.2d\n"
+ ".inst 0x4e86a414 // smmla v20.4s, v0.16b, v6.16b\n"
+ "ldr q24, [x28, #0x80]\n"
+ ".inst 0x4e87a411 // smmla v17.4s, v0.16b, v7.16b\n"
+ "ldr q30, [x28, #0x90]\n"
+ ".inst 0x4e88a415 // smmla v21.4s, v0.16b, v8.16b\n"
+ "ldr q29, [x28, #0xa0]\n"
+ ".inst 0x4e89a412 // smmla v18.4s, v0.16b, v9.16b\n"
+ "ldr q28, [x28, #0xb0]\n"
+ ".inst 0x4e8aa416 // smmla v22.4s, v0.16b, v10.16b\n"
+ "ldr q27, [x28, #0xc0]\n"
+ ".inst 0x4e84a413 // smmla v19.4s, v0.16b, v4.16b\n"
+ "ldr q26, [x28, #0xd0]\n"
+ ".inst 0x4e99a417 // smmla v23.4s, v0.16b, v25.16b\n"
+ "ldr q25, [x28, #0xe0]\n"
+ ".inst 0x4e98a430 // smmla v16.4s, v1.16b, v24.16b\n"
+ "ldr q24, [x28, #0xf0]\n"
+ "sub x25, x25, #0x10\n"
+ ".inst 0x4e9ea434 // smmla v20.4s, v1.16b, v30.16b\n"
+ ".inst 0x4e9da431 // smmla v17.4s, v1.16b, v29.16b\n"
+ "add x24, x24, #0x10\n"
+ "add x28, x28, #0x100\n"
+ ".inst 0x4e9ca435 // smmla v21.4s, v1.16b, v28.16b\n"
+ ".inst 0x4e9ba432 // smmla v18.4s, v1.16b, v27.16b\n"
+ ".inst 0x4e9aa436 // smmla v22.4s, v1.16b, v26.16b\n"
+ ".inst 0x4e99a433 // smmla v19.4s, v1.16b, v25.16b\n"
+ ".inst 0x4e98a437 // smmla v23.4s, v1.16b, v24.16b\n"
+ "tbnz %x[flags], #31, 10f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f942b // sdot v11.4s, v1.16b, v15.16b\n"
+ "10:" // Height 1: Multiply loop: unique 2: skip row sum
+ "prfm pldl1keep, [x24, #0x80]\n"
+ "11:" // Height 1: Multiply loop: Main loop skip
+ "cbz x25, 20f\n"
+ "cmp x25, #0x8\n"
+ "blt 14f\n"
+ "12:" // Height 1: Multiply loop: Odd block loop
+ "ldr d25, [x24], #0x8\n"
+ "trn1 v0.2d, v25.2d, v24.2d\n"
+ "tbnz %x[flags], #31, 13f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ "13:" // Height 1: Multiply loop: unique 3: skip row sum
+ "ldr q24, [x28, #0x0]\n"
+ "ldr q26, [x28, #0x10]\n"
+ ".inst 0x4e98a410 // smmla v16.4s, v0.16b, v24.16b\n"
+ "sub x25, x25, #0x8\n"
+ "ldr q25, [x28, #0x20]\n"
+ "ldr q24, [x28, #0x30]\n"
+ "cmp x25, #0x8\n"
+ ".inst 0x4e9aa414 // smmla v20.4s, v0.16b, v26.16b\n"
+ "ldr q27, [x28, #0x40]\n"
+ "ldr q26, [x28, #0x50]\n"
+ ".inst 0x4e99a411 // smmla v17.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a415 // smmla v21.4s, v0.16b, v24.16b\n"
+ "ldr q25, [x28, #0x60]\n"
+ "ldr q24, [x28, #0x70]\n"
+ ".inst 0x4e9ba412 // smmla v18.4s, v0.16b, v27.16b\n"
+ ".inst 0x4e9aa416 // smmla v22.4s, v0.16b, v26.16b\n"
+ ".inst 0x4e99a413 // smmla v19.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a417 // smmla v23.4s, v0.16b, v24.16b\n"
+ "add x28, x28, #0x80\n"
+ "bge 12b\n"
+ "14:" // Height 1: Multiply loop: Skip odd blocks
+ "cbz x25, 20f\n"
+ "tbz x25, #2, 16f\n"
+ "ldr s1, [x24], #0x4\n"
+ "tbz x25, #1, 15f\n"
+ "ld1 { v1.h }[2], [x24], #0x2\n"
+ "tbz x25, #0, 18f\n"
+ "ld1 { v1.b }[6], [x24]\n"
+ "b 18f\n"
+ "15:" // Height 1: Multiply loop: Ragged operand read: partial_1_4
+ "tbz x25, #0, 18f\n"
+ "ld1 { v1.b }[4], [x24]\n"
+ "b 18f\n"
+ "16:" // Height 1: Multiply loop: Ragged operand read: partial_2_0
+ "tbz x25, #1, 17f\n"
+ "ldr h1, [x24], #0x2\n"
+ "tbz x25, #0, 18f\n"
+ "ld1 { v1.b }[2], [x24]\n"
+ "b 18f\n"
+ "17:" // Height 1: Multiply loop: Ragged operand read: partial_1_0
+ "ldr b1, [x24, #0x0]\n"
+ "18:" // Height 1: Multiply loop: Ragged operand read: Done
+ "trn1 v0.2d, v1.2d, v24.2d\n"
+ "tbnz %x[flags], #31, 19f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ "19:" // Height 1: Multiply loop: unique 4: skip row sum
+ "ldr q25, [x28, #0x0]\n"
+ "ldr q24, [x28, #0x10]\n"
+ ".inst 0x4e99a410 // smmla v16.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a414 // smmla v20.4s, v0.16b, v24.16b\n"
+ "ldr q25, [x28, #0x20]\n"
+ "ldr q24, [x28, #0x30]\n"
+ ".inst 0x4e99a411 // smmla v17.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a415 // smmla v21.4s, v0.16b, v24.16b\n"
+ "ldr q25, [x28, #0x40]\n"
+ "ldr q24, [x28, #0x50]\n"
+ ".inst 0x4e99a412 // smmla v18.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a416 // smmla v22.4s, v0.16b, v24.16b\n"
+ "ldr q25, [x28, #0x60]\n"
+ "ldr q24, [x28, #0x70]\n"
+ ".inst 0x4e99a413 // smmla v19.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a417 // smmla v23.4s, v0.16b, v24.16b\n"
+ "add x28, x28, #0x80\n"
+ "20:" // Height 1: Multiply loop: No odd multiplies
+ "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+ "add x26, x26, #0x1\n"
+ "cmp x26, x20\n"
+ "bne 4b\n"
+ "uzp1 v16.2d, v16.2d, v20.2d\n"
+ "uzp1 v17.2d, v17.2d, v21.2d\n"
+ "prfm pstl1keep, [x27, #0x0]\n"
+ "uzp1 v18.2d, v18.2d, v22.2d\n"
+ "uzp1 v19.2d, v19.2d, v23.2d\n"
+ "mov v23.16b, v16.16b\n"
+ "tbnz %x[flags], #31, 21f\n"
+ "add x20, %x[qp], %[b_offset]\n"
+ "ld1r { v16.4s }, [x20]\n"
+ "addp v11.4s, v11.4s, v11.4s\n"
+ "neg v16.4s, v16.4s\n"
+ "dup v11.4s, v11.s[0]\n"
+ "mul v11.4s, v11.4s, v16.4s\n"
+ "21:" // Height 1: skip row sum fixup
+ "ldr q24, [x10, #0x0]\n"
+ "ldr q22, [x10, #0x10]\n"
+ "add v23.4s, v23.4s, v11.4s\n"
+ "add v17.4s, v17.4s, v11.4s\n"
+ "ldr q21, [x10, #0x20]\n"
+ "ldr q20, [x10, #0x30]\n"
+ "add v18.4s, v18.4s, v11.4s\n"
+ "add v19.4s, v19.4s, v11.4s\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1r { v16.4s }, [x20]\n"
+ "orr %x[flags], %x[flags], #0x80000000\n"
+ "add v23.4s, v23.4s, v24.4s\n"
+ "add v17.4s, v17.4s, v22.4s\n"
+ "add v18.4s, v18.4s, v21.4s\n"
+ "add x20, %x[qp], %[per_layer_right_shift]\n"
+ "ld1r { v0.4s }, [x20]\n"
+ "add v19.4s, v19.4s, v20.4s\n"
+ "sqrdmulh v23.4s, v23.4s, v16.4s\n"
+ "add x10, x10, #0x40\n"
+ "sqrdmulh v17.4s, v17.4s, v16.4s\n"
+ "sqrdmulh v18.4s, v18.4s, v16.4s\n"
+ "sqrdmulh v19.4s, v19.4s, v16.4s\n"
+ "tbz %x[flags], #5, 22f\n"
+ "and v22.16b, v23.16b, v0.16b\n"
+ "and v21.16b, v17.16b, v0.16b\n"
+ "and v20.16b, v18.16b, v0.16b\n"
+ "and v16.16b, v19.16b, v0.16b\n"
+ "sshr v22.4s, v22.4s, #0x1f\n"
+ "sshr v21.4s, v21.4s, #0x1f\n"
+ "sshr v20.4s, v20.4s, #0x1f\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "sqadd v23.4s, v23.4s, v22.4s\n"
+ "sqadd v17.4s, v17.4s, v21.4s\n"
+ "sqadd v18.4s, v18.4s, v20.4s\n"
+ "sqadd v19.4s, v19.4s, v16.4s\n"
+ "22:" // Height 1: no shift correction
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1r { v21.4s }, [x20]\n"
+ "srshl v23.4s, v23.4s, v0.4s\n"
+ "srshl v17.4s, v17.4s, v0.4s\n"
+ "srshl v18.4s, v18.4s, v0.4s\n"
+ "srshl v19.4s, v19.4s, v0.4s\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1r { v20.4s }, [x20]\n"
+ "add v23.4s, v23.4s, v21.4s\n"
+ "add v17.4s, v17.4s, v21.4s\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1r { v16.4s }, [x20]\n"
+ "add v18.4s, v18.4s, v21.4s\n"
+ "add v19.4s, v19.4s, v21.4s\n"
+ "cmp x9, #0x10\n"
+ "smin v23.4s, v23.4s, v20.4s\n"
+ "smin v17.4s, v17.4s, v20.4s\n"
+ "smin v18.4s, v18.4s, v20.4s\n"
+ "smin v19.4s, v19.4s, v20.4s\n"
+ "smax v23.4s, v23.4s, v16.4s\n"
+ "smax v17.4s, v17.4s, v16.4s\n"
+ "smax v18.4s, v18.4s, v16.4s\n"
+ "smax v19.4s, v19.4s, v16.4s\n"
+ "uzp1 v23.8h, v23.8h, v17.8h\n"
+ "uzp1 v16.8h, v18.8h, v19.8h\n"
+ "uzp1 v23.16b, v23.16b, v16.16b\n"
+ "bge 31f\n"
+ "tbz x9, #3, 26f\n"
+ "str d23, [x27], #0x8\n"
+ "tbz x9, #2, 24f\n"
+ "st1 { v23.s }[2], [x27], #0x4\n"
+ "tbz x9, #1, 23f\n"
+ "st1 { v23.h }[6], [x27], #0x2\n"
+ "tbz x9, #0, 30f\n"
+ "st1 { v23.b }[14], [x27]\n"
+ "b 30f\n"
+ "23:" // Height 1: Partial direct writeback: partial_1_12
+ "tbz x9, #0, 30f\n"
+ "st1 { v23.b }[12], [x27]\n"
+ "b 30f\n"
+ "24:" // Height 1: Partial direct writeback: partial_2_8
+ "tbz x9, #1, 25f\n"
+ "st1 { v23.h }[4], [x27], #0x2\n"
+ "tbz x9, #0, 30f\n"
+ "st1 { v23.b }[10], [x27]\n"
+ "b 30f\n"
+ "25:" // Height 1: Partial direct writeback: partial_1_8
+ "tbz x9, #0, 30f\n"
+ "st1 { v23.b }[8], [x27]\n"
+ "b 30f\n"
+ "26:" // Height 1: Partial direct writeback: partial_4_0
+ "tbz x9, #2, 28f\n"
+ "str s23, [x27], #0x4\n"
+ "tbz x9, #1, 27f\n"
+ "st1 { v23.h }[2], [x27], #0x2\n"
+ "tbz x9, #0, 30f\n"
+ "st1 { v23.b }[6], [x27]\n"
+ "b 30f\n"
+ "27:" // Height 1: Partial direct writeback: partial_1_4
+ "tbz x9, #0, 30f\n"
+ "st1 { v23.b }[4], [x27]\n"
+ "b 30f\n"
+ "28:" // Height 1: Partial direct writeback: partial_2_0
+ "tbz x9, #1, 29f\n"
+ "str h23, [x27], #0x2\n"
+ "tbz x9, #0, 30f\n"
+ "st1 { v23.b }[2], [x27]\n"
+ "b 30f\n"
+ "29:" // Height 1: Partial direct writeback: partial_1_0
+ "str b23, [x27, #0x0]\n"
+ "30:" // Height 1: Partial direct writeback: Done
+ "b 32f\n"
+ "31:" // Height 1: Full writeback
+ "str q23, [x27, #0x0]\n"
+ "add x27, x27, #0x10\n"
+ "32:" // Height 1: Writeback done
+ "subs x9, x9, #0x10\n"
+ "bgt 2b\n"
+ "b 130f\n"
+ "33:" // Height 2
+ "mov x10, %x[col_bias]\n"
+ "movi v11.4s, #0x0\n"
+ "movi v12.4s, #0x0\n"
+ "bic %x[flags], %x[flags], #0x80000000\n"
+ "movi v15.16b, #0x1\n"
+ "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
+ "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+ "mov x27, %x[output_ptr]\n"
+ "34:" // Height 2: Column loop
+ "movi v16.4s, #0x0\n"
+ "movi v17.4s, #0x0\n"
+ "movi v18.4s, #0x0\n"
+ "movi v19.4s, #0x0\n"
+ "movi v20.4s, #0x0\n"
+ "movi v21.4s, #0x0\n"
+ "movi v22.4s, #0x0\n"
+ "movi v23.4s, #0x0\n"
+ "35:" // Height 2: setup done
+ "mov x26, #0x0\n"
+ "36:" // Height 2: String loop
+ "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+ "ldr w25, [x20, x26, LSL #0x2]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 37f\n"
+ "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x24, [x20, #0x0]\n"
+ "ldr x23, [x20, #0x8]\n"
+ "cbnz x26, 38f\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+ "add x24, x24, x20\n"
+ "add x23, x23, x20\n"
+ "b 38f\n"
+ "37:" // Height 2: setup direct input
+ "mov x24, %x[input_ptr]\n"
+ "add x23, x24, x21\n"
+ "38:" // Height 2: input setup done
+ "cmp x25, #0x10\n"
+ "blt 43f\n"
+ "ldr q1, [x24, #0x0]\n"
+ "ldr q2, [x23, #0x0]\n"
+ "cmp x25, #0x20\n"
+ "ldr q5, [x28, #0x0]\n"
+ "ldr q6, [x28, #0x10]\n"
+ "ldr q7, [x28, #0x20]\n"
+ "ldr q8, [x28, #0x30]\n"
+ "ldr q9, [x28, #0x40]\n"
+ "ldr q10, [x28, #0x50]\n"
+ "ldr q4, [x28, #0x60]\n"
+ "blt 41f\n"
+ "39:" // Height 2: Multiply loop: Main loop head
+ "trn1 v0.2d, v1.2d, v2.2d\n"
+ ".inst 0x4e85a410 // smmla v16.4s, v0.16b, v5.16b\n"
+ "ldr q25, [x28, #0x70]\n"
+ ".inst 0x4e86a414 // smmla v20.4s, v0.16b, v6.16b\n"
+ "ldr q24, [x28, #0x80]\n"
+ "trn2 v1.2d, v1.2d, v2.2d\n"
+ ".inst 0x4e87a411 // smmla v17.4s, v0.16b, v7.16b\n"
+ "ldr q30, [x28, #0x90]\n"
+ ".inst 0x4e88a415 // smmla v21.4s, v0.16b, v8.16b\n"
+ "ldr q29, [x28, #0xa0]\n"
+ ".inst 0x4e89a412 // smmla v18.4s, v0.16b, v9.16b\n"
+ "ldr q28, [x28, #0xb0]\n"
+ ".inst 0x4e8aa416 // smmla v22.4s, v0.16b, v10.16b\n"
+ "ldr q27, [x28, #0xc0]\n"
+ ".inst 0x4e84a413 // smmla v19.4s, v0.16b, v4.16b\n"
+ "ldr q26, [x28, #0xd0]\n"
+ ".inst 0x4e99a417 // smmla v23.4s, v0.16b, v25.16b\n"
+ "ldr q25, [x28, #0xe0]\n"
+ ".inst 0x4e98a430 // smmla v16.4s, v1.16b, v24.16b\n"
+ "ldr q24, [x28, #0xf0]\n"
+ ".inst 0x4e9ea434 // smmla v20.4s, v1.16b, v30.16b\n"
+ "add x24, x24, #0x10\n"
+ ".inst 0x4e9da431 // smmla v17.4s, v1.16b, v29.16b\n"
+ "add x23, x23, #0x10\n"
+ ".inst 0x4e9ca435 // smmla v21.4s, v1.16b, v28.16b\n"
+ "add x28, x28, #0x100\n"
+ ".inst 0x4e9ba432 // smmla v18.4s, v1.16b, v27.16b\n"
+ ".inst 0x4e9aa436 // smmla v22.4s, v1.16b, v26.16b\n"
+ ".inst 0x4e99a433 // smmla v19.4s, v1.16b, v25.16b\n"
+ ".inst 0x4e98a437 // smmla v23.4s, v1.16b, v24.16b\n"
+ "tbnz %x[flags], #31, 40f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f942b // sdot v11.4s, v1.16b, v15.16b\n"
+ "40:" // Height 2: Multiply loop: unique 5: skip row sum
+ "ldr q1, [x24, #0x0]\n"
+ "ldr q2, [x23, #0x0]\n"
+ "sub x25, x25, #0x10\n"
+ "cmp x25, #0x20\n"
+ "ldr q5, [x28, #0x0]\n"
+ "ldr q6, [x28, #0x10]\n"
+ "ldr q7, [x28, #0x20]\n"
+ "ldr q8, [x28, #0x30]\n"
+ "ldr q9, [x28, #0x40]\n"
+ "ldr q10, [x28, #0x50]\n"
+ "ldr q4, [x28, #0x60]\n"
+ "prfm pldl1keep, [x24, #0x80]\n"
+ "prfm pldl1keep, [x23, #0x80]\n"
+ "bge 39b\n"
+ "41:" // Height 2: Multiply loop: Single iteration only
+ "trn1 v0.2d, v1.2d, v2.2d\n"
+ ".inst 0x4e85a410 // smmla v16.4s, v0.16b, v5.16b\n"
+ "ldr q25, [x28, #0x70]\n"
+ ".inst 0x4e86a414 // smmla v20.4s, v0.16b, v6.16b\n"
+ "ldr q24, [x28, #0x80]\n"
+ "trn2 v1.2d, v1.2d, v2.2d\n"
+ ".inst 0x4e87a411 // smmla v17.4s, v0.16b, v7.16b\n"
+ "ldr q30, [x28, #0x90]\n"
+ ".inst 0x4e88a415 // smmla v21.4s, v0.16b, v8.16b\n"
+ "ldr q29, [x28, #0xa0]\n"
+ ".inst 0x4e89a412 // smmla v18.4s, v0.16b, v9.16b\n"
+ "ldr q28, [x28, #0xb0]\n"
+ ".inst 0x4e8aa416 // smmla v22.4s, v0.16b, v10.16b\n"
+ "ldr q27, [x28, #0xc0]\n"
+ ".inst 0x4e84a413 // smmla v19.4s, v0.16b, v4.16b\n"
+ "ldr q26, [x28, #0xd0]\n"
+ ".inst 0x4e99a417 // smmla v23.4s, v0.16b, v25.16b\n"
+ "ldr q25, [x28, #0xe0]\n"
+ ".inst 0x4e98a430 // smmla v16.4s, v1.16b, v24.16b\n"
+ "ldr q24, [x28, #0xf0]\n"
+ "sub x25, x25, #0x10\n"
+ ".inst 0x4e9ea434 // smmla v20.4s, v1.16b, v30.16b\n"
+ ".inst 0x4e9da431 // smmla v17.4s, v1.16b, v29.16b\n"
+ "add x24, x24, #0x10\n"
+ "add x23, x23, #0x10\n"
+ ".inst 0x4e9ca435 // smmla v21.4s, v1.16b, v28.16b\n"
+ ".inst 0x4e9ba432 // smmla v18.4s, v1.16b, v27.16b\n"
+ "add x28, x28, #0x100\n"
+ ".inst 0x4e9aa436 // smmla v22.4s, v1.16b, v26.16b\n"
+ ".inst 0x4e99a433 // smmla v19.4s, v1.16b, v25.16b\n"
+ ".inst 0x4e98a437 // smmla v23.4s, v1.16b, v24.16b\n"
+ "tbnz %x[flags], #31, 42f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f942b // sdot v11.4s, v1.16b, v15.16b\n"
+ "42:" // Height 2: Multiply loop: unique 6: skip row sum
+ "prfm pldl1keep, [x24, #0x80]\n"
+ "prfm pldl1keep, [x23, #0x80]\n"
+ "43:" // Height 2: Multiply loop: Main loop skip
+ "cbz x25, 52f\n"
+ "cmp x25, #0x8\n"
+ "blt 46f\n"
+ "44:" // Height 2: Multiply loop: Odd block loop
+ "ldr d25, [x24], #0x8\n"
+ "ldr d24, [x23], #0x8\n"
+ "trn1 v0.2d, v25.2d, v24.2d\n"
+ "tbnz %x[flags], #31, 45f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ "45:" // Height 2: Multiply loop: unique 7: skip row sum
+ "ldr q24, [x28, #0x0]\n"
+ "ldr q26, [x28, #0x10]\n"
+ ".inst 0x4e98a410 // smmla v16.4s, v0.16b, v24.16b\n"
+ "sub x25, x25, #0x8\n"
+ "ldr q25, [x28, #0x20]\n"
+ "ldr q24, [x28, #0x30]\n"
+ "cmp x25, #0x8\n"
+ ".inst 0x4e9aa414 // smmla v20.4s, v0.16b, v26.16b\n"
+ "ldr q27, [x28, #0x40]\n"
+ "ldr q26, [x28, #0x50]\n"
+ ".inst 0x4e99a411 // smmla v17.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a415 // smmla v21.4s, v0.16b, v24.16b\n"
+ "ldr q25, [x28, #0x60]\n"
+ "ldr q24, [x28, #0x70]\n"
+ ".inst 0x4e9ba412 // smmla v18.4s, v0.16b, v27.16b\n"
+ ".inst 0x4e9aa416 // smmla v22.4s, v0.16b, v26.16b\n"
+ ".inst 0x4e99a413 // smmla v19.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a417 // smmla v23.4s, v0.16b, v24.16b\n"
+ "add x28, x28, #0x80\n"
+ "bge 44b\n"
+ "46:" // Height 2: Multiply loop: Skip odd blocks
+ "cbz x25, 52f\n"
+ "tbz x25, #2, 48f\n"
+ "ldr s1, [x24], #0x4\n"
+ "ldr s2, [x23], #0x4\n"
+ "tbz x25, #1, 47f\n"
+ "ld1 { v1.h }[2], [x24], #0x2\n"
+ "ld1 { v2.h }[2], [x23], #0x2\n"
+ "tbz x25, #0, 50f\n"
+ "ld1 { v1.b }[6], [x24]\n"
+ "ld1 { v2.b }[6], [x23]\n"
+ "b 50f\n"
+ "47:" // Height 2: Multiply loop: Ragged operand read: partial_1_4
+ "tbz x25, #0, 50f\n"
+ "ld1 { v1.b }[4], [x24]\n"
+ "ld1 { v2.b }[4], [x23]\n"
+ "b 50f\n"
+ "48:" // Height 2: Multiply loop: Ragged operand read: partial_2_0
+ "tbz x25, #1, 49f\n"
+ "ldr h1, [x24], #0x2\n"
+ "ldr h2, [x23], #0x2\n"
+ "tbz x25, #0, 50f\n"
+ "ld1 { v1.b }[2], [x24]\n"
+ "ld1 { v2.b }[2], [x23]\n"
+ "b 50f\n"
+ "49:" // Height 2: Multiply loop: Ragged operand read: partial_1_0
+ "ldr b1, [x24, #0x0]\n"
+ "ldr b2, [x23, #0x0]\n"
+ "50:" // Height 2: Multiply loop: Ragged operand read: Done
+ "trn1 v0.2d, v1.2d, v2.2d\n"
+ "tbnz %x[flags], #31, 51f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ "51:" // Height 2: Multiply loop: unique 8: skip row sum
+ "ldr q25, [x28, #0x0]\n"
+ "ldr q24, [x28, #0x10]\n"
+ ".inst 0x4e99a410 // smmla v16.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a414 // smmla v20.4s, v0.16b, v24.16b\n"
+ "ldr q25, [x28, #0x20]\n"
+ "ldr q24, [x28, #0x30]\n"
+ ".inst 0x4e99a411 // smmla v17.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a415 // smmla v21.4s, v0.16b, v24.16b\n"
+ "ldr q25, [x28, #0x40]\n"
+ "ldr q24, [x28, #0x50]\n"
+ ".inst 0x4e99a412 // smmla v18.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a416 // smmla v22.4s, v0.16b, v24.16b\n"
+ "ldr q25, [x28, #0x60]\n"
+ "ldr q24, [x28, #0x70]\n"
+ ".inst 0x4e99a413 // smmla v19.4s, v0.16b, v25.16b\n"
+ ".inst 0x4e98a417 // smmla v23.4s, v0.16b, v24.16b\n"
+ "add x28, x28, #0x80\n"
+ "52:" // Height 2: Multiply loop: No odd multiplies
+ "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+ "add x26, x26, #0x1\n"
+ "cmp x26, x20\n"
+ "bne 36b\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+ "uzp1 v24.2d, v16.2d, v20.2d\n"
+ "add x23, x27, x20\n"
+ "uzp2 v16.2d, v16.2d, v20.2d\n"
+ "uzp1 v20.2d, v17.2d, v21.2d\n"
+ "uzp2 v17.2d, v17.2d, v21.2d\n"
+ "prfm pstl1keep, [x27, #0x0]\n"
+ "prfm pstl1keep, [x23, #0x0]\n"
+ "uzp1 v21.2d, v18.2d, v22.2d\n"
+ "uzp2 v18.2d, v18.2d, v22.2d\n"
+ "uzp1 v22.2d, v19.2d, v23.2d\n"
+ "uzp2 v19.2d, v19.2d, v23.2d\n"
+ "mov v23.16b, v24.16b\n"
+ "tbnz %x[flags], #31, 53f\n"
+ "add x20, %x[qp], %[b_offset]\n"
+ "ld1r { v24.4s }, [x20]\n"
+ "addp v11.4s, v11.4s, v11.4s\n"
+ "neg v24.4s, v24.4s\n"
+ "dup v12.4s, v11.s[3]\n"
+ "dup v11.4s, v11.s[0]\n"
+ "mul v11.4s, v11.4s, v24.4s\n"
+ "mul v12.4s, v12.4s, v24.4s\n"
+ "53:" // Height 2: skip row sum fixup
+ "ldr q28, [x10, #0x0]\n"
+ "ldr q27, [x10, #0x10]\n"
+ "add v23.4s, v23.4s, v11.4s\n"
+ "add v20.4s, v20.4s, v11.4s\n"
+ "ldr q26, [x10, #0x20]\n"
+ "ldr q25, [x10, #0x30]\n"
+ "add v21.4s, v21.4s, v11.4s\n"
+ "add v22.4s, v22.4s, v11.4s\n"
+ "add v16.4s, v16.4s, v12.4s\n"
+ "add v17.4s, v17.4s, v12.4s\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1r { v24.4s }, [x20]\n"
+ "add v18.4s, v18.4s, v12.4s\n"
+ "add v19.4s, v19.4s, v12.4s\n"
+ "orr %x[flags], %x[flags], #0x80000000\n"
+ "add x20, %x[qp], %[per_layer_right_shift]\n"
+ "add v23.4s, v23.4s, v28.4s\n"
+ "add v20.4s, v20.4s, v27.4s\n"
+ "add x10, x10, #0x40\n"
+ "add v21.4s, v21.4s, v26.4s\n"
+ "add v22.4s, v22.4s, v25.4s\n"
+ "add v16.4s, v16.4s, v28.4s\n"
+ "ld1r { v0.4s }, [x20]\n"
+ "add v17.4s, v17.4s, v27.4s\n"
+ "add v18.4s, v18.4s, v26.4s\n"
+ "add v19.4s, v19.4s, v25.4s\n"
+ "sqrdmulh v23.4s, v23.4s, v24.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v24.4s\n"
+ "sqrdmulh v21.4s, v21.4s, v24.4s\n"
+ "sqrdmulh v22.4s, v22.4s, v24.4s\n"
+ "sqrdmulh v16.4s, v16.4s, v24.4s\n"
+ "sqrdmulh v17.4s, v17.4s, v24.4s\n"
+ "sqrdmulh v18.4s, v18.4s, v24.4s\n"
+ "sqrdmulh v19.4s, v19.4s, v24.4s\n"
+ "tbz %x[flags], #5, 54f\n"
+ "and v24.16b, v23.16b, v0.16b\n"
+ "sshr v24.4s, v24.4s, #0x1f\n"
+ "sqadd v23.4s, v23.4s, v24.4s\n"
+ "and v30.16b, v20.16b, v0.16b\n"
+ "and v29.16b, v21.16b, v0.16b\n"
+ "and v28.16b, v22.16b, v0.16b\n"
+ "and v27.16b, v16.16b, v0.16b\n"
+ "and v26.16b, v17.16b, v0.16b\n"
+ "and v25.16b, v18.16b, v0.16b\n"
+ "and v24.16b, v19.16b, v0.16b\n"
+ "sshr v30.4s, v30.4s, #0x1f\n"
+ "sshr v29.4s, v29.4s, #0x1f\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
+ "sshr v27.4s, v27.4s, #0x1f\n"
+ "sshr v26.4s, v26.4s, #0x1f\n"
+ "sshr v25.4s, v25.4s, #0x1f\n"
+ "sshr v24.4s, v24.4s, #0x1f\n"
+ "sqadd v20.4s, v20.4s, v30.4s\n"
+ "sqadd v21.4s, v21.4s, v29.4s\n"
+ "sqadd v22.4s, v22.4s, v28.4s\n"
+ "sqadd v16.4s, v16.4s, v27.4s\n"
+ "sqadd v17.4s, v17.4s, v26.4s\n"
+ "sqadd v18.4s, v18.4s, v25.4s\n"
+ "sqadd v19.4s, v19.4s, v24.4s\n"
+ "54:" // Height 2: no shift correction
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1r { v26.4s }, [x20]\n"
+ "srshl v23.4s, v23.4s, v0.4s\n"
+ "srshl v20.4s, v20.4s, v0.4s\n"
+ "srshl v21.4s, v21.4s, v0.4s\n"
+ "srshl v22.4s, v22.4s, v0.4s\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1r { v25.4s }, [x20]\n"
+ "srshl v16.4s, v16.4s, v0.4s\n"
+ "srshl v17.4s, v17.4s, v0.4s\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1r { v24.4s }, [x20]\n"
+ "srshl v18.4s, v18.4s, v0.4s\n"
+ "srshl v19.4s, v19.4s, v0.4s\n"
+ "cmp x9, #0x10\n"
+ "add v23.4s, v23.4s, v26.4s\n"
+ "add v20.4s, v20.4s, v26.4s\n"
+ "add v21.4s, v21.4s, v26.4s\n"
+ "add v22.4s, v22.4s, v26.4s\n"
+ "add v16.4s, v16.4s, v26.4s\n"
+ "add v17.4s, v17.4s, v26.4s\n"
+ "add v18.4s, v18.4s, v26.4s\n"
+ "add v19.4s, v19.4s, v26.4s\n"
+ "smin v23.4s, v23.4s, v25.4s\n"
+ "smin v20.4s, v20.4s, v25.4s\n"
+ "smin v21.4s, v21.4s, v25.4s\n"
+ "smin v22.4s, v22.4s, v25.4s\n"
+ "smin v16.4s, v16.4s, v25.4s\n"
+ "smin v17.4s, v17.4s, v25.4s\n"
+ "smin v18.4s, v18.4s, v25.4s\n"
+ "smin v19.4s, v19.4s, v25.4s\n"
+ "smax v23.4s, v23.4s, v24.4s\n"
+ "smax v20.4s, v20.4s, v24.4s\n"
+ "smax v21.4s, v21.4s, v24.4s\n"
+ "smax v22.4s, v22.4s, v24.4s\n"
+ "smax v16.4s, v16.4s, v24.4s\n"
+ "smax v17.4s, v17.4s, v24.4s\n"
+ "smax v18.4s, v18.4s, v24.4s\n"
+ "smax v19.4s, v19.4s, v24.4s\n"
+ "uzp1 v23.8h, v23.8h, v20.8h\n"
+ "uzp1 v20.8h, v21.8h, v22.8h\n"
+ "uzp1 v16.8h, v16.8h, v17.8h\n"
+ "uzp1 v17.8h, v18.8h, v19.8h\n"
+ "uzp1 v23.16b, v23.16b, v20.16b\n"
+ "uzp1 v16.16b, v16.16b, v17.16b\n"
+ "bge 63f\n"
+ "tbz x9, #3, 58f\n"
+ "str d23, [x27], #0x8\n"
+ "str d16, [x23], #0x8\n"
+ "tbz x9, #2, 56f\n"
+ "st1 { v23.s }[2], [x27], #0x4\n"
+ "st1 { v16.s }[2], [x23], #0x4\n"
+ "tbz x9, #1, 55f\n"
+ "st1 { v23.h }[6], [x27], #0x2\n"
+ "st1 { v16.h }[6], [x23], #0x2\n"
+ "tbz x9, #0, 62f\n"
+ "st1 { v23.b }[14], [x27]\n"
+ "st1 { v16.b }[14], [x23]\n"
+ "b 62f\n"
+ "55:" // Height 2: Partial direct writeback: partial_1_12
+ "tbz x9, #0, 62f\n"
+ "st1 { v23.b }[12], [x27]\n"
+ "st1 { v16.b }[12], [x23]\n"
+ "b 62f\n"
+ "56:" // Height 2: Partial direct writeback: partial_2_8
+ "tbz x9, #1, 57f\n"
+ "st1 { v23.h }[4], [x27], #0x2\n"
+ "st1 { v16.h }[4], [x23], #0x2\n"
+ "tbz x9, #0, 62f\n"
+ "st1 { v23.b }[10], [x27]\n"
+ "st1 { v16.b }[10], [x23]\n"
+ "b 62f\n"
+ "57:" // Height 2: Partial direct writeback: partial_1_8
+ "tbz x9, #0, 62f\n"
+ "st1 { v23.b }[8], [x27]\n"
+ "st1 { v16.b }[8], [x23]\n"
+ "b 62f\n"
+ "58:" // Height 2: Partial direct writeback: partial_4_0
+ "tbz x9, #2, 60f\n"
+ "str s23, [x27], #0x4\n"
+ "str s16, [x23], #0x4\n"
+ "tbz x9, #1, 59f\n"
+ "st1 { v23.h }[2], [x27], #0x2\n"
+ "st1 { v16.h }[2], [x23], #0x2\n"
+ "tbz x9, #0, 62f\n"
+ "st1 { v23.b }[6], [x27]\n"
+ "st1 { v16.b }[6], [x23]\n"
+ "b 62f\n"
+ "59:" // Height 2: Partial direct writeback: partial_1_4
+ "tbz x9, #0, 62f\n"
+ "st1 { v23.b }[4], [x27]\n"
+ "st1 { v16.b }[4], [x23]\n"
+ "b 62f\n"
+ "60:" // Height 2: Partial direct writeback: partial_2_0
+ "tbz x9, #1, 61f\n"
+ "str h23, [x27], #0x2\n"
+ "str h16, [x23], #0x2\n"
+ "tbz x9, #0, 62f\n"
+ "st1 { v23.b }[2], [x27]\n"
+ "st1 { v16.b }[2], [x23]\n"
+ "b 62f\n"
+ "61:" // Height 2: Partial direct writeback: partial_1_0
+ "str b23, [x27, #0x0]\n"
+ "str b16, [x23, #0x0]\n"
+ "62:" // Height 2: Partial direct writeback: Done
+ "b 64f\n"
+ "63:" // Height 2: Full writeback
+ "str q23, [x27, #0x0]\n"
+ "add x27, x27, #0x10\n"
+ "str q16, [x23, #0x0]\n"
+ "64:" // Height 2: Writeback done
+ "subs x9, x9, #0x10\n"
+ "bgt 34b\n"
+ "b 130f\n"
+ "65:" // Height 3
+ "mov x10, %x[col_bias]\n"
+ "movi v11.4s, #0x0\n"
+ "movi v12.4s, #0x0\n"
+ "bic %x[flags], %x[flags], #0x80000000\n"
+ "movi v13.4s, #0x0\n"
+ "movi v15.16b, #0x1\n"
+ "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
+ "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+ "mov x27, %x[output_ptr]\n"
+ "66:" // Height 3: Column loop
+ "movi v16.4s, #0x0\n"
+ "movi v17.4s, #0x0\n"
+ "movi v18.4s, #0x0\n"
+ "movi v19.4s, #0x0\n"
+ "movi v20.4s, #0x0\n"
+ "movi v21.4s, #0x0\n"
+ "movi v22.4s, #0x0\n"
+ "movi v23.4s, #0x0\n"
+ "movi v24.4s, #0x0\n"
+ "movi v25.4s, #0x0\n"
+ "movi v26.4s, #0x0\n"
+ "movi v27.4s, #0x0\n"
+ "movi v28.4s, #0x0\n"
+ "movi v29.4s, #0x0\n"
+ "movi v30.4s, #0x0\n"
+ "movi v31.4s, #0x0\n"
+ "67:" // Height 3: setup done
+ "mov x26, #0x0\n"
+ "68:" // Height 3: String loop
+ "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+ "ldr w25, [x20, x26, LSL #0x2]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 69f\n"
+ "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x24, [x20, #0x0]\n"
+ "ldr x23, [x20, #0x8]\n"
+ "ldr x22, [x20, #0x10]\n"
+ "cbnz x26, 70f\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+ "add x24, x24, x20\n"
+ "add x23, x23, x20\n"
+ "add x22, x22, x20\n"
+ "b 70f\n"
+ "69:" // Height 3: setup direct input
+ "mov x24, %x[input_ptr]\n"
+ "add x23, x24, x21\n"
+ "add x22, x23, x21\n"
+ "70:" // Height 3: input setup done
+ "cmp x25, #0x10\n"
+ "blt 75f\n"
+ "ldr q1, [x24, #0x0]\n"
+ "ldr q2, [x23, #0x0]\n"
+ "cmp x25, #0x20\n"
+ "ldr q3, [x22, #0x0]\n"
+ "ldr q5, [x28, #0x0]\n"
+ "ldr q6, [x28, #0x10]\n"
+ "ldr q7, [x28, #0x20]\n"
+ "ldr q8, [x28, #0x30]\n"
+ "ldr q9, [x28, #0x40]\n"
+ "ldr q10, [x28, #0x50]\n"
+ "blt 73f\n"
+ "71:" // Height 3: Multiply loop: Main loop head
+ "trn1 v0.2d, v1.2d, v2.2d\n"
+ "trn2 v1.2d, v1.2d, v2.2d\n"
+ ".inst 0x4e85a410 // smmla v16.4s, v0.16b, v5.16b\n"
+ "trn1 v2.2d, v3.2d, v4.2d\n"
+ ".inst 0x4e85a458 // smmla v24.4s, v2.16b, v5.16b\n"
+ "ldr q14, [x28, #0x70]\n"
+ ".inst 0x4e86a414 // smmla v20.4s, v0.16b, v6.16b\n"
+ "trn2 v3.2d, v3.2d, v4.2d\n"
+ "ldr q5, [x28, #0x60]\n"
+ ".inst 0x4e86a45c // smmla v28.4s, v2.16b, v6.16b\n"
+ "ldr q4, [x28, #0x80]\n"
+ ".inst 0x4e87a411 // smmla v17.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e87a459 // smmla v25.4s, v2.16b, v7.16b\n"
+ "ldr q7, [x28, #0x90]\n"
+ "add x24, x24, #0x10\n"
+ ".inst 0x4e88a415 // smmla v21.4s, v0.16b, v8.16b\n"
+ ".inst 0x4e88a45d // smmla v29.4s, v2.16b, v8.16b\n"
+ "ldr q8, [x28, #0xa0]\n"
+ "add x23, x23, #0x10\n"
+ ".inst 0x4e89a412 // smmla v18.4s, v0.16b, v9.16b\n"
+ ".inst 0x4e89a45a // smmla v26.4s, v2.16b, v9.16b\n"
+ "ldr q9, [x28, #0xb0]\n"
+ "add x22, x22, #0x10\n"
+ ".inst 0x4e8aa416 // smmla v22.4s, v0.16b, v10.16b\n"
+ ".inst 0x4e8aa45e // smmla v30.4s, v2.16b, v10.16b\n"
+ "ldr q10, [x28, #0xc0]\n"
+ ".inst 0x4e85a413 // smmla v19.4s, v0.16b, v5.16b\n"
+ ".inst 0x4e85a45b // smmla v27.4s, v2.16b, v5.16b\n"
+ "ldr q6, [x28, #0xd0]\n"
+ ".inst 0x4e8ea417 // smmla v23.4s, v0.16b, v14.16b\n"
+ ".inst 0x4e8ea45f // smmla v31.4s, v2.16b, v14.16b\n"
+ "ldr q5, [x28, #0xe0]\n"
+ ".inst 0x4e84a430 // smmla v16.4s, v1.16b, v4.16b\n"
+ ".inst 0x4e84a478 // smmla v24.4s, v3.16b, v4.16b\n"
+ "ldr q4, [x28, #0xf0]\n"
+ "add x28, x28, #0x100\n"
+ ".inst 0x4e87a434 // smmla v20.4s, v1.16b, v7.16b\n"
+ ".inst 0x4e87a47c // smmla v28.4s, v3.16b, v7.16b\n"
+ ".inst 0x4e88a431 // smmla v17.4s, v1.16b, v8.16b\n"
+ ".inst 0x4e88a479 // smmla v25.4s, v3.16b, v8.16b\n"
+ ".inst 0x4e89a435 // smmla v21.4s, v1.16b, v9.16b\n"
+ ".inst 0x4e89a47d // smmla v29.4s, v3.16b, v9.16b\n"
+ ".inst 0x4e8aa432 // smmla v18.4s, v1.16b, v10.16b\n"
+ ".inst 0x4e8aa47a // smmla v26.4s, v3.16b, v10.16b\n"
+ ".inst 0x4e86a436 // smmla v22.4s, v1.16b, v6.16b\n"
+ ".inst 0x4e86a47e // smmla v30.4s, v3.16b, v6.16b\n"
+ ".inst 0x4e85a433 // smmla v19.4s, v1.16b, v5.16b\n"
+ ".inst 0x4e85a47b // smmla v27.4s, v3.16b, v5.16b\n"
+ ".inst 0x4e84a437 // smmla v23.4s, v1.16b, v4.16b\n"
+ ".inst 0x4e84a47f // smmla v31.4s, v3.16b, v4.16b\n"
+ "tbnz %x[flags], #31, 72f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n"
+ ".inst 0x4e8f942b // sdot v11.4s, v1.16b, v15.16b\n"
+ ".inst 0x4e8f946d // sdot v13.4s, v3.16b, v15.16b\n"
+ "72:" // Height 3: Multiply loop: unique 9: skip row sum
+ "ldr q1, [x24, #0x0]\n"
+ "ldr q2, [x23, #0x0]\n"
+ "sub x25, x25, #0x10\n"
+ "cmp x25, #0x20\n"
+ "ldr q3, [x22, #0x0]\n"
+ "ldr q5, [x28, #0x0]\n"
+ "ldr q6, [x28, #0x10]\n"
+ "ldr q7, [x28, #0x20]\n"
+ "ldr q8, [x28, #0x30]\n"
+ "ldr q9, [x28, #0x40]\n"
+ "ldr q10, [x28, #0x50]\n"
+ "prfm pldl1keep, [x24, #0x80]\n"
+ "prfm pldl1keep, [x23, #0x80]\n"
+ "prfm pldl1keep, [x22, #0x80]\n"
+ "bge 71b\n"
+ "73:" // Height 3: Multiply loop: Single iteration only
+ "trn1 v0.2d, v1.2d, v2.2d\n"
+ "trn2 v1.2d, v1.2d, v2.2d\n"
+ ".inst 0x4e85a410 // smmla v16.4s, v0.16b, v5.16b\n"
+ "trn1 v2.2d, v3.2d, v4.2d\n"
+ ".inst 0x4e85a458 // smmla v24.4s, v2.16b, v5.16b\n"
+ "ldr q14, [x28, #0x70]\n"
+ ".inst 0x4e86a414 // smmla v20.4s, v0.16b, v6.16b\n"
+ "trn2 v3.2d, v3.2d, v4.2d\n"
+ "ldr q5, [x28, #0x60]\n"
+ ".inst 0x4e86a45c // smmla v28.4s, v2.16b, v6.16b\n"
+ "ldr q4, [x28, #0x80]\n"
+ ".inst 0x4e87a411 // smmla v17.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e87a459 // smmla v25.4s, v2.16b, v7.16b\n"
+ "ldr q7, [x28, #0x90]\n"
+ "sub x25, x25, #0x10\n"
+ ".inst 0x4e88a415 // smmla v21.4s, v0.16b, v8.16b\n"
+ ".inst 0x4e88a45d // smmla v29.4s, v2.16b, v8.16b\n"
+ "ldr q8, [x28, #0xa0]\n"
+ "add x24, x24, #0x10\n"
+ ".inst 0x4e89a412 // smmla v18.4s, v0.16b, v9.16b\n"
+ ".inst 0x4e89a45a // smmla v26.4s, v2.16b, v9.16b\n"
+ "ldr q9, [x28, #0xb0]\n"
+ "add x23, x23, #0x10\n"
+ ".inst 0x4e8aa416 // smmla v22.4s, v0.16b, v10.16b\n"
+ ".inst 0x4e8aa45e // smmla v30.4s, v2.16b, v10.16b\n"
+ "ldr q10, [x28, #0xc0]\n"
+ "add x22, x22, #0x10\n"
+ ".inst 0x4e85a413 // smmla v19.4s, v0.16b, v5.16b\n"
+ ".inst 0x4e85a45b // smmla v27.4s, v2.16b, v5.16b\n"
+ "ldr q6, [x28, #0xd0]\n"
+ ".inst 0x4e8ea417 // smmla v23.4s, v0.16b, v14.16b\n"
+ ".inst 0x4e8ea45f // smmla v31.4s, v2.16b, v14.16b\n"
+ "ldr q5, [x28, #0xe0]\n"
+ ".inst 0x4e84a430 // smmla v16.4s, v1.16b, v4.16b\n"
+ ".inst 0x4e84a478 // smmla v24.4s, v3.16b, v4.16b\n"
+ "ldr q4, [x28, #0xf0]\n"
+ "add x28, x28, #0x100\n"
+ ".inst 0x4e87a434 // smmla v20.4s, v1.16b, v7.16b\n"
+ ".inst 0x4e87a47c // smmla v28.4s, v3.16b, v7.16b\n"
+ ".inst 0x4e88a431 // smmla v17.4s, v1.16b, v8.16b\n"
+ ".inst 0x4e88a479 // smmla v25.4s, v3.16b, v8.16b\n"
+ ".inst 0x4e89a435 // smmla v21.4s, v1.16b, v9.16b\n"
+ ".inst 0x4e89a47d // smmla v29.4s, v3.16b, v9.16b\n"
+ ".inst 0x4e8aa432 // smmla v18.4s, v1.16b, v10.16b\n"
+ ".inst 0x4e8aa47a // smmla v26.4s, v3.16b, v10.16b\n"
+ ".inst 0x4e86a436 // smmla v22.4s, v1.16b, v6.16b\n"
+ ".inst 0x4e86a47e // smmla v30.4s, v3.16b, v6.16b\n"
+ ".inst 0x4e85a433 // smmla v19.4s, v1.16b, v5.16b\n"
+ ".inst 0x4e85a47b // smmla v27.4s, v3.16b, v5.16b\n"
+ ".inst 0x4e84a437 // smmla v23.4s, v1.16b, v4.16b\n"
+ ".inst 0x4e84a47f // smmla v31.4s, v3.16b, v4.16b\n"
+ "tbnz %x[flags], #31, 74f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n"
+ ".inst 0x4e8f942b // sdot v11.4s, v1.16b, v15.16b\n"
+ ".inst 0x4e8f946d // sdot v13.4s, v3.16b, v15.16b\n"
+ "74:" // Height 3: Multiply loop: unique 10: skip row sum
+ "prfm pldl1keep, [x24, #0x80]\n"
+ "prfm pldl1keep, [x23, #0x80]\n"
+ "prfm pldl1keep, [x22, #0x80]\n"
+ "75:" // Height 3: Multiply loop: Main loop skip
+ "cbz x25, 84f\n"
+ "cmp x25, #0x8\n"
+ "blt 78f\n"
+ "76:" // Height 3: Multiply loop: Odd block loop
+ "ldr d1, [x24], #0x8\n"
+ "ldr d0, [x23], #0x8\n"
+ "trn1 v0.2d, v1.2d, v0.2d\n"
+ "ldr d1, [x22], #0x8\n"
+ "trn1 v2.2d, v1.2d, v2.2d\n"
+ "tbnz %x[flags], #31, 77f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n"
+ "77:" // Height 3: Multiply loop: unique 11: skip row sum
+ "ldr q3, [x28, #0x0]\n"
+ "ldr q1, [x28, #0x10]\n"
+ ".inst 0x4e83a410 // smmla v16.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e83a458 // smmla v24.4s, v2.16b, v3.16b\n"
+ "ldr q7, [x28, #0x20]\n"
+ "ldr q6, [x28, #0x30]\n"
+ "sub x25, x25, #0x8\n"
+ "cmp x25, #0x8\n"
+ "ldr q5, [x28, #0x40]\n"
+ "ldr q4, [x28, #0x50]\n"
+ ".inst 0x4e81a414 // smmla v20.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e81a45c // smmla v28.4s, v2.16b, v1.16b\n"
+ "ldr q3, [x28, #0x60]\n"
+ "ldr q1, [x28, #0x70]\n"
+ ".inst 0x4e87a411 // smmla v17.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e87a459 // smmla v25.4s, v2.16b, v7.16b\n"
+ ".inst 0x4e86a415 // smmla v21.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e86a45d // smmla v29.4s, v2.16b, v6.16b\n"
+ "add x28, x28, #0x80\n"
+ ".inst 0x4e85a412 // smmla v18.4s, v0.16b, v5.16b\n"
+ ".inst 0x4e85a45a // smmla v26.4s, v2.16b, v5.16b\n"
+ ".inst 0x4e84a416 // smmla v22.4s, v0.16b, v4.16b\n"
+ ".inst 0x4e84a45e // smmla v30.4s, v2.16b, v4.16b\n"
+ ".inst 0x4e83a413 // smmla v19.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e83a45b // smmla v27.4s, v2.16b, v3.16b\n"
+ ".inst 0x4e81a417 // smmla v23.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e81a45f // smmla v31.4s, v2.16b, v1.16b\n"
+ "bge 76b\n"
+ "78:" // Height 3: Multiply loop: Skip odd blocks
+ "cbz x25, 84f\n"
+ "tbz x25, #2, 80f\n"
+ "ldr s1, [x24], #0x4\n"
+ "ldr s2, [x23], #0x4\n"
+ "ldr s3, [x22], #0x4\n"
+ "tbz x25, #1, 79f\n"
+ "ld1 { v1.h }[2], [x24], #0x2\n"
+ "ld1 { v2.h }[2], [x23], #0x2\n"
+ "ld1 { v3.h }[2], [x22], #0x2\n"
+ "tbz x25, #0, 82f\n"
+ "ld1 { v1.b }[6], [x24]\n"
+ "ld1 { v2.b }[6], [x23]\n"
+ "ld1 { v3.b }[6], [x22]\n"
+ "b 82f\n"
+ "79:" // Height 3: Multiply loop: Ragged operand read: partial_1_4
+ "tbz x25, #0, 82f\n"
+ "ld1 { v1.b }[4], [x24]\n"
+ "ld1 { v2.b }[4], [x23]\n"
+ "ld1 { v3.b }[4], [x22]\n"
+ "b 82f\n"
+ "80:" // Height 3: Multiply loop: Ragged operand read: partial_2_0
+ "tbz x25, #1, 81f\n"
+ "ldr h1, [x24], #0x2\n"
+ "ldr h2, [x23], #0x2\n"
+ "ldr h3, [x22], #0x2\n"
+ "tbz x25, #0, 82f\n"
+ "ld1 { v1.b }[2], [x24]\n"
+ "ld1 { v2.b }[2], [x23]\n"
+ "ld1 { v3.b }[2], [x22]\n"
+ "b 82f\n"
+ "81:" // Height 3: Multiply loop: Ragged operand read: partial_1_0
+ "ldr b1, [x24, #0x0]\n"
+ "ldr b2, [x23, #0x0]\n"
+ "ldr b3, [x22, #0x0]\n"
+ "82:" // Height 3: Multiply loop: Ragged operand read: Done
+ "trn1 v0.2d, v1.2d, v2.2d\n"
+ "trn1 v2.2d, v3.2d, v4.2d\n"
+ "tbnz %x[flags], #31, 83f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n"
+ "83:" // Height 3: Multiply loop: unique 12: skip row sum
+ "ldr q1, [x28, #0x0]\n"
+ "ldr q3, [x28, #0x10]\n"
+ ".inst 0x4e81a410 // smmla v16.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e81a458 // smmla v24.4s, v2.16b, v1.16b\n"
+ "ldr q1, [x28, #0x20]\n"
+ "ldr q6, [x28, #0x30]\n"
+ ".inst 0x4e83a414 // smmla v20.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e83a45c // smmla v28.4s, v2.16b, v3.16b\n"
+ "ldr q5, [x28, #0x40]\n"
+ "ldr q4, [x28, #0x50]\n"
+ ".inst 0x4e81a411 // smmla v17.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e81a459 // smmla v25.4s, v2.16b, v1.16b\n"
+ "ldr q3, [x28, #0x60]\n"
+ "ldr q1, [x28, #0x70]\n"
+ ".inst 0x4e86a415 // smmla v21.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e86a45d // smmla v29.4s, v2.16b, v6.16b\n"
+ ".inst 0x4e85a412 // smmla v18.4s, v0.16b, v5.16b\n"
+ ".inst 0x4e85a45a // smmla v26.4s, v2.16b, v5.16b\n"
+ "add x28, x28, #0x80\n"
+ ".inst 0x4e84a416 // smmla v22.4s, v0.16b, v4.16b\n"
+ ".inst 0x4e84a45e // smmla v30.4s, v2.16b, v4.16b\n"
+ ".inst 0x4e83a413 // smmla v19.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e83a45b // smmla v27.4s, v2.16b, v3.16b\n"
+ ".inst 0x4e81a417 // smmla v23.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e81a45f // smmla v31.4s, v2.16b, v1.16b\n"
+ "84:" // Height 3: Multiply loop: No odd multiplies
+ "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+ "add x26, x26, #0x1\n"
+ "cmp x26, x20\n"
+ "bne 68b\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+ "uzp1 v0.2d, v16.2d, v20.2d\n"
+ "add x23, x27, x20\n"
+ "add x22, x23, x20\n"
+ "uzp2 v16.2d, v16.2d, v20.2d\n"
+ "uzp1 v20.2d, v17.2d, v21.2d\n"
+ "prfm pstl1keep, [x27, #0x0]\n"
+ "prfm pstl1keep, [x23, #0x0]\n"
+ "uzp2 v17.2d, v17.2d, v21.2d\n"
+ "uzp1 v21.2d, v18.2d, v22.2d\n"
+ "prfm pstl1keep, [x22, #0x0]\n"
+ "uzp2 v18.2d, v18.2d, v22.2d\n"
+ "uzp1 v22.2d, v19.2d, v23.2d\n"
+ "uzp2 v19.2d, v19.2d, v23.2d\n"
+ "uzp1 v24.2d, v24.2d, v28.2d\n"
+ "uzp1 v25.2d, v25.2d, v29.2d\n"
+ "uzp1 v26.2d, v26.2d, v30.2d\n"
+ "uzp1 v27.2d, v27.2d, v31.2d\n"
+ "mov v31.16b, v0.16b\n"
+ "tbnz %x[flags], #31, 85f\n"
+ "add x20, %x[qp], %[b_offset]\n"
+ "ld1r { v23.4s }, [x20]\n"
+ "addp v11.4s, v11.4s, v11.4s\n"
+ "addp v13.4s, v13.4s, v13.4s\n"
+ "neg v23.4s, v23.4s\n"
+ "dup v12.4s, v11.s[3]\n"
+ "dup v11.4s, v11.s[0]\n"
+ "dup v13.4s, v13.s[0]\n"
+ "mul v11.4s, v11.4s, v23.4s\n"
+ "mul v12.4s, v12.4s, v23.4s\n"
+ "mul v13.4s, v13.4s, v23.4s\n"
+ "85:" // Height 3: skip row sum fixup
+ "ldr q0, [x10, #0x0]\n"
+ "ldr q30, [x10, #0x10]\n"
+ "add v31.4s, v31.4s, v11.4s\n"
+ "add v20.4s, v20.4s, v11.4s\n"
+ "ldr q29, [x10, #0x20]\n"
+ "ldr q28, [x10, #0x30]\n"
+ "add v21.4s, v21.4s, v11.4s\n"
+ "add v22.4s, v22.4s, v11.4s\n"
+ "add v16.4s, v16.4s, v12.4s\n"
+ "add v17.4s, v17.4s, v12.4s\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1r { v23.4s }, [x20]\n"
+ "add v18.4s, v18.4s, v12.4s\n"
+ "add v19.4s, v19.4s, v12.4s\n"
+ "orr %x[flags], %x[flags], #0x80000000\n"
+ "add x20, %x[qp], %[per_layer_right_shift]\n"
+ "add v24.4s, v24.4s, v13.4s\n"
+ "add v25.4s, v25.4s, v13.4s\n"
+ "add x10, x10, #0x40\n"
+ "add v26.4s, v26.4s, v13.4s\n"
+ "add v27.4s, v27.4s, v13.4s\n"
+ "add v31.4s, v31.4s, v0.4s\n"
+ "add v20.4s, v20.4s, v30.4s\n"
+ "add v21.4s, v21.4s, v29.4s\n"
+ "add v22.4s, v22.4s, v28.4s\n"
+ "add v16.4s, v16.4s, v0.4s\n"
+ "add v17.4s, v17.4s, v30.4s\n"
+ "add v18.4s, v18.4s, v29.4s\n"
+ "add v19.4s, v19.4s, v28.4s\n"
+ "add v24.4s, v24.4s, v0.4s\n"
+ "ld1r { v0.4s }, [x20]\n"
+ "add v25.4s, v25.4s, v30.4s\n"
+ "add v26.4s, v26.4s, v29.4s\n"
+ "add v27.4s, v27.4s, v28.4s\n"
+ "sqrdmulh v31.4s, v31.4s, v23.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v23.4s\n"
+ "sqrdmulh v21.4s, v21.4s, v23.4s\n"
+ "sqrdmulh v22.4s, v22.4s, v23.4s\n"
+ "sqrdmulh v16.4s, v16.4s, v23.4s\n"
+ "sqrdmulh v17.4s, v17.4s, v23.4s\n"
+ "sqrdmulh v18.4s, v18.4s, v23.4s\n"
+ "sqrdmulh v19.4s, v19.4s, v23.4s\n"
+ "sqrdmulh v24.4s, v24.4s, v23.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v23.4s\n"
+ "sqrdmulh v26.4s, v26.4s, v23.4s\n"
+ "sqrdmulh v27.4s, v27.4s, v23.4s\n"
+ "tbz %x[flags], #5, 86f\n"
+ "and v1.16b, v31.16b, v0.16b\n"
+ "and v30.16b, v20.16b, v0.16b\n"
+ "and v29.16b, v21.16b, v0.16b\n"
+ "and v28.16b, v22.16b, v0.16b\n"
+ "and v23.16b, v16.16b, v0.16b\n"
+ "sshr v1.4s, v1.4s, #0x1f\n"
+ "sshr v30.4s, v30.4s, #0x1f\n"
+ "sshr v29.4s, v29.4s, #0x1f\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
+ "sshr v23.4s, v23.4s, #0x1f\n"
+ "sqadd v31.4s, v31.4s, v1.4s\n"
+ "sqadd v20.4s, v20.4s, v30.4s\n"
+ "sqadd v21.4s, v21.4s, v29.4s\n"
+ "sqadd v22.4s, v22.4s, v28.4s\n"
+ "sqadd v16.4s, v16.4s, v23.4s\n"
+ "and v3.16b, v17.16b, v0.16b\n"
+ "and v2.16b, v18.16b, v0.16b\n"
+ "and v1.16b, v19.16b, v0.16b\n"
+ "and v30.16b, v24.16b, v0.16b\n"
+ "and v29.16b, v25.16b, v0.16b\n"
+ "and v28.16b, v26.16b, v0.16b\n"
+ "and v23.16b, v27.16b, v0.16b\n"
+ "sshr v3.4s, v3.4s, #0x1f\n"
+ "sshr v2.4s, v2.4s, #0x1f\n"
+ "sshr v1.4s, v1.4s, #0x1f\n"
+ "sshr v30.4s, v30.4s, #0x1f\n"
+ "sshr v29.4s, v29.4s, #0x1f\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
+ "sshr v23.4s, v23.4s, #0x1f\n"
+ "sqadd v17.4s, v17.4s, v3.4s\n"
+ "sqadd v18.4s, v18.4s, v2.4s\n"
+ "sqadd v19.4s, v19.4s, v1.4s\n"
+ "sqadd v24.4s, v24.4s, v30.4s\n"
+ "sqadd v25.4s, v25.4s, v29.4s\n"
+ "sqadd v26.4s, v26.4s, v28.4s\n"
+ "sqadd v27.4s, v27.4s, v23.4s\n"
+ "86:" // Height 3: no shift correction
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1r { v29.4s }, [x20]\n"
+ "srshl v31.4s, v31.4s, v0.4s\n"
+ "srshl v20.4s, v20.4s, v0.4s\n"
+ "srshl v21.4s, v21.4s, v0.4s\n"
+ "srshl v22.4s, v22.4s, v0.4s\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1r { v28.4s }, [x20]\n"
+ "srshl v16.4s, v16.4s, v0.4s\n"
+ "srshl v17.4s, v17.4s, v0.4s\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1r { v23.4s }, [x20]\n"
+ "srshl v18.4s, v18.4s, v0.4s\n"
+ "srshl v19.4s, v19.4s, v0.4s\n"
+ "cmp x9, #0x10\n"
+ "srshl v24.4s, v24.4s, v0.4s\n"
+ "srshl v25.4s, v25.4s, v0.4s\n"
+ "srshl v26.4s, v26.4s, v0.4s\n"
+ "srshl v27.4s, v27.4s, v0.4s\n"
+ "add v31.4s, v31.4s, v29.4s\n"
+ "add v20.4s, v20.4s, v29.4s\n"
+ "add v21.4s, v21.4s, v29.4s\n"
+ "add v22.4s, v22.4s, v29.4s\n"
+ "add v16.4s, v16.4s, v29.4s\n"
+ "add v17.4s, v17.4s, v29.4s\n"
+ "add v18.4s, v18.4s, v29.4s\n"
+ "add v19.4s, v19.4s, v29.4s\n"
+ "add v24.4s, v24.4s, v29.4s\n"
+ "add v25.4s, v25.4s, v29.4s\n"
+ "add v26.4s, v26.4s, v29.4s\n"
+ "add v27.4s, v27.4s, v29.4s\n"
+ "smin v31.4s, v31.4s, v28.4s\n"
+ "smin v20.4s, v20.4s, v28.4s\n"
+ "smin v21.4s, v21.4s, v28.4s\n"
+ "smin v22.4s, v22.4s, v28.4s\n"
+ "smin v16.4s, v16.4s, v28.4s\n"
+ "smin v17.4s, v17.4s, v28.4s\n"
+ "smin v18.4s, v18.4s, v28.4s\n"
+ "smin v19.4s, v19.4s, v28.4s\n"
+ "smin v24.4s, v24.4s, v28.4s\n"
+ "smin v25.4s, v25.4s, v28.4s\n"
+ "smin v26.4s, v26.4s, v28.4s\n"
+ "smin v27.4s, v27.4s, v28.4s\n"
+ "smax v31.4s, v31.4s, v23.4s\n"
+ "smax v20.4s, v20.4s, v23.4s\n"
+ "smax v21.4s, v21.4s, v23.4s\n"
+ "smax v22.4s, v22.4s, v23.4s\n"
+ "smax v16.4s, v16.4s, v23.4s\n"
+ "smax v17.4s, v17.4s, v23.4s\n"
+ "smax v18.4s, v18.4s, v23.4s\n"
+ "smax v19.4s, v19.4s, v23.4s\n"
+ "smax v24.4s, v24.4s, v23.4s\n"
+ "smax v25.4s, v25.4s, v23.4s\n"
+ "smax v26.4s, v26.4s, v23.4s\n"
+ "smax v27.4s, v27.4s, v23.4s\n"
+ "uzp1 v31.8h, v31.8h, v20.8h\n"
+ "uzp1 v20.8h, v21.8h, v22.8h\n"
+ "uzp1 v16.8h, v16.8h, v17.8h\n"
+ "uzp1 v18.8h, v18.8h, v19.8h\n"
+ "uzp1 v24.8h, v24.8h, v25.8h\n"
+ "uzp1 v17.8h, v26.8h, v27.8h\n"
+ "uzp1 v31.16b, v31.16b, v20.16b\n"
+ "uzp1 v16.16b, v16.16b, v18.16b\n"
+ "uzp1 v24.16b, v24.16b, v17.16b\n"
+ "bge 95f\n"
+ "tbz x9, #3, 90f\n"
+ "str d31, [x27], #0x8\n"
+ "str d16, [x23], #0x8\n"
+ "str d24, [x22], #0x8\n"
+ "tbz x9, #2, 88f\n"
+ "st1 { v31.s }[2], [x27], #0x4\n"
+ "st1 { v16.s }[2], [x23], #0x4\n"
+ "st1 { v24.s }[2], [x22], #0x4\n"
+ "tbz x9, #1, 87f\n"
+ "st1 { v31.h }[6], [x27], #0x2\n"
+ "st1 { v16.h }[6], [x23], #0x2\n"
+ "st1 { v24.h }[6], [x22], #0x2\n"
+ "tbz x9, #0, 94f\n"
+ "st1 { v31.b }[14], [x27]\n"
+ "st1 { v16.b }[14], [x23]\n"
+ "st1 { v24.b }[14], [x22]\n"
+ "b 94f\n"
+ "87:" // Height 3: Partial direct writeback: partial_1_12
+ "tbz x9, #0, 94f\n"
+ "st1 { v31.b }[12], [x27]\n"
+ "st1 { v16.b }[12], [x23]\n"
+ "st1 { v24.b }[12], [x22]\n"
+ "b 94f\n"
+ "88:" // Height 3: Partial direct writeback: partial_2_8
+ "tbz x9, #1, 89f\n"
+ "st1 { v31.h }[4], [x27], #0x2\n"
+ "st1 { v16.h }[4], [x23], #0x2\n"
+ "st1 { v24.h }[4], [x22], #0x2\n"
+ "tbz x9, #0, 94f\n"
+ "st1 { v31.b }[10], [x27]\n"
+ "st1 { v16.b }[10], [x23]\n"
+ "st1 { v24.b }[10], [x22]\n"
+ "b 94f\n"
+ "89:" // Height 3: Partial direct writeback: partial_1_8
+ "tbz x9, #0, 94f\n"
+ "st1 { v31.b }[8], [x27]\n"
+ "st1 { v16.b }[8], [x23]\n"
+ "st1 { v24.b }[8], [x22]\n"
+ "b 94f\n"
+ "90:" // Height 3: Partial direct writeback: partial_4_0
+ "tbz x9, #2, 92f\n"
+ "str s31, [x27], #0x4\n"
+ "str s16, [x23], #0x4\n"
+ "str s24, [x22], #0x4\n"
+ "tbz x9, #1, 91f\n"
+ "st1 { v31.h }[2], [x27], #0x2\n"
+ "st1 { v16.h }[2], [x23], #0x2\n"
+ "st1 { v24.h }[2], [x22], #0x2\n"
+ "tbz x9, #0, 94f\n"
+ "st1 { v31.b }[6], [x27]\n"
+ "st1 { v16.b }[6], [x23]\n"
+ "st1 { v24.b }[6], [x22]\n"
+ "b 94f\n"
+ "91:" // Height 3: Partial direct writeback: partial_1_4
+ "tbz x9, #0, 94f\n"
+ "st1 { v31.b }[4], [x27]\n"
+ "st1 { v16.b }[4], [x23]\n"
+ "st1 { v24.b }[4], [x22]\n"
+ "b 94f\n"
+ "92:" // Height 3: Partial direct writeback: partial_2_0
+ "tbz x9, #1, 93f\n"
+ "str h31, [x27], #0x2\n"
+ "str h16, [x23], #0x2\n"
+ "str h24, [x22], #0x2\n"
+ "tbz x9, #0, 94f\n"
+ "st1 { v31.b }[2], [x27]\n"
+ "st1 { v16.b }[2], [x23]\n"
+ "st1 { v24.b }[2], [x22]\n"
+ "b 94f\n"
+ "93:" // Height 3: Partial direct writeback: partial_1_0
+ "str b31, [x27, #0x0]\n"
+ "str b16, [x23, #0x0]\n"
+ "str b24, [x22, #0x0]\n"
+ "94:" // Height 3: Partial direct writeback: Done
+ "b 96f\n"
+ "95:" // Height 3: Full writeback
+ "str q31, [x27, #0x0]\n"
+ "add x27, x27, #0x10\n"
+ "str q16, [x23, #0x0]\n"
+ "str q24, [x22, #0x0]\n"
+ "96:" // Height 3: Writeback done
+ "subs x9, x9, #0x10\n"
+ "bgt 66b\n"
+ "b 130f\n"
+ "97:" // Height 4
+ "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
+ "mov x20, #0x4\n"
+ "mov x10, %x[col_bias]\n"
+ "movi v11.4s, #0x0\n"
+ "movi v12.4s, #0x0\n"
+ "movi v13.4s, #0x0\n"
+ "bic %x[flags], %x[flags], #0x80000000\n"
+ "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
+ "movi v14.4s, #0x0\n"
+ "movi v15.16b, #0x1\n"
+ "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
+ "mov x27, %x[output_ptr]\n"
+ "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
+ "98:" // Height 4: Column loop
+ "movi v16.4s, #0x0\n"
+ "movi v17.4s, #0x0\n"
+ "movi v18.4s, #0x0\n"
+ "movi v19.4s, #0x0\n"
+ "movi v20.4s, #0x0\n"
+ "movi v21.4s, #0x0\n"
+ "movi v22.4s, #0x0\n"
+ "movi v23.4s, #0x0\n"
+ "movi v24.4s, #0x0\n"
+ "movi v25.4s, #0x0\n"
+ "movi v26.4s, #0x0\n"
+ "movi v27.4s, #0x0\n"
+ "movi v28.4s, #0x0\n"
+ "movi v29.4s, #0x0\n"
+ "movi v30.4s, #0x0\n"
+ "movi v31.4s, #0x0\n"
+ "99:" // Height 4: setup done
+ "mov x26, #0x0\n"
+ "100:" // Height 4: String loop
+ "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
+ "ldr w25, [x20, x26, LSL #0x2]\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 101f\n"
+ "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
+ "add x20, x20, x21, LSL #3\n"
+ "ldr x24, [x20, #0x0]\n"
+ "ldr x23, [x20, #0x8]\n"
+ "ldr x22, [x20, #0x10]\n"
+ "ldr x21, [x20, #0x18]\n"
+ "cbnz x26, 102f\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
+ "add x24, x24, x20\n"
+ "add x23, x23, x20\n"
+ "add x22, x22, x20\n"
+ "add x21, x21, x20\n"
+ "b 102f\n"
+ "101:" // Height 4: setup direct input
+ "mov x24, %x[input_ptr]\n"
+ "add x23, x24, x21\n"
+ "add x22, x23, x21\n"
+ "add x21, x22, x21\n"
+ "102:" // Height 4: input setup done
+ "cmp x25, #0x10\n"
+ "blt 107f\n"
+ "ldr q1, [x24, #0x0]\n"
+ "ldr q2, [x23, #0x0]\n"
+ "cmp x25, #0x20\n"
+ "ldr q3, [x22, #0x0]\n"
+ "ldr q4, [x21, #0x0]\n"
+ "ldr q5, [x28, #0x0]\n"
+ "ldr q6, [x28, #0x10]\n"
+ "ldr q7, [x28, #0x20]\n"
+ "ldr q8, [x28, #0x30]\n"
+ "ldr q9, [x28, #0x40]\n"
+ "ldr q10, [x28, #0x50]\n"
+ "blt 105f\n"
+ "103:" // Height 4: Multiply loop: Main loop head
+ "trn1 v0.2d, v1.2d, v2.2d\n"
+ "trn2 v1.2d, v1.2d, v2.2d\n"
+ ".inst 0x4e85a410 // smmla v16.4s, v0.16b, v5.16b\n"
+ "add x24, x24, #0x10\n"
+ "trn1 v2.2d, v3.2d, v4.2d\n"
+ "trn2 v3.2d, v3.2d, v4.2d\n"
+ "ldr q4, [x28, #0x60]\n"
+ ".inst 0x4e85a458 // smmla v24.4s, v2.16b, v5.16b\n"
+ "ldr q5, [x28, #0x70]\n"
+ ".inst 0x4e86a414 // smmla v20.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e86a45c // smmla v28.4s, v2.16b, v6.16b\n"
+ "ldr q6, [x28, #0x80]\n"
+ ".inst 0x4e87a411 // smmla v17.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e87a459 // smmla v25.4s, v2.16b, v7.16b\n"
+ "ldr q7, [x28, #0x90]\n"
+ "add x23, x23, #0x10\n"
+ ".inst 0x4e88a415 // smmla v21.4s, v0.16b, v8.16b\n"
+ ".inst 0x4e88a45d // smmla v29.4s, v2.16b, v8.16b\n"
+ "ldr q8, [x28, #0xa0]\n"
+ "add x22, x22, #0x10\n"
+ ".inst 0x4e89a412 // smmla v18.4s, v0.16b, v9.16b\n"
+ ".inst 0x4e89a45a // smmla v26.4s, v2.16b, v9.16b\n"
+ "ldr q9, [x28, #0xb0]\n"
+ "add x21, x21, #0x10\n"
+ ".inst 0x4e8aa416 // smmla v22.4s, v0.16b, v10.16b\n"
+ ".inst 0x4e8aa45e // smmla v30.4s, v2.16b, v10.16b\n"
+ "ldr q10, [x28, #0xc0]\n"
+ ".inst 0x4e84a413 // smmla v19.4s, v0.16b, v4.16b\n"
+ ".inst 0x4e84a45b // smmla v27.4s, v2.16b, v4.16b\n"
+ "ldr q4, [x28, #0xd0]\n"
+ ".inst 0x4e85a417 // smmla v23.4s, v0.16b, v5.16b\n"
+ ".inst 0x4e85a45f // smmla v31.4s, v2.16b, v5.16b\n"
+ "ldr q5, [x28, #0xe0]\n"
+ ".inst 0x4e86a430 // smmla v16.4s, v1.16b, v6.16b\n"
+ ".inst 0x4e86a478 // smmla v24.4s, v3.16b, v6.16b\n"
+ "ldr q6, [x28, #0xf0]\n"
+ "add x28, x28, #0x100\n"
+ ".inst 0x4e87a434 // smmla v20.4s, v1.16b, v7.16b\n"
+ ".inst 0x4e87a47c // smmla v28.4s, v3.16b, v7.16b\n"
+ ".inst 0x4e88a431 // smmla v17.4s, v1.16b, v8.16b\n"
+ ".inst 0x4e88a479 // smmla v25.4s, v3.16b, v8.16b\n"
+ ".inst 0x4e89a435 // smmla v21.4s, v1.16b, v9.16b\n"
+ ".inst 0x4e89a47d // smmla v29.4s, v3.16b, v9.16b\n"
+ ".inst 0x4e8aa432 // smmla v18.4s, v1.16b, v10.16b\n"
+ ".inst 0x4e8aa47a // smmla v26.4s, v3.16b, v10.16b\n"
+ ".inst 0x4e84a436 // smmla v22.4s, v1.16b, v4.16b\n"
+ ".inst 0x4e84a47e // smmla v30.4s, v3.16b, v4.16b\n"
+ ".inst 0x4e85a433 // smmla v19.4s, v1.16b, v5.16b\n"
+ ".inst 0x4e85a47b // smmla v27.4s, v3.16b, v5.16b\n"
+ ".inst 0x4e86a437 // smmla v23.4s, v1.16b, v6.16b\n"
+ ".inst 0x4e86a47f // smmla v31.4s, v3.16b, v6.16b\n"
+ "tbnz %x[flags], #31, 104f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n"
+ ".inst 0x4e8f942b // sdot v11.4s, v1.16b, v15.16b\n"
+ ".inst 0x4e8f946d // sdot v13.4s, v3.16b, v15.16b\n"
+ "104:" // Height 4: Multiply loop: unique 13: skip row sum
+ "ldr q1, [x24, #0x0]\n"
+ "ldr q2, [x23, #0x0]\n"
+ "sub x25, x25, #0x10\n"
+ "cmp x25, #0x20\n"
+ "ldr q3, [x22, #0x0]\n"
+ "ldr q4, [x21, #0x0]\n"
+ "ldr q5, [x28, #0x0]\n"
+ "ldr q6, [x28, #0x10]\n"
+ "ldr q7, [x28, #0x20]\n"
+ "ldr q8, [x28, #0x30]\n"
+ "ldr q9, [x28, #0x40]\n"
+ "ldr q10, [x28, #0x50]\n"
+ "prfm pldl1keep, [x24, #0x80]\n"
+ "prfm pldl1keep, [x23, #0x80]\n"
+ "prfm pldl1keep, [x22, #0x80]\n"
+ "prfm pldl1keep, [x21, #0x80]\n"
+ "bge 103b\n"
+ "105:" // Height 4: Multiply loop: Single iteration only
+ "trn1 v0.2d, v1.2d, v2.2d\n"
+ "trn2 v1.2d, v1.2d, v2.2d\n"
+ ".inst 0x4e85a410 // smmla v16.4s, v0.16b, v5.16b\n"
+ "sub x25, x25, #0x10\n"
+ "trn1 v2.2d, v3.2d, v4.2d\n"
+ "trn2 v3.2d, v3.2d, v4.2d\n"
+ "ldr q4, [x28, #0x60]\n"
+ ".inst 0x4e85a458 // smmla v24.4s, v2.16b, v5.16b\n"
+ "ldr q5, [x28, #0x70]\n"
+ ".inst 0x4e86a414 // smmla v20.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e86a45c // smmla v28.4s, v2.16b, v6.16b\n"
+ "ldr q6, [x28, #0x80]\n"
+ ".inst 0x4e87a411 // smmla v17.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e87a459 // smmla v25.4s, v2.16b, v7.16b\n"
+ "ldr q7, [x28, #0x90]\n"
+ "add x24, x24, #0x10\n"
+ ".inst 0x4e88a415 // smmla v21.4s, v0.16b, v8.16b\n"
+ ".inst 0x4e88a45d // smmla v29.4s, v2.16b, v8.16b\n"
+ "ldr q8, [x28, #0xa0]\n"
+ "add x23, x23, #0x10\n"
+ ".inst 0x4e89a412 // smmla v18.4s, v0.16b, v9.16b\n"
+ ".inst 0x4e89a45a // smmla v26.4s, v2.16b, v9.16b\n"
+ "ldr q9, [x28, #0xb0]\n"
+ "add x22, x22, #0x10\n"
+ ".inst 0x4e8aa416 // smmla v22.4s, v0.16b, v10.16b\n"
+ ".inst 0x4e8aa45e // smmla v30.4s, v2.16b, v10.16b\n"
+ "ldr q10, [x28, #0xc0]\n"
+ "add x21, x21, #0x10\n"
+ ".inst 0x4e84a413 // smmla v19.4s, v0.16b, v4.16b\n"
+ ".inst 0x4e84a45b // smmla v27.4s, v2.16b, v4.16b\n"
+ "ldr q4, [x28, #0xd0]\n"
+ ".inst 0x4e85a417 // smmla v23.4s, v0.16b, v5.16b\n"
+ ".inst 0x4e85a45f // smmla v31.4s, v2.16b, v5.16b\n"
+ "ldr q5, [x28, #0xe0]\n"
+ ".inst 0x4e86a430 // smmla v16.4s, v1.16b, v6.16b\n"
+ ".inst 0x4e86a478 // smmla v24.4s, v3.16b, v6.16b\n"
+ "ldr q6, [x28, #0xf0]\n"
+ "add x28, x28, #0x100\n"
+ ".inst 0x4e87a434 // smmla v20.4s, v1.16b, v7.16b\n"
+ ".inst 0x4e87a47c // smmla v28.4s, v3.16b, v7.16b\n"
+ ".inst 0x4e88a431 // smmla v17.4s, v1.16b, v8.16b\n"
+ ".inst 0x4e88a479 // smmla v25.4s, v3.16b, v8.16b\n"
+ ".inst 0x4e89a435 // smmla v21.4s, v1.16b, v9.16b\n"
+ ".inst 0x4e89a47d // smmla v29.4s, v3.16b, v9.16b\n"
+ ".inst 0x4e8aa432 // smmla v18.4s, v1.16b, v10.16b\n"
+ ".inst 0x4e8aa47a // smmla v26.4s, v3.16b, v10.16b\n"
+ ".inst 0x4e84a436 // smmla v22.4s, v1.16b, v4.16b\n"
+ ".inst 0x4e84a47e // smmla v30.4s, v3.16b, v4.16b\n"
+ ".inst 0x4e85a433 // smmla v19.4s, v1.16b, v5.16b\n"
+ ".inst 0x4e85a47b // smmla v27.4s, v3.16b, v5.16b\n"
+ ".inst 0x4e86a437 // smmla v23.4s, v1.16b, v6.16b\n"
+ ".inst 0x4e86a47f // smmla v31.4s, v3.16b, v6.16b\n"
+ "tbnz %x[flags], #31, 106f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n"
+ ".inst 0x4e8f942b // sdot v11.4s, v1.16b, v15.16b\n"
+ ".inst 0x4e8f946d // sdot v13.4s, v3.16b, v15.16b\n"
+ "106:" // Height 4: Multiply loop: unique 14: skip row sum
+ "prfm pldl1keep, [x24, #0x80]\n"
+ "prfm pldl1keep, [x23, #0x80]\n"
+ "prfm pldl1keep, [x22, #0x80]\n"
+ "prfm pldl1keep, [x21, #0x80]\n"
+ "107:" // Height 4: Multiply loop: Main loop skip
+ "cbz x25, 116f\n"
+ "cmp x25, #0x8\n"
+ "blt 110f\n"
+ "108:" // Height 4: Multiply loop: Odd block loop
+ "ldr d1, [x24], #0x8\n"
+ "ldr d0, [x23], #0x8\n"
+ "trn1 v0.2d, v1.2d, v0.2d\n"
+ "ldr d2, [x22], #0x8\n"
+ "ldr d1, [x21], #0x8\n"
+ "trn1 v2.2d, v2.2d, v1.2d\n"
+ "tbnz %x[flags], #31, 109f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n"
+ "109:" // Height 4: Multiply loop: unique 15: skip row sum
+ "ldr q3, [x28, #0x0]\n"
+ "ldr q1, [x28, #0x10]\n"
+ ".inst 0x4e83a410 // smmla v16.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e83a458 // smmla v24.4s, v2.16b, v3.16b\n"
+ "ldr q7, [x28, #0x20]\n"
+ "ldr q6, [x28, #0x30]\n"
+ "sub x25, x25, #0x8\n"
+ "cmp x25, #0x8\n"
+ "ldr q5, [x28, #0x40]\n"
+ "ldr q4, [x28, #0x50]\n"
+ ".inst 0x4e81a414 // smmla v20.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e81a45c // smmla v28.4s, v2.16b, v1.16b\n"
+ "ldr q3, [x28, #0x60]\n"
+ "ldr q1, [x28, #0x70]\n"
+ ".inst 0x4e87a411 // smmla v17.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e87a459 // smmla v25.4s, v2.16b, v7.16b\n"
+ ".inst 0x4e86a415 // smmla v21.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e86a45d // smmla v29.4s, v2.16b, v6.16b\n"
+ "add x28, x28, #0x80\n"
+ ".inst 0x4e85a412 // smmla v18.4s, v0.16b, v5.16b\n"
+ ".inst 0x4e85a45a // smmla v26.4s, v2.16b, v5.16b\n"
+ ".inst 0x4e84a416 // smmla v22.4s, v0.16b, v4.16b\n"
+ ".inst 0x4e84a45e // smmla v30.4s, v2.16b, v4.16b\n"
+ ".inst 0x4e83a413 // smmla v19.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e83a45b // smmla v27.4s, v2.16b, v3.16b\n"
+ ".inst 0x4e81a417 // smmla v23.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e81a45f // smmla v31.4s, v2.16b, v1.16b\n"
+ "bge 108b\n"
+ "110:" // Height 4: Multiply loop: Skip odd blocks
+ "cbz x25, 116f\n"
+ "tbz x25, #2, 112f\n"
+ "ldr s1, [x24], #0x4\n"
+ "ldr s2, [x23], #0x4\n"
+ "ldr s3, [x22], #0x4\n"
+ "ldr s9, [x21], #0x4\n"
+ "tbz x25, #1, 111f\n"
+ "ld1 { v1.h }[2], [x24], #0x2\n"
+ "ld1 { v2.h }[2], [x23], #0x2\n"
+ "ld1 { v3.h }[2], [x22], #0x2\n"
+ "ld1 { v9.h }[2], [x21], #0x2\n"
+ "tbz x25, #0, 114f\n"
+ "ld1 { v1.b }[6], [x24]\n"
+ "ld1 { v2.b }[6], [x23]\n"
+ "ld1 { v3.b }[6], [x22]\n"
+ "ld1 { v9.b }[6], [x21]\n"
+ "b 114f\n"
+ "111:" // Height 4: Multiply loop: Ragged operand read: partial_1_4
+ "tbz x25, #0, 114f\n"
+ "ld1 { v1.b }[4], [x24]\n"
+ "ld1 { v2.b }[4], [x23]\n"
+ "ld1 { v3.b }[4], [x22]\n"
+ "ld1 { v9.b }[4], [x21]\n"
+ "b 114f\n"
+ "112:" // Height 4: Multiply loop: Ragged operand read: partial_2_0
+ "tbz x25, #1, 113f\n"
+ "ldr h1, [x24], #0x2\n"
+ "ldr h2, [x23], #0x2\n"
+ "ldr h3, [x22], #0x2\n"
+ "ldr h9, [x21], #0x2\n"
+ "tbz x25, #0, 114f\n"
+ "ld1 { v1.b }[2], [x24]\n"
+ "ld1 { v2.b }[2], [x23]\n"
+ "ld1 { v3.b }[2], [x22]\n"
+ "ld1 { v9.b }[2], [x21]\n"
+ "b 114f\n"
+ "113:" // Height 4: Multiply loop: Ragged operand read: partial_1_0
+ "ldr b1, [x24, #0x0]\n"
+ "ldr b2, [x23, #0x0]\n"
+ "ldr b3, [x22, #0x0]\n"
+ "ldr b9, [x21, #0x0]\n"
+ "114:" // Height 4: Multiply loop: Ragged operand read: Done
+ "trn1 v0.2d, v1.2d, v2.2d\n"
+ "trn1 v2.2d, v3.2d, v9.2d\n"
+ "tbnz %x[flags], #31, 115f\n"
+ ".inst 0x4e8f940b // sdot v11.4s, v0.16b, v15.16b\n"
+ ".inst 0x4e8f944d // sdot v13.4s, v2.16b, v15.16b\n"
+ "115:" // Height 4: Multiply loop: unique 16: skip row sum
+ "ldr q1, [x28, #0x0]\n"
+ "ldr q3, [x28, #0x10]\n"
+ ".inst 0x4e81a410 // smmla v16.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e81a458 // smmla v24.4s, v2.16b, v1.16b\n"
+ "ldr q1, [x28, #0x20]\n"
+ "ldr q6, [x28, #0x30]\n"
+ ".inst 0x4e83a414 // smmla v20.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e83a45c // smmla v28.4s, v2.16b, v3.16b\n"
+ "ldr q5, [x28, #0x40]\n"
+ "ldr q4, [x28, #0x50]\n"
+ ".inst 0x4e81a411 // smmla v17.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e81a459 // smmla v25.4s, v2.16b, v1.16b\n"
+ "ldr q3, [x28, #0x60]\n"
+ "ldr q1, [x28, #0x70]\n"
+ ".inst 0x4e86a415 // smmla v21.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e86a45d // smmla v29.4s, v2.16b, v6.16b\n"
+ ".inst 0x4e85a412 // smmla v18.4s, v0.16b, v5.16b\n"
+ ".inst 0x4e85a45a // smmla v26.4s, v2.16b, v5.16b\n"
+ "add x28, x28, #0x80\n"
+ ".inst 0x4e84a416 // smmla v22.4s, v0.16b, v4.16b\n"
+ ".inst 0x4e84a45e // smmla v30.4s, v2.16b, v4.16b\n"
+ ".inst 0x4e83a413 // smmla v19.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e83a45b // smmla v27.4s, v2.16b, v3.16b\n"
+ ".inst 0x4e81a417 // smmla v23.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e81a45f // smmla v31.4s, v2.16b, v1.16b\n"
+ "116:" // Height 4: Multiply loop: No odd multiplies
+ "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
+ "add x26, x26, #0x1\n"
+ "cmp x26, x20\n"
+ "bne 100b\n"
+ "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
+ "uzp1 v0.2d, v16.2d, v20.2d\n"
+ "add x23, x27, x20\n"
+ "add x22, x23, x20\n"
+ "add x21, x22, x20\n"
+ "uzp2 v16.2d, v16.2d, v20.2d\n"
+ "uzp1 v20.2d, v17.2d, v21.2d\n"
+ "prfm pstl1keep, [x27, #0x0]\n"
+ "uzp2 v17.2d, v17.2d, v21.2d\n"
+ "uzp1 v21.2d, v18.2d, v22.2d\n"
+ "prfm pstl1keep, [x23, #0x0]\n"
+ "prfm pstl1keep, [x22, #0x0]\n"
+ "uzp2 v18.2d, v18.2d, v22.2d\n"
+ "uzp1 v22.2d, v19.2d, v23.2d\n"
+ "prfm pstl1keep, [x21, #0x0]\n"
+ "uzp2 v19.2d, v19.2d, v23.2d\n"
+ "uzp1 v23.2d, v24.2d, v28.2d\n"
+ "uzp2 v24.2d, v24.2d, v28.2d\n"
+ "uzp1 v28.2d, v25.2d, v29.2d\n"
+ "uzp2 v25.2d, v25.2d, v29.2d\n"
+ "uzp1 v29.2d, v26.2d, v30.2d\n"
+ "uzp2 v26.2d, v26.2d, v30.2d\n"
+ "uzp1 v30.2d, v27.2d, v31.2d\n"
+ "uzp2 v27.2d, v27.2d, v31.2d\n"
+ "mov v31.16b, v0.16b\n"
+ "tbnz %x[flags], #31, 117f\n"
+ "add x20, %x[qp], %[b_offset]\n"
+ "ld1r { v0.4s }, [x20]\n"
+ "addp v11.4s, v11.4s, v11.4s\n"
+ "addp v13.4s, v13.4s, v13.4s\n"
+ "neg v0.4s, v0.4s\n"
+ "dup v12.4s, v11.s[3]\n"
+ "dup v11.4s, v11.s[0]\n"
+ "dup v14.4s, v13.s[3]\n"
+ "dup v13.4s, v13.s[0]\n"
+ "mul v11.4s, v11.4s, v0.4s\n"
+ "mul v12.4s, v12.4s, v0.4s\n"
+ "mul v13.4s, v13.4s, v0.4s\n"
+ "mul v14.4s, v14.4s, v0.4s\n"
+ "117:" // Height 4: skip row sum fixup
+ "ldr q0, [x10, #0x0]\n"
+ "ldr q4, [x10, #0x10]\n"
+ "add v31.4s, v31.4s, v11.4s\n"
+ "add v20.4s, v20.4s, v11.4s\n"
+ "ldr q3, [x10, #0x20]\n"
+ "ldr q2, [x10, #0x30]\n"
+ "add v21.4s, v21.4s, v11.4s\n"
+ "add v22.4s, v22.4s, v11.4s\n"
+ "add v16.4s, v16.4s, v12.4s\n"
+ "add v17.4s, v17.4s, v12.4s\n"
+ "add x20, %x[qp], %[per_layer_mul]\n"
+ "ld1r { v1.4s }, [x20]\n"
+ "add v18.4s, v18.4s, v12.4s\n"
+ "add v19.4s, v19.4s, v12.4s\n"
+ "orr %x[flags], %x[flags], #0x80000000\n"
+ "add x20, %x[qp], %[per_layer_right_shift]\n"
+ "add v23.4s, v23.4s, v13.4s\n"
+ "add v28.4s, v28.4s, v13.4s\n"
+ "add x10, x10, #0x40\n"
+ "add v29.4s, v29.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v13.4s\n"
+ "add v24.4s, v24.4s, v14.4s\n"
+ "add v25.4s, v25.4s, v14.4s\n"
+ "add v26.4s, v26.4s, v14.4s\n"
+ "add v27.4s, v27.4s, v14.4s\n"
+ "add v31.4s, v31.4s, v0.4s\n"
+ "add v20.4s, v20.4s, v4.4s\n"
+ "add v21.4s, v21.4s, v3.4s\n"
+ "add v22.4s, v22.4s, v2.4s\n"
+ "add v16.4s, v16.4s, v0.4s\n"
+ "add v17.4s, v17.4s, v4.4s\n"
+ "add v18.4s, v18.4s, v3.4s\n"
+ "add v19.4s, v19.4s, v2.4s\n"
+ "add v23.4s, v23.4s, v0.4s\n"
+ "add v28.4s, v28.4s, v4.4s\n"
+ "add v29.4s, v29.4s, v3.4s\n"
+ "add v30.4s, v30.4s, v2.4s\n"
+ "add v24.4s, v24.4s, v0.4s\n"
+ "ld1r { v0.4s }, [x20]\n"
+ "add v25.4s, v25.4s, v4.4s\n"
+ "add v26.4s, v26.4s, v3.4s\n"
+ "add v27.4s, v27.4s, v2.4s\n"
+ "sqrdmulh v31.4s, v31.4s, v1.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v1.4s\n"
+ "sqrdmulh v21.4s, v21.4s, v1.4s\n"
+ "sqrdmulh v22.4s, v22.4s, v1.4s\n"
+ "sqrdmulh v16.4s, v16.4s, v1.4s\n"
+ "sqrdmulh v17.4s, v17.4s, v1.4s\n"
+ "sqrdmulh v18.4s, v18.4s, v1.4s\n"
+ "sqrdmulh v19.4s, v19.4s, v1.4s\n"
+ "sqrdmulh v23.4s, v23.4s, v1.4s\n"
+ "sqrdmulh v28.4s, v28.4s, v1.4s\n"
+ "sqrdmulh v29.4s, v29.4s, v1.4s\n"
+ "sqrdmulh v30.4s, v30.4s, v1.4s\n"
+ "sqrdmulh v24.4s, v24.4s, v1.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v1.4s\n"
+ "sqrdmulh v26.4s, v26.4s, v1.4s\n"
+ "sqrdmulh v27.4s, v27.4s, v1.4s\n"
+ "tbz %x[flags], #5, 118f\n"
+ "and v2.16b, v31.16b, v0.16b\n"
+ "and v1.16b, v20.16b, v0.16b\n"
+ "sshr v2.4s, v2.4s, #0x1f\n"
+ "sshr v1.4s, v1.4s, #0x1f\n"
+ "sqadd v31.4s, v31.4s, v2.4s\n"
+ "sqadd v20.4s, v20.4s, v1.4s\n"
+ "and v7.16b, v21.16b, v0.16b\n"
+ "and v6.16b, v22.16b, v0.16b\n"
+ "and v5.16b, v16.16b, v0.16b\n"
+ "and v4.16b, v17.16b, v0.16b\n"
+ "and v3.16b, v18.16b, v0.16b\n"
+ "and v2.16b, v19.16b, v0.16b\n"
+ "and v1.16b, v23.16b, v0.16b\n"
+ "sshr v7.4s, v7.4s, #0x1f\n"
+ "sshr v6.4s, v6.4s, #0x1f\n"
+ "sshr v5.4s, v5.4s, #0x1f\n"
+ "sshr v4.4s, v4.4s, #0x1f\n"
+ "sshr v3.4s, v3.4s, #0x1f\n"
+ "sshr v2.4s, v2.4s, #0x1f\n"
+ "sshr v1.4s, v1.4s, #0x1f\n"
+ "sqadd v21.4s, v21.4s, v7.4s\n"
+ "sqadd v22.4s, v22.4s, v6.4s\n"
+ "sqadd v16.4s, v16.4s, v5.4s\n"
+ "sqadd v17.4s, v17.4s, v4.4s\n"
+ "sqadd v18.4s, v18.4s, v3.4s\n"
+ "sqadd v19.4s, v19.4s, v2.4s\n"
+ "sqadd v23.4s, v23.4s, v1.4s\n"
+ "and v7.16b, v28.16b, v0.16b\n"
+ "and v6.16b, v29.16b, v0.16b\n"
+ "and v5.16b, v30.16b, v0.16b\n"
+ "and v4.16b, v24.16b, v0.16b\n"
+ "and v3.16b, v25.16b, v0.16b\n"
+ "and v2.16b, v26.16b, v0.16b\n"
+ "and v1.16b, v27.16b, v0.16b\n"
+ "sshr v7.4s, v7.4s, #0x1f\n"
+ "sshr v6.4s, v6.4s, #0x1f\n"
+ "sshr v5.4s, v5.4s, #0x1f\n"
+ "sshr v4.4s, v4.4s, #0x1f\n"
+ "sshr v3.4s, v3.4s, #0x1f\n"
+ "sshr v2.4s, v2.4s, #0x1f\n"
+ "sshr v1.4s, v1.4s, #0x1f\n"
+ "sqadd v28.4s, v28.4s, v7.4s\n"
+ "sqadd v29.4s, v29.4s, v6.4s\n"
+ "sqadd v30.4s, v30.4s, v5.4s\n"
+ "sqadd v24.4s, v24.4s, v4.4s\n"
+ "sqadd v25.4s, v25.4s, v3.4s\n"
+ "sqadd v26.4s, v26.4s, v2.4s\n"
+ "sqadd v27.4s, v27.4s, v1.4s\n"
+ "118:" // Height 4: no shift correction
+ "add x20, %x[qp], %[c_offset]\n"
+ "ld1r { v3.4s }, [x20]\n"
+ "srshl v31.4s, v31.4s, v0.4s\n"
+ "srshl v20.4s, v20.4s, v0.4s\n"
+ "srshl v21.4s, v21.4s, v0.4s\n"
+ "srshl v22.4s, v22.4s, v0.4s\n"
+ "add x20, %x[qp], %[maxval]\n"
+ "ld1r { v2.4s }, [x20]\n"
+ "srshl v16.4s, v16.4s, v0.4s\n"
+ "srshl v17.4s, v17.4s, v0.4s\n"
+ "add x20, %x[qp], %[minval]\n"
+ "ld1r { v1.4s }, [x20]\n"
+ "srshl v18.4s, v18.4s, v0.4s\n"
+ "srshl v19.4s, v19.4s, v0.4s\n"
+ "cmp x9, #0x10\n"
+ "srshl v23.4s, v23.4s, v0.4s\n"
+ "srshl v28.4s, v28.4s, v0.4s\n"
+ "srshl v29.4s, v29.4s, v0.4s\n"
+ "srshl v30.4s, v30.4s, v0.4s\n"
+ "srshl v24.4s, v24.4s, v0.4s\n"
+ "srshl v25.4s, v25.4s, v0.4s\n"
+ "srshl v26.4s, v26.4s, v0.4s\n"
+ "srshl v27.4s, v27.4s, v0.4s\n"
+ "add v31.4s, v31.4s, v3.4s\n"
+ "add v20.4s, v20.4s, v3.4s\n"
+ "add v21.4s, v21.4s, v3.4s\n"
+ "add v22.4s, v22.4s, v3.4s\n"
+ "add v16.4s, v16.4s, v3.4s\n"
+ "add v17.4s, v17.4s, v3.4s\n"
+ "add v18.4s, v18.4s, v3.4s\n"
+ "add v19.4s, v19.4s, v3.4s\n"
+ "add v23.4s, v23.4s, v3.4s\n"
+ "add v28.4s, v28.4s, v3.4s\n"
+ "add v29.4s, v29.4s, v3.4s\n"
+ "add v30.4s, v30.4s, v3.4s\n"
+ "add v24.4s, v24.4s, v3.4s\n"
+ "add v25.4s, v25.4s, v3.4s\n"
+ "add v26.4s, v26.4s, v3.4s\n"
+ "add v27.4s, v27.4s, v3.4s\n"
+ "smin v31.4s, v31.4s, v2.4s\n"
+ "smin v20.4s, v20.4s, v2.4s\n"
+ "smin v21.4s, v21.4s, v2.4s\n"
+ "smin v22.4s, v22.4s, v2.4s\n"
+ "smin v16.4s, v16.4s, v2.4s\n"
+ "smin v17.4s, v17.4s, v2.4s\n"
+ "smin v18.4s, v18.4s, v2.4s\n"
+ "smin v19.4s, v19.4s, v2.4s\n"
+ "smin v23.4s, v23.4s, v2.4s\n"
+ "smin v28.4s, v28.4s, v2.4s\n"
+ "smin v29.4s, v29.4s, v2.4s\n"
+ "smin v30.4s, v30.4s, v2.4s\n"
+ "smin v24.4s, v24.4s, v2.4s\n"
+ "smin v25.4s, v25.4s, v2.4s\n"
+ "smin v26.4s, v26.4s, v2.4s\n"
+ "smin v27.4s, v27.4s, v2.4s\n"
+ "smax v31.4s, v31.4s, v1.4s\n"
+ "smax v20.4s, v20.4s, v1.4s\n"
+ "smax v21.4s, v21.4s, v1.4s\n"
+ "smax v22.4s, v22.4s, v1.4s\n"
+ "smax v16.4s, v16.4s, v1.4s\n"
+ "smax v17.4s, v17.4s, v1.4s\n"
+ "smax v18.4s, v18.4s, v1.4s\n"
+ "smax v19.4s, v19.4s, v1.4s\n"
+ "smax v23.4s, v23.4s, v1.4s\n"
+ "smax v28.4s, v28.4s, v1.4s\n"
+ "smax v29.4s, v29.4s, v1.4s\n"
+ "smax v30.4s, v30.4s, v1.4s\n"
+ "smax v24.4s, v24.4s, v1.4s\n"
+ "smax v25.4s, v25.4s, v1.4s\n"
+ "smax v26.4s, v26.4s, v1.4s\n"
+ "smax v27.4s, v27.4s, v1.4s\n"
+ "uzp1 v31.8h, v31.8h, v20.8h\n"
+ "uzp1 v20.8h, v21.8h, v22.8h\n"
+ "uzp1 v16.8h, v16.8h, v17.8h\n"
+ "uzp1 v19.8h, v18.8h, v19.8h\n"
+ "uzp1 v23.8h, v23.8h, v28.8h\n"
+ "uzp1 v18.8h, v29.8h, v30.8h\n"
+ "uzp1 v24.8h, v24.8h, v25.8h\n"
+ "uzp1 v17.8h, v26.8h, v27.8h\n"
+ "uzp1 v31.16b, v31.16b, v20.16b\n"
+ "uzp1 v16.16b, v16.16b, v19.16b\n"
+ "uzp1 v23.16b, v23.16b, v18.16b\n"
+ "uzp1 v24.16b, v24.16b, v17.16b\n"
+ "bge 127f\n"
+ "tbz x9, #3, 122f\n"
+ "str d31, [x27], #0x8\n"
+ "str d16, [x23], #0x8\n"
+ "str d23, [x22], #0x8\n"
+ "str d24, [x21], #0x8\n"
+ "tbz x9, #2, 120f\n"
+ "st1 { v31.s }[2], [x27], #0x4\n"
+ "st1 { v16.s }[2], [x23], #0x4\n"
+ "st1 { v23.s }[2], [x22], #0x4\n"
+ "st1 { v24.s }[2], [x21], #0x4\n"
+ "tbz x9, #1, 119f\n"
+ "st1 { v31.h }[6], [x27], #0x2\n"
+ "st1 { v16.h }[6], [x23], #0x2\n"
+ "st1 { v23.h }[6], [x22], #0x2\n"
+ "st1 { v24.h }[6], [x21], #0x2\n"
+ "tbz x9, #0, 126f\n"
+ "st1 { v31.b }[14], [x27]\n"
+ "st1 { v16.b }[14], [x23]\n"
+ "st1 { v23.b }[14], [x22]\n"
+ "st1 { v24.b }[14], [x21]\n"
+ "b 126f\n"
+ "119:" // Height 4: Partial direct writeback: partial_1_12
+ "tbz x9, #0, 126f\n"
+ "st1 { v31.b }[12], [x27]\n"
+ "st1 { v16.b }[12], [x23]\n"
+ "st1 { v23.b }[12], [x22]\n"
+ "st1 { v24.b }[12], [x21]\n"
+ "b 126f\n"
+ "120:" // Height 4: Partial direct writeback: partial_2_8
+ "tbz x9, #1, 121f\n"
+ "st1 { v31.h }[4], [x27], #0x2\n"
+ "st1 { v16.h }[4], [x23], #0x2\n"
+ "st1 { v23.h }[4], [x22], #0x2\n"
+ "st1 { v24.h }[4], [x21], #0x2\n"
+ "tbz x9, #0, 126f\n"
+ "st1 { v31.b }[10], [x27]\n"
+ "st1 { v16.b }[10], [x23]\n"
+ "st1 { v23.b }[10], [x22]\n"
+ "st1 { v24.b }[10], [x21]\n"
+ "b 126f\n"
+ "121:" // Height 4: Partial direct writeback: partial_1_8
+ "tbz x9, #0, 126f\n"
+ "st1 { v31.b }[8], [x27]\n"
+ "st1 { v16.b }[8], [x23]\n"
+ "st1 { v23.b }[8], [x22]\n"
+ "st1 { v24.b }[8], [x21]\n"
+ "b 126f\n"
+ "122:" // Height 4: Partial direct writeback: partial_4_0
+ "tbz x9, #2, 124f\n"
+ "str s31, [x27], #0x4\n"
+ "str s16, [x23], #0x4\n"
+ "str s23, [x22], #0x4\n"
+ "str s24, [x21], #0x4\n"
+ "tbz x9, #1, 123f\n"
+ "st1 { v31.h }[2], [x27], #0x2\n"
+ "st1 { v16.h }[2], [x23], #0x2\n"
+ "st1 { v23.h }[2], [x22], #0x2\n"
+ "st1 { v24.h }[2], [x21], #0x2\n"
+ "tbz x9, #0, 126f\n"
+ "st1 { v31.b }[6], [x27]\n"
+ "st1 { v16.b }[6], [x23]\n"
+ "st1 { v23.b }[6], [x22]\n"
+ "st1 { v24.b }[6], [x21]\n"
+ "b 126f\n"
+ "123:" // Height 4: Partial direct writeback: partial_1_4
+ "tbz x9, #0, 126f\n"
+ "st1 { v31.b }[4], [x27]\n"
+ "st1 { v16.b }[4], [x23]\n"
+ "st1 { v23.b }[4], [x22]\n"
+ "st1 { v24.b }[4], [x21]\n"
+ "b 126f\n"
+ "124:" // Height 4: Partial direct writeback: partial_2_0
+ "tbz x9, #1, 125f\n"
+ "str h31, [x27], #0x2\n"
+ "str h16, [x23], #0x2\n"
+ "str h23, [x22], #0x2\n"
+ "str h24, [x21], #0x2\n"
+ "tbz x9, #0, 126f\n"
+ "st1 { v31.b }[2], [x27]\n"
+ "st1 { v16.b }[2], [x23]\n"
+ "st1 { v23.b }[2], [x22]\n"
+ "st1 { v24.b }[2], [x21]\n"
+ "b 126f\n"
+ "125:" // Height 4: Partial direct writeback: partial_1_0
+ "str b31, [x27, #0x0]\n"
+ "str b16, [x23, #0x0]\n"
+ "str b23, [x22, #0x0]\n"
+ "str b24, [x21, #0x0]\n"
+ "126:" // Height 4: Partial direct writeback: Done
+ "b 128f\n"
+ "127:" // Height 4: Full writeback
+ "str q31, [x27, #0x0]\n"
+ "add x27, x27, #0x10\n"
+ "str q16, [x23, #0x0]\n"
+ "str q23, [x22, #0x0]\n"
+ "str q24, [x21, #0x0]\n"
+ "128:" // Height 4: Writeback done
+ "subs x9, x9, #0x10\n"
+ "bgt 98b\n"
+ "subs %x[M], %x[M], #0x4\n"
+ "beq 130f\n"
+ "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "tbz %x[flags], #3, 129f\n"
+ "add x21, x21, #0x4\n"
+ "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
+ "b 1b\n"
+ "129:" // Update direct input
+ "mov x20, #0x4\n"
+ "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
+ "b 1b\n"
+ "130:" // Exit
+ : [M] "+&r" (M), [flags] "+&r" (flags), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr)
+ : [args_ptr] "r" (&ka), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
+ : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+ );
+}
+
+} // namespace arm_gemm
+#endif // __aarch64__
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-03 20:37:06 +0000