| /* |
| * Copyright (c) 2019-2021 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_u8qa_dot_4x16 ( |
| unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<uint8_t> A_arg, |
| size_t M, size_t N, const uint8_t *B_ptr, IndirectOutputArg<uint8_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 uint8_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 91f\n" |
| "cmp %x[M], #0x2\n" |
| "bgt 61f\n" |
| "beq 31f\n" |
| "movi v11.4s, #0x0\n" |
| "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" |
| "movi v15.16b, #0x1\n" |
| "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" |
| "mov x27, %x[col_bias]\n" |
| "bic %x[flags], %x[flags], #0x80000000\n" |
| "mov x26, %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" |
| "3:" // Height 1: setup done |
| "mov x25, #0x0\n" |
| "4:" // Height 1: String loop |
| "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" |
| "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" |
| "ldr w24, [x20, x25, LSL #0x2]\n" |
| "tbz %x[flags], #3, 5f\n" |
| "ldr x20, [%x[input_ptr], x25, LSL #0x3]\n" |
| "add x20, x20, x19, LSL #3\n" |
| "ldr x23, [x20, #0x0]\n" |
| "cbnz x25, 6f\n" |
| "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" |
| "add x23, x23, x19\n" |
| "b 6f\n" |
| "5:" // Height 1: setup direct input |
| "mov x23, %x[input_ptr]\n" |
| "6:" // Height 1: input setup done |
| "cmp x24, #0x10\n" |
| "blt 11f\n" |
| "ldr q0, [x23, #0x0]\n" |
| "ldr q4, [x28, #0x0]\n" |
| "cmp x24, #0x20\n" |
| "blt 9f\n" |
| "7:" // Height 1: Multiply loop: Main loop head |
| ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" |
| "ldr q5, [x28, #0x10]\n" |
| "add x23, x23, #0x10\n" |
| ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" |
| "ldr q6, [x28, #0x20]\n" |
| "ldr q7, [x28, #0x30]\n" |
| ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" |
| "ldr q8, [x28, #0x40]\n" |
| ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" |
| "ldr q9, [x28, #0x50]\n" |
| "ldr q10, [x28, #0x60]\n" |
| ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" |
| "ldr q4, [x28, #0x70]\n" |
| "ldr q5, [x28, #0x80]\n" |
| ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" |
| ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" |
| "ldr q6, [x28, #0x90]\n" |
| "ldr q7, [x28, #0xa0]\n" |
| ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" |
| "ldr q8, [x28, #0xb0]\n" |
| ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" |
| "ldr q9, [x28, #0xc0]\n" |
| "ldr q10, [x28, #0xd0]\n" |
| ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" |
| ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" |
| "ldr q4, [x28, #0xe0]\n" |
| "ldr q5, [x28, #0xf0]\n" |
| ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" |
| "add x28, x28, #0x100\n" |
| ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" |
| ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" |
| ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" |
| ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" |
| "tbnz %x[flags], #31, 8f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| "8:" // Height 1: Multiply loop: unique 1: skip row sum |
| "prfm pldl1keep, [x23, #0x80]\n" |
| "sub x24, x24, #0x10\n" |
| "ldr q0, [x23, #0x0]\n" |
| "cmp x24, #0x20\n" |
| "ldr q4, [x28, #0x0]\n" |
| "bge 7b\n" |
| "9:" // Height 1: Multiply loop: Single iteration only |
| ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" |
| "ldr q5, [x28, #0x10]\n" |
| "sub x24, x24, #0x10\n" |
| ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" |
| "ldr q6, [x28, #0x20]\n" |
| "add x23, x23, #0x10\n" |
| ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" |
| "ldr q7, [x28, #0x30]\n" |
| "ldr q8, [x28, #0x40]\n" |
| ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" |
| "ldr q9, [x28, #0x50]\n" |
| ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" |
| "ldr q10, [x28, #0x60]\n" |
| "ldr q4, [x28, #0x70]\n" |
| ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" |
| "ldr q5, [x28, #0x80]\n" |
| "ldr q6, [x28, #0x90]\n" |
| ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" |
| ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" |
| "ldr q7, [x28, #0xa0]\n" |
| "ldr q8, [x28, #0xb0]\n" |
| ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" |
| "ldr q9, [x28, #0xc0]\n" |
| ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" |
| "ldr q10, [x28, #0xd0]\n" |
| "ldr q4, [x28, #0xe0]\n" |
| ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" |
| ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" |
| "ldr q5, [x28, #0xf0]\n" |
| "add x28, x28, #0x100\n" |
| ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" |
| ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" |
| ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" |
| ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" |
| "tbnz %x[flags], #31, 10f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| "10:" // Height 1: Multiply loop: unique 2: skip row sum |
| "prfm pldl1keep, [x23, #0x80]\n" |
| "11:" // Height 1: Multiply loop: Main loop skip |
| "cbz x24, 18f\n" |
| "cmp x24, #0x4\n" |
| "blt 14f\n" |
| "12:" // Height 1: Multiply loop: Odd block loop |
| "ldr s0, [x23], #0x4\n" |
| "tbnz %x[flags], #31, 13f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| "13:" // Height 1: Multiply loop: unique 3: skip row sum |
| "ldr q6, [x28, #0x0]\n" |
| ".inst 0x6f80e0d0 // udot v16.4s, v6.16b, v0.4b[0]\n" |
| "ldr q7, [x28, #0x10]\n" |
| "sub x24, x24, #0x4\n" |
| ".inst 0x6f80e0f1 // udot v17.4s, v7.16b, v0.4b[0]\n" |
| "ldr q8, [x28, #0x20]\n" |
| "cmp x24, #0x4\n" |
| ".inst 0x6f80e112 // udot v18.4s, v8.16b, v0.4b[0]\n" |
| "ldr q9, [x28, #0x30]\n" |
| "add x28, x28, #0x40\n" |
| ".inst 0x6f80e133 // udot v19.4s, v9.16b, v0.4b[0]\n" |
| "bge 12b\n" |
| "cbz x24, 18f\n" |
| "14:" // Height 1: Multiply loop: Skip odd blocks |
| "tbz x24, #1, 15f\n" |
| "ldr h0, [x23], #0x2\n" |
| "tbz x24, #0, 16f\n" |
| "ld1 { v0.b }[2], [x23]\n" |
| "b 16f\n" |
| "15:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 |
| "ldr b0, [x23, #0x0]\n" |
| "16:" // Height 1: Multiply loop: Ragged operand read: Done |
| "tbnz %x[flags], #31, 17f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| "17:" // Height 1: Multiply loop: unique 4: skip row sum |
| "ldr q10, [x28, #0x0]\n" |
| ".inst 0x6f80e150 // udot v16.4s, v10.16b, v0.4b[0]\n" |
| "ldr q4, [x28, #0x10]\n" |
| "ldr q5, [x28, #0x20]\n" |
| ".inst 0x6f80e091 // udot v17.4s, v4.16b, v0.4b[0]\n" |
| "ldr q6, [x28, #0x30]\n" |
| "add x28, x28, #0x40\n" |
| ".inst 0x6f80e0b2 // udot v18.4s, v5.16b, v0.4b[0]\n" |
| ".inst 0x6f80e0d3 // udot v19.4s, v6.16b, v0.4b[0]\n" |
| "18:" // Height 1: Multiply loop: No odd multiplies |
| "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" |
| "add x25, x25, #0x1\n" |
| "cmp x25, x19\n" |
| "bne 4b\n" |
| "prfm pstl1keep, [x26, #0x0]\n" |
| "tbnz %x[flags], #31, 19f\n" |
| "addp v11.4s, v11.4s, v11.4s\n" |
| "add x22, %x[qp], %[b_offset]\n" |
| "addp v11.4s, v11.4s, v11.4s\n" |
| "ld1r { v1.4s }, [x22]\n" |
| "neg v1.4s, v1.4s\n" |
| "mul v11.4s, v11.4s, v1.4s\n" |
| "19:" // Height 1: skip row sum fixup |
| "add v16.4s, v16.4s, v11.4s\n" |
| "ldr q0, [x27, #0x0]\n" |
| "orr %x[flags], %x[flags], #0x80000000\n" |
| "add v17.4s, v17.4s, v11.4s\n" |
| "ldr q1, [x27, #0x10]\n" |
| "add x23, %x[qp], %[per_layer_right_shift]\n" |
| "add v18.4s, v18.4s, v11.4s\n" |
| "ldr q2, [x27, #0x20]\n" |
| "add x22, %x[qp], %[per_layer_mul]\n" |
| "add v19.4s, v19.4s, v11.4s\n" |
| "ldr q3, [x27, #0x30]\n" |
| "add x27, x27, #0x40\n" |
| "add v16.4s, v16.4s, v0.4s\n" |
| "ld1r { v0.4s }, [x23]\n" |
| "ld1r { v4.4s }, [x22]\n" |
| "add v17.4s, v17.4s, v1.4s\n" |
| "add v18.4s, v18.4s, v2.4s\n" |
| "add v19.4s, v19.4s, v3.4s\n" |
| "sqrdmulh v16.4s, v16.4s, v4.4s\n" |
| "sqrdmulh v17.4s, v17.4s, v4.4s\n" |
| "sqrdmulh v18.4s, v18.4s, v4.4s\n" |
| "sqrdmulh v19.4s, v19.4s, v4.4s\n" |
| "tbz %x[flags], #5, 20f\n" |
| "and v4.16b, v16.16b, v0.16b\n" |
| "and v5.16b, v17.16b, v0.16b\n" |
| "and v6.16b, v18.16b, v0.16b\n" |
| "and v7.16b, v19.16b, v0.16b\n" |
| "sshr v4.4s, v4.4s, #0x1f\n" |
| "sshr v5.4s, v5.4s, #0x1f\n" |
| "sshr v6.4s, v6.4s, #0x1f\n" |
| "sqadd v16.4s, v16.4s, v4.4s\n" |
| "sqadd v17.4s, v17.4s, v5.4s\n" |
| "sqadd v18.4s, v18.4s, v6.4s\n" |
| "sshr v7.4s, v7.4s, #0x1f\n" |
| "sqadd v19.4s, v19.4s, v7.4s\n" |
| "20:" // Height 1: no shift correction |
| "srshl v16.4s, v16.4s, v0.4s\n" |
| "add x22, %x[qp], %[c_offset]\n" |
| "ld1r { v4.4s }, [x22]\n" |
| "srshl v17.4s, v17.4s, v0.4s\n" |
| "add x22, %x[qp], %[minval]\n" |
| "srshl v18.4s, v18.4s, v0.4s\n" |
| "ld1r { v5.4s }, [x22]\n" |
| "add x22, %x[qp], %[maxval]\n" |
| "srshl v19.4s, v19.4s, v0.4s\n" |
| "ld1r { v6.4s }, [x22]\n" |
| "cmp x9, #0x10\n" |
| "add v16.4s, v16.4s, v4.4s\n" |
| "add v17.4s, v17.4s, v4.4s\n" |
| "add v18.4s, v18.4s, v4.4s\n" |
| "add v19.4s, v19.4s, v4.4s\n" |
| "smin v16.4s, v16.4s, v6.4s\n" |
| "smin v17.4s, v17.4s, v6.4s\n" |
| "smin v18.4s, v18.4s, v6.4s\n" |
| "smax v16.4s, v16.4s, v5.4s\n" |
| "smax v17.4s, v17.4s, v5.4s\n" |
| "smax v18.4s, v18.4s, v5.4s\n" |
| "smin v19.4s, v19.4s, v6.4s\n" |
| "uzp1 v16.8h, v16.8h, v17.8h\n" |
| "smax v19.4s, v19.4s, v5.4s\n" |
| "uzp1 v17.8h, v18.8h, v19.8h\n" |
| "uzp1 v16.16b, v16.16b, v17.16b\n" |
| "bge 29f\n" |
| "tbz x9, #3, 24f\n" |
| "str d16, [x26], #0x8\n" |
| "tbz x9, #2, 22f\n" |
| "st1 { v16.s }[2], [x26], #0x4\n" |
| "tbz x9, #1, 21f\n" |
| "st1 { v16.h }[6], [x26], #0x2\n" |
| "tbz x9, #0, 28f\n" |
| "st1 { v16.b }[14], [x26]\n" |
| "b 28f\n" |
| "21:" // Height 1: Partial direct writeback: partial_1_12 |
| "tbz x9, #0, 28f\n" |
| "st1 { v16.b }[12], [x26]\n" |
| "b 28f\n" |
| "22:" // Height 1: Partial direct writeback: partial_2_8 |
| "tbz x9, #1, 23f\n" |
| "st1 { v16.h }[4], [x26], #0x2\n" |
| "tbz x9, #0, 28f\n" |
| "st1 { v16.b }[10], [x26]\n" |
| "b 28f\n" |
| "23:" // Height 1: Partial direct writeback: partial_1_8 |
| "tbz x9, #0, 28f\n" |
| "st1 { v16.b }[8], [x26]\n" |
| "b 28f\n" |
| "24:" // Height 1: Partial direct writeback: partial_4_0 |
| "tbz x9, #2, 26f\n" |
| "str s16, [x26], #0x4\n" |
| "tbz x9, #1, 25f\n" |
| "st1 { v16.h }[2], [x26], #0x2\n" |
| "tbz x9, #0, 28f\n" |
| "st1 { v16.b }[6], [x26]\n" |
| "b 28f\n" |
| "25:" // Height 1: Partial direct writeback: partial_1_4 |
| "tbz x9, #0, 28f\n" |
| "st1 { v16.b }[4], [x26]\n" |
| "b 28f\n" |
| "26:" // Height 1: Partial direct writeback: partial_2_0 |
| "tbz x9, #1, 27f\n" |
| "str h16, [x26], #0x2\n" |
| "tbz x9, #0, 28f\n" |
| "st1 { v16.b }[2], [x26]\n" |
| "b 28f\n" |
| "27:" // Height 1: Partial direct writeback: partial_1_0 |
| "str b16, [x26, #0x0]\n" |
| "28:" // Height 1: Partial direct writeback: Done |
| "b 30f\n" |
| "29:" // Height 1: Full writeback |
| "str q16, [x26, #0x0]\n" |
| "add x26, x26, #0x10\n" |
| "30:" // Height 1: Writeback done |
| "subs x9, x9, #0x10\n" |
| "bgt 2b\n" |
| "b 122f\n" |
| "31:" // Height 2 |
| "movi v11.4s, #0x0\n" |
| "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" |
| "mov x27, %x[col_bias]\n" |
| "movi v12.4s, #0x0\n" |
| "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" |
| "bic %x[flags], %x[flags], #0x80000000\n" |
| "movi v15.16b, #0x1\n" |
| "mov x26, %x[output_ptr]\n" |
| "32:" // 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" |
| "33:" // Height 2: setup done |
| "mov x25, #0x0\n" |
| "34:" // Height 2: String loop |
| "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" |
| "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" |
| "ldr w24, [x20, x25, LSL #0x2]\n" |
| "tbz %x[flags], #3, 35f\n" |
| "ldr x20, [%x[input_ptr], x25, LSL #0x3]\n" |
| "add x20, x20, x19, LSL #3\n" |
| "ldr x23, [x20, #0x0]\n" |
| "ldr x22, [x20, #0x8]\n" |
| "cbnz x25, 36f\n" |
| "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" |
| "add x23, x23, x19\n" |
| "add x22, x22, x19\n" |
| "b 36f\n" |
| "35:" // Height 2: setup direct input |
| "mov x23, %x[input_ptr]\n" |
| "add x22, x23, x19\n" |
| "36:" // Height 2: input setup done |
| "cmp x24, #0x10\n" |
| "blt 41f\n" |
| "ldr q0, [x23, #0x0]\n" |
| "ldr q1, [x22, #0x0]\n" |
| "cmp x24, #0x20\n" |
| "ldr q4, [x28, #0x0]\n" |
| "blt 39f\n" |
| "37:" // Height 2: Multiply loop: Main loop head |
| ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" |
| "ldr q5, [x28, #0x10]\n" |
| "add x23, x23, #0x10\n" |
| ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" |
| "ldr q6, [x28, #0x20]\n" |
| "add x22, x22, #0x10\n" |
| ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" |
| "ldr q7, [x28, #0x30]\n" |
| ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" |
| "ldr q8, [x28, #0x40]\n" |
| ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" |
| "ldr q9, [x28, #0x50]\n" |
| ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" |
| "ldr q10, [x28, #0x60]\n" |
| "ldr q4, [x28, #0x70]\n" |
| ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" |
| "ldr q5, [x28, #0x80]\n" |
| ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" |
| "ldr q6, [x28, #0x90]\n" |
| ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" |
| "ldr q7, [x28, #0xa0]\n" |
| ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" |
| "ldr q8, [x28, #0xb0]\n" |
| ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" |
| "ldr q9, [x28, #0xc0]\n" |
| ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" |
| "ldr q10, [x28, #0xd0]\n" |
| ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" |
| "ldr q4, [x28, #0xe0]\n" |
| ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" |
| "ldr q5, [x28, #0xf0]\n" |
| "add x28, x28, #0x100\n" |
| ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" |
| ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" |
| ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" |
| ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" |
| ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" |
| ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" |
| ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" |
| ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e8b7 // udot v23.4s, v5.16b, v1.4b[3]\n" |
| "tbnz %x[flags], #31, 38f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| "38:" // Height 2: Multiply loop: unique 5: skip row sum |
| "prfm pldl1keep, [x23, #0x80]\n" |
| "sub x24, x24, #0x10\n" |
| "prfm pldl1keep, [x22, #0x80]\n" |
| "cmp x24, #0x20\n" |
| "ldr q0, [x23, #0x0]\n" |
| "ldr q1, [x22, #0x0]\n" |
| "ldr q4, [x28, #0x0]\n" |
| "bge 37b\n" |
| "39:" // Height 2: Multiply loop: Single iteration only |
| ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" |
| "ldr q5, [x28, #0x10]\n" |
| "sub x24, x24, #0x10\n" |
| ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" |
| "ldr q6, [x28, #0x20]\n" |
| "add x23, x23, #0x10\n" |
| ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" |
| "ldr q7, [x28, #0x30]\n" |
| "add x22, x22, #0x10\n" |
| ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" |
| "ldr q8, [x28, #0x40]\n" |
| ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" |
| "ldr q9, [x28, #0x50]\n" |
| ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" |
| "ldr q10, [x28, #0x60]\n" |
| "ldr q4, [x28, #0x70]\n" |
| ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" |
| "ldr q5, [x28, #0x80]\n" |
| ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" |
| "ldr q6, [x28, #0x90]\n" |
| ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" |
| "ldr q7, [x28, #0xa0]\n" |
| ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" |
| "ldr q8, [x28, #0xb0]\n" |
| ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" |
| "ldr q9, [x28, #0xc0]\n" |
| ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" |
| "ldr q10, [x28, #0xd0]\n" |
| ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" |
| "ldr q4, [x28, #0xe0]\n" |
| ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" |
| "ldr q5, [x28, #0xf0]\n" |
| "add x28, x28, #0x100\n" |
| ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" |
| ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" |
| ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" |
| ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" |
| ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" |
| ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" |
| ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" |
| ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e8b7 // udot v23.4s, v5.16b, v1.4b[3]\n" |
| "tbnz %x[flags], #31, 40f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| "40:" // Height 2: Multiply loop: unique 6: skip row sum |
| "prfm pldl1keep, [x23, #0x80]\n" |
| "prfm pldl1keep, [x22, #0x80]\n" |
| "41:" // Height 2: Multiply loop: Main loop skip |
| "cbz x24, 48f\n" |
| "cmp x24, #0x4\n" |
| "blt 44f\n" |
| "42:" // Height 2: Multiply loop: Odd block loop |
| "ldr s0, [x23], #0x4\n" |
| "ldr s1, [x22], #0x4\n" |
| "tbnz %x[flags], #31, 43f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| "43:" // Height 2: Multiply loop: unique 7: skip row sum |
| "ldr q6, [x28, #0x0]\n" |
| ".inst 0x6f80e0d0 // udot v16.4s, v6.16b, v0.4b[0]\n" |
| "ldr q7, [x28, #0x10]\n" |
| "sub x24, x24, #0x4\n" |
| ".inst 0x6f81e0d4 // udot v20.4s, v6.16b, v1.4b[0]\n" |
| "ldr q8, [x28, #0x20]\n" |
| "cmp x24, #0x4\n" |
| ".inst 0x6f80e0f1 // udot v17.4s, v7.16b, v0.4b[0]\n" |
| "ldr q9, [x28, #0x30]\n" |
| "add x28, x28, #0x40\n" |
| ".inst 0x6f81e0f5 // udot v21.4s, v7.16b, v1.4b[0]\n" |
| ".inst 0x6f80e112 // udot v18.4s, v8.16b, v0.4b[0]\n" |
| ".inst 0x6f81e116 // udot v22.4s, v8.16b, v1.4b[0]\n" |
| ".inst 0x6f80e133 // udot v19.4s, v9.16b, v0.4b[0]\n" |
| ".inst 0x6f81e137 // udot v23.4s, v9.16b, v1.4b[0]\n" |
| "bge 42b\n" |
| "cbz x24, 48f\n" |
| "44:" // Height 2: Multiply loop: Skip odd blocks |
| "tbz x24, #1, 45f\n" |
| "ldr h0, [x23], #0x2\n" |
| "ldr h1, [x22], #0x2\n" |
| "tbz x24, #0, 46f\n" |
| "ld1 { v0.b }[2], [x23]\n" |
| "ld1 { v1.b }[2], [x22]\n" |
| "b 46f\n" |
| "45:" // Height 2: Multiply loop: Ragged operand read: partial_1_0 |
| "ldr b0, [x23, #0x0]\n" |
| "ldr b1, [x22, #0x0]\n" |
| "46:" // Height 2: Multiply loop: Ragged operand read: Done |
| "tbnz %x[flags], #31, 47f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| "47:" // Height 2: Multiply loop: unique 8: skip row sum |
| "ldr q10, [x28, #0x0]\n" |
| ".inst 0x6f80e150 // udot v16.4s, v10.16b, v0.4b[0]\n" |
| "ldr q4, [x28, #0x10]\n" |
| ".inst 0x6f81e154 // udot v20.4s, v10.16b, v1.4b[0]\n" |
| "ldr q5, [x28, #0x20]\n" |
| "ldr q6, [x28, #0x30]\n" |
| ".inst 0x6f80e091 // udot v17.4s, v4.16b, v0.4b[0]\n" |
| "add x28, x28, #0x40\n" |
| ".inst 0x6f81e095 // udot v21.4s, v4.16b, v1.4b[0]\n" |
| ".inst 0x6f80e0b2 // udot v18.4s, v5.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0b6 // udot v22.4s, v5.16b, v1.4b[0]\n" |
| ".inst 0x6f80e0d3 // udot v19.4s, v6.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0d7 // udot v23.4s, v6.16b, v1.4b[0]\n" |
| "48:" // Height 2: Multiply loop: No odd multiplies |
| "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" |
| "add x25, x25, #0x1\n" |
| "cmp x25, x19\n" |
| "bne 34b\n" |
| "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" |
| "prfm pstl1keep, [x26, #0x0]\n" |
| "add x21, x26, x19\n" |
| "prfm pstl1keep, [x21, #0x0]\n" |
| "tbnz %x[flags], #31, 49f\n" |
| "addp v11.4s, v11.4s, v11.4s\n" |
| "add x22, %x[qp], %[b_offset]\n" |
| "ld1r { v2.4s }, [x22]\n" |
| "addp v12.4s, v12.4s, v12.4s\n" |
| "addp v11.4s, v11.4s, v11.4s\n" |
| "neg v2.4s, v2.4s\n" |
| "addp v12.4s, v12.4s, v12.4s\n" |
| "mul v11.4s, v11.4s, v2.4s\n" |
| "mul v12.4s, v12.4s, v2.4s\n" |
| "49:" // Height 2: skip row sum fixup |
| "add v16.4s, v16.4s, v11.4s\n" |
| "ldr q0, [x27, #0x0]\n" |
| "orr %x[flags], %x[flags], #0x80000000\n" |
| "add v17.4s, v17.4s, v11.4s\n" |
| "ldr q1, [x27, #0x10]\n" |
| "add x23, %x[qp], %[per_layer_right_shift]\n" |
| "add v18.4s, v18.4s, v11.4s\n" |
| "ldr q2, [x27, #0x20]\n" |
| "add x22, %x[qp], %[per_layer_mul]\n" |
| "add v19.4s, v19.4s, v11.4s\n" |
| "ldr q3, [x27, #0x30]\n" |
| "add x27, x27, #0x40\n" |
| "add v20.4s, v20.4s, v12.4s\n" |
| "ld1r { v4.4s }, [x22]\n" |
| "add v21.4s, v21.4s, v12.4s\n" |
| "add v22.4s, v22.4s, v12.4s\n" |
| "add v23.4s, v23.4s, v12.4s\n" |
| "add v16.4s, v16.4s, v0.4s\n" |
| "add v17.4s, v17.4s, v1.4s\n" |
| "add v18.4s, v18.4s, v2.4s\n" |
| "add v19.4s, v19.4s, v3.4s\n" |
| "add v20.4s, v20.4s, v0.4s\n" |
| "ld1r { v0.4s }, [x23]\n" |
| "add v21.4s, v21.4s, v1.4s\n" |
| "add v22.4s, v22.4s, v2.4s\n" |
| "add v23.4s, v23.4s, v3.4s\n" |
| "sqrdmulh v16.4s, v16.4s, v4.4s\n" |
| "sqrdmulh v17.4s, v17.4s, v4.4s\n" |
| "sqrdmulh v18.4s, v18.4s, v4.4s\n" |
| "sqrdmulh v19.4s, v19.4s, v4.4s\n" |
| "sqrdmulh v20.4s, v20.4s, v4.4s\n" |
| "sqrdmulh v21.4s, v21.4s, v4.4s\n" |
| "sqrdmulh v22.4s, v22.4s, v4.4s\n" |
| "sqrdmulh v23.4s, v23.4s, v4.4s\n" |
| "tbz %x[flags], #5, 50f\n" |
| "and v4.16b, v16.16b, v0.16b\n" |
| "and v5.16b, v17.16b, v0.16b\n" |
| "and v6.16b, v18.16b, v0.16b\n" |
| "sshr v4.4s, v4.4s, #0x1f\n" |
| "sshr v5.4s, v5.4s, #0x1f\n" |
| "sshr v6.4s, v6.4s, #0x1f\n" |
| "sqadd v16.4s, v16.4s, v4.4s\n" |
| "sqadd v17.4s, v17.4s, v5.4s\n" |
| "sqadd v18.4s, v18.4s, v6.4s\n" |
| "and v7.16b, v19.16b, v0.16b\n" |
| "and v8.16b, v20.16b, v0.16b\n" |
| "and v9.16b, v21.16b, v0.16b\n" |
| "sshr v7.4s, v7.4s, #0x1f\n" |
| "sshr v8.4s, v8.4s, #0x1f\n" |
| "sshr v9.4s, v9.4s, #0x1f\n" |
| "sqadd v19.4s, v19.4s, v7.4s\n" |
| "sqadd v20.4s, v20.4s, v8.4s\n" |
| "sqadd v21.4s, v21.4s, v9.4s\n" |
| "and v10.16b, v22.16b, v0.16b\n" |
| "and v4.16b, v23.16b, v0.16b\n" |
| "sshr v10.4s, v10.4s, #0x1f\n" |
| "sshr v4.4s, v4.4s, #0x1f\n" |
| "sqadd v22.4s, v22.4s, v10.4s\n" |
| "sqadd v23.4s, v23.4s, v4.4s\n" |
| "50:" // Height 2: no shift correction |
| "srshl v16.4s, v16.4s, v0.4s\n" |
| "add x22, %x[qp], %[c_offset]\n" |
| "ld1r { v4.4s }, [x22]\n" |
| "srshl v17.4s, v17.4s, v0.4s\n" |
| "add x22, %x[qp], %[minval]\n" |
| "srshl v18.4s, v18.4s, v0.4s\n" |
| "ld1r { v5.4s }, [x22]\n" |
| "add x22, %x[qp], %[maxval]\n" |
| "srshl v19.4s, v19.4s, v0.4s\n" |
| "ld1r { v6.4s }, [x22]\n" |
| "cmp x9, #0x10\n" |
| "srshl v20.4s, v20.4s, v0.4s\n" |
| "srshl v21.4s, v21.4s, v0.4s\n" |
| "add v16.4s, v16.4s, v4.4s\n" |
| "add v17.4s, v17.4s, v4.4s\n" |
| "add v18.4s, v18.4s, v4.4s\n" |
| "smin v16.4s, v16.4s, v6.4s\n" |
| "smin v17.4s, v17.4s, v6.4s\n" |
| "smin v18.4s, v18.4s, v6.4s\n" |
| "smax v16.4s, v16.4s, v5.4s\n" |
| "smax v17.4s, v17.4s, v5.4s\n" |
| "smax v18.4s, v18.4s, v5.4s\n" |
| "add v19.4s, v19.4s, v4.4s\n" |
| "add v20.4s, v20.4s, v4.4s\n" |
| "add v21.4s, v21.4s, v4.4s\n" |
| "smin v19.4s, v19.4s, v6.4s\n" |
| "smin v20.4s, v20.4s, v6.4s\n" |
| "smin v21.4s, v21.4s, v6.4s\n" |
| "smax v19.4s, v19.4s, v5.4s\n" |
| "smax v20.4s, v20.4s, v5.4s\n" |
| "smax v21.4s, v21.4s, v5.4s\n" |
| "srshl v22.4s, v22.4s, v0.4s\n" |
| "srshl v23.4s, v23.4s, v0.4s\n" |
| "uzp1 v16.8h, v16.8h, v17.8h\n" |
| "uzp1 v17.8h, v18.8h, v19.8h\n" |
| "add v22.4s, v22.4s, v4.4s\n" |
| "add v23.4s, v23.4s, v4.4s\n" |
| "uzp1 v20.8h, v20.8h, v21.8h\n" |
| "smin v22.4s, v22.4s, v6.4s\n" |
| "smin v23.4s, v23.4s, v6.4s\n" |
| "uzp1 v16.16b, v16.16b, v17.16b\n" |
| "smax v22.4s, v22.4s, v5.4s\n" |
| "smax v23.4s, v23.4s, v5.4s\n" |
| "uzp1 v21.8h, v22.8h, v23.8h\n" |
| "uzp1 v20.16b, v20.16b, v21.16b\n" |
| "bge 59f\n" |
| "tbz x9, #3, 54f\n" |
| "str d16, [x26], #0x8\n" |
| "str d20, [x21], #0x8\n" |
| "tbz x9, #2, 52f\n" |
| "st1 { v16.s }[2], [x26], #0x4\n" |
| "st1 { v20.s }[2], [x21], #0x4\n" |
| "tbz x9, #1, 51f\n" |
| "st1 { v16.h }[6], [x26], #0x2\n" |
| "st1 { v20.h }[6], [x21], #0x2\n" |
| "tbz x9, #0, 58f\n" |
| "st1 { v16.b }[14], [x26]\n" |
| "st1 { v20.b }[14], [x21]\n" |
| "b 58f\n" |
| "51:" // Height 2: Partial direct writeback: partial_1_12 |
| "tbz x9, #0, 58f\n" |
| "st1 { v16.b }[12], [x26]\n" |
| "st1 { v20.b }[12], [x21]\n" |
| "b 58f\n" |
| "52:" // Height 2: Partial direct writeback: partial_2_8 |
| "tbz x9, #1, 53f\n" |
| "st1 { v16.h }[4], [x26], #0x2\n" |
| "st1 { v20.h }[4], [x21], #0x2\n" |
| "tbz x9, #0, 58f\n" |
| "st1 { v16.b }[10], [x26]\n" |
| "st1 { v20.b }[10], [x21]\n" |
| "b 58f\n" |
| "53:" // Height 2: Partial direct writeback: partial_1_8 |
| "tbz x9, #0, 58f\n" |
| "st1 { v16.b }[8], [x26]\n" |
| "st1 { v20.b }[8], [x21]\n" |
| "b 58f\n" |
| "54:" // Height 2: Partial direct writeback: partial_4_0 |
| "tbz x9, #2, 56f\n" |
| "str s16, [x26], #0x4\n" |
| "str s20, [x21], #0x4\n" |
| "tbz x9, #1, 55f\n" |
| "st1 { v16.h }[2], [x26], #0x2\n" |
| "st1 { v20.h }[2], [x21], #0x2\n" |
| "tbz x9, #0, 58f\n" |
| "st1 { v16.b }[6], [x26]\n" |
| "st1 { v20.b }[6], [x21]\n" |
| "b 58f\n" |
| "55:" // Height 2: Partial direct writeback: partial_1_4 |
| "tbz x9, #0, 58f\n" |
| "st1 { v16.b }[4], [x26]\n" |
| "st1 { v20.b }[4], [x21]\n" |
| "b 58f\n" |
| "56:" // Height 2: Partial direct writeback: partial_2_0 |
| "tbz x9, #1, 57f\n" |
| "str h16, [x26], #0x2\n" |
| "str h20, [x21], #0x2\n" |
| "tbz x9, #0, 58f\n" |
| "st1 { v16.b }[2], [x26]\n" |
| "st1 { v20.b }[2], [x21]\n" |
| "b 58f\n" |
| "57:" // Height 2: Partial direct writeback: partial_1_0 |
| "str b16, [x26, #0x0]\n" |
| "str b20, [x21, #0x0]\n" |
| "58:" // Height 2: Partial direct writeback: Done |
| "b 60f\n" |
| "59:" // Height 2: Full writeback |
| "str q16, [x26, #0x0]\n" |
| "add x26, x26, #0x10\n" |
| "str q20, [x21, #0x0]\n" |
| "60:" // Height 2: Writeback done |
| "subs x9, x9, #0x10\n" |
| "bgt 32b\n" |
| "b 122f\n" |
| "61:" // Height 3 |
| "movi v11.4s, #0x0\n" |
| "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" |
| "mov x27, %x[col_bias]\n" |
| "movi v12.4s, #0x0\n" |
| "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" |
| "bic %x[flags], %x[flags], #0x80000000\n" |
| "movi v13.4s, #0x0\n" |
| "mov x26, %x[output_ptr]\n" |
| "movi v15.16b, #0x1\n" |
| "62:" // 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" |
| "63:" // Height 3: setup done |
| "mov x25, #0x0\n" |
| "64:" // Height 3: String loop |
| "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" |
| "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" |
| "ldr w24, [x20, x25, LSL #0x2]\n" |
| "tbz %x[flags], #3, 65f\n" |
| "ldr x20, [%x[input_ptr], x25, LSL #0x3]\n" |
| "add x20, x20, x19, LSL #3\n" |
| "ldr x23, [x20, #0x0]\n" |
| "ldr x22, [x20, #0x8]\n" |
| "ldr x21, [x20, #0x10]\n" |
| "cbnz x25, 66f\n" |
| "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" |
| "add x23, x23, x19\n" |
| "add x22, x22, x19\n" |
| "add x21, x21, x19\n" |
| "b 66f\n" |
| "65:" // Height 3: setup direct input |
| "mov x23, %x[input_ptr]\n" |
| "add x22, x23, x19\n" |
| "add x21, x22, x19\n" |
| "66:" // Height 3: input setup done |
| "cmp x24, #0x10\n" |
| "blt 71f\n" |
| "ldr q0, [x23, #0x0]\n" |
| "ldr q1, [x22, #0x0]\n" |
| "cmp x24, #0x20\n" |
| "ldr q2, [x21, #0x0]\n" |
| "ldr q4, [x28, #0x0]\n" |
| "blt 69f\n" |
| "67:" // Height 3: Multiply loop: Main loop head |
| ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" |
| "ldr q5, [x28, #0x10]\n" |
| "add x23, x23, #0x10\n" |
| ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" |
| "ldr q6, [x28, #0x20]\n" |
| "add x22, x22, #0x10\n" |
| ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" |
| "ldr q7, [x28, #0x30]\n" |
| "add x21, x21, #0x10\n" |
| ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" |
| "ldr q8, [x28, #0x40]\n" |
| ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" |
| "ldr q9, [x28, #0x50]\n" |
| ".inst 0x6f82e0b9 // udot v25.4s, v5.16b, v2.4b[0]\n" |
| "ldr q10, [x28, #0x60]\n" |
| ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" |
| "ldr q4, [x28, #0x70]\n" |
| ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" |
| "ldr q5, [x28, #0x80]\n" |
| ".inst 0x6f82e0da // udot v26.4s, v6.16b, v2.4b[0]\n" |
| "ldr q6, [x28, #0x90]\n" |
| ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0fb // udot v27.4s, v7.16b, v2.4b[0]\n" |
| "ldr q7, [x28, #0xa0]\n" |
| ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e118 // udot v24.4s, v8.16b, v2.4b[1]\n" |
| "ldr q8, [x28, #0xb0]\n" |
| ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e139 // udot v25.4s, v9.16b, v2.4b[1]\n" |
| "ldr q9, [x28, #0xc0]\n" |
| ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e15a // udot v26.4s, v10.16b, v2.4b[1]\n" |
| "ldr q10, [x28, #0xd0]\n" |
| ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e09b // udot v27.4s, v4.16b, v2.4b[1]\n" |
| "ldr q4, [x28, #0xe0]\n" |
| ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8b8 // udot v24.4s, v5.16b, v2.4b[2]\n" |
| "ldr q5, [x28, #0xf0]\n" |
| "add x28, x28, #0x100\n" |
| ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8d9 // udot v25.4s, v6.16b, v2.4b[2]\n" |
| ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8fa // udot v26.4s, v7.16b, v2.4b[2]\n" |
| ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" |
| ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" |
| ".inst 0x6f82e91b // udot v27.4s, v8.16b, v2.4b[2]\n" |
| ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e938 // udot v24.4s, v9.16b, v2.4b[3]\n" |
| ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e959 // udot v25.4s, v10.16b, v2.4b[3]\n" |
| ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e89a // udot v26.4s, v4.16b, v2.4b[3]\n" |
| ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e8b7 // udot v23.4s, v5.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e8bb // udot v27.4s, v5.16b, v2.4b[3]\n" |
| "tbnz %x[flags], #31, 68f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" |
| "68:" // Height 3: Multiply loop: unique 9: skip row sum |
| "prfm pldl1keep, [x23, #0x80]\n" |
| "sub x24, x24, #0x10\n" |
| "prfm pldl1keep, [x22, #0x80]\n" |
| "cmp x24, #0x20\n" |
| "prfm pldl1keep, [x21, #0x80]\n" |
| "ldr q0, [x23, #0x0]\n" |
| "ldr q1, [x22, #0x0]\n" |
| "ldr q2, [x21, #0x0]\n" |
| "ldr q4, [x28, #0x0]\n" |
| "bge 67b\n" |
| "69:" // Height 3: Multiply loop: Single iteration only |
| ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" |
| "ldr q5, [x28, #0x10]\n" |
| "sub x24, x24, #0x10\n" |
| ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" |
| "ldr q6, [x28, #0x20]\n" |
| "add x23, x23, #0x10\n" |
| ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" |
| "ldr q7, [x28, #0x30]\n" |
| "add x22, x22, #0x10\n" |
| ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" |
| "ldr q8, [x28, #0x40]\n" |
| "add x21, x21, #0x10\n" |
| ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" |
| "ldr q9, [x28, #0x50]\n" |
| ".inst 0x6f82e0b9 // udot v25.4s, v5.16b, v2.4b[0]\n" |
| "ldr q10, [x28, #0x60]\n" |
| ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" |
| "ldr q4, [x28, #0x70]\n" |
| ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" |
| "ldr q5, [x28, #0x80]\n" |
| ".inst 0x6f82e0da // udot v26.4s, v6.16b, v2.4b[0]\n" |
| "ldr q6, [x28, #0x90]\n" |
| ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0fb // udot v27.4s, v7.16b, v2.4b[0]\n" |
| "ldr q7, [x28, #0xa0]\n" |
| ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e118 // udot v24.4s, v8.16b, v2.4b[1]\n" |
| "ldr q8, [x28, #0xb0]\n" |
| ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e139 // udot v25.4s, v9.16b, v2.4b[1]\n" |
| "ldr q9, [x28, #0xc0]\n" |
| ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e15a // udot v26.4s, v10.16b, v2.4b[1]\n" |
| "ldr q10, [x28, #0xd0]\n" |
| ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e09b // udot v27.4s, v4.16b, v2.4b[1]\n" |
| "ldr q4, [x28, #0xe0]\n" |
| ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8b8 // udot v24.4s, v5.16b, v2.4b[2]\n" |
| "ldr q5, [x28, #0xf0]\n" |
| "add x28, x28, #0x100\n" |
| ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8d9 // udot v25.4s, v6.16b, v2.4b[2]\n" |
| ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8fa // udot v26.4s, v7.16b, v2.4b[2]\n" |
| ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" |
| ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" |
| ".inst 0x6f82e91b // udot v27.4s, v8.16b, v2.4b[2]\n" |
| ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e938 // udot v24.4s, v9.16b, v2.4b[3]\n" |
| ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e959 // udot v25.4s, v10.16b, v2.4b[3]\n" |
| ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e89a // udot v26.4s, v4.16b, v2.4b[3]\n" |
| ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e8b7 // udot v23.4s, v5.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e8bb // udot v27.4s, v5.16b, v2.4b[3]\n" |
| "tbnz %x[flags], #31, 70f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" |
| "70:" // Height 3: Multiply loop: unique 10: skip row sum |
| "prfm pldl1keep, [x23, #0x80]\n" |
| "prfm pldl1keep, [x22, #0x80]\n" |
| "prfm pldl1keep, [x21, #0x80]\n" |
| "71:" // Height 3: Multiply loop: Main loop skip |
| "cbz x24, 78f\n" |
| "cmp x24, #0x4\n" |
| "blt 74f\n" |
| "72:" // Height 3: Multiply loop: Odd block loop |
| "ldr s0, [x23], #0x4\n" |
| "ldr s1, [x22], #0x4\n" |
| "ldr s2, [x21], #0x4\n" |
| "tbnz %x[flags], #31, 73f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" |
| "73:" // Height 3: Multiply loop: unique 11: skip row sum |
| "ldr q6, [x28, #0x0]\n" |
| ".inst 0x6f80e0d0 // udot v16.4s, v6.16b, v0.4b[0]\n" |
| "ldr q7, [x28, #0x10]\n" |
| "sub x24, x24, #0x4\n" |
| ".inst 0x6f81e0d4 // udot v20.4s, v6.16b, v1.4b[0]\n" |
| "ldr q8, [x28, #0x20]\n" |
| "cmp x24, #0x4\n" |
| ".inst 0x6f82e0d8 // udot v24.4s, v6.16b, v2.4b[0]\n" |
| "ldr q9, [x28, #0x30]\n" |
| "add x28, x28, #0x40\n" |
| ".inst 0x6f80e0f1 // udot v17.4s, v7.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0f5 // udot v21.4s, v7.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0f9 // udot v25.4s, v7.16b, v2.4b[0]\n" |
| ".inst 0x6f80e112 // udot v18.4s, v8.16b, v0.4b[0]\n" |
| ".inst 0x6f81e116 // udot v22.4s, v8.16b, v1.4b[0]\n" |
| ".inst 0x6f82e11a // udot v26.4s, v8.16b, v2.4b[0]\n" |
| ".inst 0x6f80e133 // udot v19.4s, v9.16b, v0.4b[0]\n" |
| ".inst 0x6f81e137 // udot v23.4s, v9.16b, v1.4b[0]\n" |
| ".inst 0x6f82e13b // udot v27.4s, v9.16b, v2.4b[0]\n" |
| "bge 72b\n" |
| "cbz x24, 78f\n" |
| "74:" // Height 3: Multiply loop: Skip odd blocks |
| "tbz x24, #1, 75f\n" |
| "ldr h0, [x23], #0x2\n" |
| "ldr h1, [x22], #0x2\n" |
| "ldr h2, [x21], #0x2\n" |
| "tbz x24, #0, 76f\n" |
| "ld1 { v0.b }[2], [x23]\n" |
| "ld1 { v1.b }[2], [x22]\n" |
| "ld1 { v2.b }[2], [x21]\n" |
| "b 76f\n" |
| "75:" // Height 3: Multiply loop: Ragged operand read: partial_1_0 |
| "ldr b0, [x23, #0x0]\n" |
| "ldr b1, [x22, #0x0]\n" |
| "ldr b2, [x21, #0x0]\n" |
| "76:" // Height 3: Multiply loop: Ragged operand read: Done |
| "tbnz %x[flags], #31, 77f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" |
| "77:" // Height 3: Multiply loop: unique 12: skip row sum |
| "ldr q10, [x28, #0x0]\n" |
| ".inst 0x6f80e150 // udot v16.4s, v10.16b, v0.4b[0]\n" |
| "ldr q4, [x28, #0x10]\n" |
| ".inst 0x6f81e154 // udot v20.4s, v10.16b, v1.4b[0]\n" |
| "ldr q5, [x28, #0x20]\n" |
| ".inst 0x6f82e158 // udot v24.4s, v10.16b, v2.4b[0]\n" |
| "ldr q6, [x28, #0x30]\n" |
| "add x28, x28, #0x40\n" |
| ".inst 0x6f80e091 // udot v17.4s, v4.16b, v0.4b[0]\n" |
| ".inst 0x6f81e095 // udot v21.4s, v4.16b, v1.4b[0]\n" |
| ".inst 0x6f82e099 // udot v25.4s, v4.16b, v2.4b[0]\n" |
| ".inst 0x6f80e0b2 // udot v18.4s, v5.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0b6 // udot v22.4s, v5.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0ba // udot v26.4s, v5.16b, v2.4b[0]\n" |
| ".inst 0x6f80e0d3 // udot v19.4s, v6.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0d7 // udot v23.4s, v6.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0db // udot v27.4s, v6.16b, v2.4b[0]\n" |
| "78:" // Height 3: Multiply loop: No odd multiplies |
| "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" |
| "add x25, x25, #0x1\n" |
| "cmp x25, x19\n" |
| "bne 64b\n" |
| "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" |
| "prfm pstl1keep, [x26, #0x0]\n" |
| "add x21, x26, x19\n" |
| "prfm pstl1keep, [x21, #0x0]\n" |
| "add x20, x21, x19\n" |
| "prfm pstl1keep, [x20, #0x0]\n" |
| "tbnz %x[flags], #31, 79f\n" |
| "addp v11.4s, v11.4s, v11.4s\n" |
| "add x22, %x[qp], %[b_offset]\n" |
| "ld1r { v3.4s }, [x22]\n" |
| "addp v12.4s, v12.4s, v12.4s\n" |
| "addp v13.4s, v13.4s, v13.4s\n" |
| "addp v11.4s, v11.4s, v11.4s\n" |
| "neg v3.4s, v3.4s\n" |
| "addp v12.4s, v12.4s, v12.4s\n" |
| "addp v13.4s, v13.4s, v13.4s\n" |
| "mul v11.4s, v11.4s, v3.4s\n" |
| "mul v12.4s, v12.4s, v3.4s\n" |
| "mul v13.4s, v13.4s, v3.4s\n" |
| "79:" // Height 3: skip row sum fixup |
| "add v16.4s, v16.4s, v11.4s\n" |
| "ldr q0, [x27, #0x0]\n" |
| "orr %x[flags], %x[flags], #0x80000000\n" |
| "add v17.4s, v17.4s, v11.4s\n" |
| "ldr q1, [x27, #0x10]\n" |
| "add x23, %x[qp], %[per_layer_right_shift]\n" |
| "add v18.4s, v18.4s, v11.4s\n" |
| "ldr q2, [x27, #0x20]\n" |
| "add x22, %x[qp], %[per_layer_mul]\n" |
| "add v19.4s, v19.4s, v11.4s\n" |
| "ldr q3, [x27, #0x30]\n" |
| "add x27, x27, #0x40\n" |
| "add v20.4s, v20.4s, v12.4s\n" |
| "ld1r { v4.4s }, [x22]\n" |
| "add v21.4s, v21.4s, v12.4s\n" |
| "add v22.4s, v22.4s, v12.4s\n" |
| "add v23.4s, v23.4s, v12.4s\n" |
| "add v24.4s, v24.4s, v13.4s\n" |
| "add v25.4s, v25.4s, v13.4s\n" |
| "add v26.4s, v26.4s, v13.4s\n" |
| "add v27.4s, v27.4s, v13.4s\n" |
| "add v16.4s, v16.4s, v0.4s\n" |
| "add v17.4s, v17.4s, v1.4s\n" |
| "add v18.4s, v18.4s, v2.4s\n" |
| "add v19.4s, v19.4s, v3.4s\n" |
| "add v20.4s, v20.4s, v0.4s\n" |
| "add v21.4s, v21.4s, v1.4s\n" |
| "add v22.4s, v22.4s, v2.4s\n" |
| "add v23.4s, v23.4s, v3.4s\n" |
| "add v24.4s, v24.4s, v0.4s\n" |
| "ld1r { v0.4s }, [x23]\n" |
| "add v25.4s, v25.4s, v1.4s\n" |
| "add v26.4s, v26.4s, v2.4s\n" |
| "add v27.4s, v27.4s, v3.4s\n" |
| "sqrdmulh v16.4s, v16.4s, v4.4s\n" |
| "sqrdmulh v17.4s, v17.4s, v4.4s\n" |
| "sqrdmulh v18.4s, v18.4s, v4.4s\n" |
| "sqrdmulh v19.4s, v19.4s, v4.4s\n" |
| "sqrdmulh v20.4s, v20.4s, v4.4s\n" |
| "sqrdmulh v21.4s, v21.4s, v4.4s\n" |
| "sqrdmulh v22.4s, v22.4s, v4.4s\n" |
| "sqrdmulh v23.4s, v23.4s, v4.4s\n" |
| "sqrdmulh v24.4s, v24.4s, v4.4s\n" |
| "sqrdmulh v25.4s, v25.4s, v4.4s\n" |
| "sqrdmulh v26.4s, v26.4s, v4.4s\n" |
| "sqrdmulh v27.4s, v27.4s, v4.4s\n" |
| "tbz %x[flags], #5, 80f\n" |
| "and v4.16b, v16.16b, v0.16b\n" |
| "and v5.16b, v17.16b, v0.16b\n" |
| "and v6.16b, v18.16b, v0.16b\n" |
| "sshr v4.4s, v4.4s, #0x1f\n" |
| "sshr v5.4s, v5.4s, #0x1f\n" |
| "sshr v6.4s, v6.4s, #0x1f\n" |
| "sqadd v16.4s, v16.4s, v4.4s\n" |
| "sqadd v17.4s, v17.4s, v5.4s\n" |
| "sqadd v18.4s, v18.4s, v6.4s\n" |
| "and v7.16b, v19.16b, v0.16b\n" |
| "and v8.16b, v20.16b, v0.16b\n" |
| "and v9.16b, v21.16b, v0.16b\n" |
| "sshr v7.4s, v7.4s, #0x1f\n" |
| "sshr v8.4s, v8.4s, #0x1f\n" |
| "sshr v9.4s, v9.4s, #0x1f\n" |
| "sqadd v19.4s, v19.4s, v7.4s\n" |
| "sqadd v20.4s, v20.4s, v8.4s\n" |
| "sqadd v21.4s, v21.4s, v9.4s\n" |
| "and v10.16b, v22.16b, v0.16b\n" |
| "and v4.16b, v23.16b, v0.16b\n" |
| "and v5.16b, v24.16b, v0.16b\n" |
| "sshr v10.4s, v10.4s, #0x1f\n" |
| "sshr v4.4s, v4.4s, #0x1f\n" |
| "sshr v5.4s, v5.4s, #0x1f\n" |
| "sqadd v22.4s, v22.4s, v10.4s\n" |
| "sqadd v23.4s, v23.4s, v4.4s\n" |
| "sqadd v24.4s, v24.4s, v5.4s\n" |
| "and v6.16b, v25.16b, v0.16b\n" |
| "and v7.16b, v26.16b, v0.16b\n" |
| "and v8.16b, v27.16b, v0.16b\n" |
| "sshr v6.4s, v6.4s, #0x1f\n" |
| "sshr v7.4s, v7.4s, #0x1f\n" |
| "sshr v8.4s, v8.4s, #0x1f\n" |
| "sqadd v25.4s, v25.4s, v6.4s\n" |
| "sqadd v26.4s, v26.4s, v7.4s\n" |
| "sqadd v27.4s, v27.4s, v8.4s\n" |
| "80:" // Height 3: no shift correction |
| "srshl v16.4s, v16.4s, v0.4s\n" |
| "add x22, %x[qp], %[c_offset]\n" |
| "ld1r { v4.4s }, [x22]\n" |
| "srshl v17.4s, v17.4s, v0.4s\n" |
| "add x22, %x[qp], %[minval]\n" |
| "srshl v18.4s, v18.4s, v0.4s\n" |
| "ld1r { v5.4s }, [x22]\n" |
| "add x22, %x[qp], %[maxval]\n" |
| "srshl v19.4s, v19.4s, v0.4s\n" |
| "ld1r { v6.4s }, [x22]\n" |
| "cmp x9, #0x10\n" |
| "srshl v20.4s, v20.4s, v0.4s\n" |
| "srshl v21.4s, v21.4s, v0.4s\n" |
| "add v16.4s, v16.4s, v4.4s\n" |
| "add v17.4s, v17.4s, v4.4s\n" |
| "add v18.4s, v18.4s, v4.4s\n" |
| "smin v16.4s, v16.4s, v6.4s\n" |
| "smin v17.4s, v17.4s, v6.4s\n" |
| "smin v18.4s, v18.4s, v6.4s\n" |
| "smax v16.4s, v16.4s, v5.4s\n" |
| "smax v17.4s, v17.4s, v5.4s\n" |
| "smax v18.4s, v18.4s, v5.4s\n" |
| "add v19.4s, v19.4s, v4.4s\n" |
| "add v20.4s, v20.4s, v4.4s\n" |
| "add v21.4s, v21.4s, v4.4s\n" |
| "smin v19.4s, v19.4s, v6.4s\n" |
| "smin v20.4s, v20.4s, v6.4s\n" |
| "smin v21.4s, v21.4s, v6.4s\n" |
| "smax v19.4s, v19.4s, v5.4s\n" |
| "smax v20.4s, v20.4s, v5.4s\n" |
| "smax v21.4s, v21.4s, v5.4s\n" |
| "srshl v22.4s, v22.4s, v0.4s\n" |
| "srshl v23.4s, v23.4s, v0.4s\n" |
| "srshl v24.4s, v24.4s, v0.4s\n" |
| "srshl v25.4s, v25.4s, v0.4s\n" |
| "add v22.4s, v22.4s, v4.4s\n" |
| "add v23.4s, v23.4s, v4.4s\n" |
| "add v24.4s, v24.4s, v4.4s\n" |
| "smin v22.4s, v22.4s, v6.4s\n" |
| "smin v23.4s, v23.4s, v6.4s\n" |
| "smin v24.4s, v24.4s, v6.4s\n" |
| "smax v22.4s, v22.4s, v5.4s\n" |
| "smax v23.4s, v23.4s, v5.4s\n" |
| "smax v24.4s, v24.4s, v5.4s\n" |
| "add v25.4s, v25.4s, v4.4s\n" |
| "srshl v26.4s, v26.4s, v0.4s\n" |
| "srshl v27.4s, v27.4s, v0.4s\n" |
| "smin v25.4s, v25.4s, v6.4s\n" |
| "uzp1 v16.8h, v16.8h, v17.8h\n" |
| "add v26.4s, v26.4s, v4.4s\n" |
| "smax v25.4s, v25.4s, v5.4s\n" |
| "add v27.4s, v27.4s, v4.4s\n" |
| "smin v26.4s, v26.4s, v6.4s\n" |
| "uzp1 v17.8h, v18.8h, v19.8h\n" |
| "smin v27.4s, v27.4s, v6.4s\n" |
| "smax v26.4s, v26.4s, v5.4s\n" |
| "uzp1 v20.8h, v20.8h, v21.8h\n" |
| "smax v27.4s, v27.4s, v5.4s\n" |
| "uzp1 v21.8h, v22.8h, v23.8h\n" |
| "uzp1 v24.8h, v24.8h, v25.8h\n" |
| "uzp1 v25.8h, v26.8h, v27.8h\n" |
| "uzp1 v16.16b, v16.16b, v17.16b\n" |
| "uzp1 v20.16b, v20.16b, v21.16b\n" |
| "uzp1 v24.16b, v24.16b, v25.16b\n" |
| "bge 89f\n" |
| "tbz x9, #3, 84f\n" |
| "str d16, [x26], #0x8\n" |
| "str d20, [x21], #0x8\n" |
| "str d24, [x20], #0x8\n" |
| "tbz x9, #2, 82f\n" |
| "st1 { v16.s }[2], [x26], #0x4\n" |
| "st1 { v20.s }[2], [x21], #0x4\n" |
| "st1 { v24.s }[2], [x20], #0x4\n" |
| "tbz x9, #1, 81f\n" |
| "st1 { v16.h }[6], [x26], #0x2\n" |
| "st1 { v20.h }[6], [x21], #0x2\n" |
| "st1 { v24.h }[6], [x20], #0x2\n" |
| "tbz x9, #0, 88f\n" |
| "st1 { v16.b }[14], [x26]\n" |
| "st1 { v20.b }[14], [x21]\n" |
| "st1 { v24.b }[14], [x20]\n" |
| "b 88f\n" |
| "81:" // Height 3: Partial direct writeback: partial_1_12 |
| "tbz x9, #0, 88f\n" |
| "st1 { v16.b }[12], [x26]\n" |
| "st1 { v20.b }[12], [x21]\n" |
| "st1 { v24.b }[12], [x20]\n" |
| "b 88f\n" |
| "82:" // Height 3: Partial direct writeback: partial_2_8 |
| "tbz x9, #1, 83f\n" |
| "st1 { v16.h }[4], [x26], #0x2\n" |
| "st1 { v20.h }[4], [x21], #0x2\n" |
| "st1 { v24.h }[4], [x20], #0x2\n" |
| "tbz x9, #0, 88f\n" |
| "st1 { v16.b }[10], [x26]\n" |
| "st1 { v20.b }[10], [x21]\n" |
| "st1 { v24.b }[10], [x20]\n" |
| "b 88f\n" |
| "83:" // Height 3: Partial direct writeback: partial_1_8 |
| "tbz x9, #0, 88f\n" |
| "st1 { v16.b }[8], [x26]\n" |
| "st1 { v20.b }[8], [x21]\n" |
| "st1 { v24.b }[8], [x20]\n" |
| "b 88f\n" |
| "84:" // Height 3: Partial direct writeback: partial_4_0 |
| "tbz x9, #2, 86f\n" |
| "str s16, [x26], #0x4\n" |
| "str s20, [x21], #0x4\n" |
| "str s24, [x20], #0x4\n" |
| "tbz x9, #1, 85f\n" |
| "st1 { v16.h }[2], [x26], #0x2\n" |
| "st1 { v20.h }[2], [x21], #0x2\n" |
| "st1 { v24.h }[2], [x20], #0x2\n" |
| "tbz x9, #0, 88f\n" |
| "st1 { v16.b }[6], [x26]\n" |
| "st1 { v20.b }[6], [x21]\n" |
| "st1 { v24.b }[6], [x20]\n" |
| "b 88f\n" |
| "85:" // Height 3: Partial direct writeback: partial_1_4 |
| "tbz x9, #0, 88f\n" |
| "st1 { v16.b }[4], [x26]\n" |
| "st1 { v20.b }[4], [x21]\n" |
| "st1 { v24.b }[4], [x20]\n" |
| "b 88f\n" |
| "86:" // Height 3: Partial direct writeback: partial_2_0 |
| "tbz x9, #1, 87f\n" |
| "str h16, [x26], #0x2\n" |
| "str h20, [x21], #0x2\n" |
| "str h24, [x20], #0x2\n" |
| "tbz x9, #0, 88f\n" |
| "st1 { v16.b }[2], [x26]\n" |
| "st1 { v20.b }[2], [x21]\n" |
| "st1 { v24.b }[2], [x20]\n" |
| "b 88f\n" |
| "87:" // Height 3: Partial direct writeback: partial_1_0 |
| "str b16, [x26, #0x0]\n" |
| "str b20, [x21, #0x0]\n" |
| "str b24, [x20, #0x0]\n" |
| "88:" // Height 3: Partial direct writeback: Done |
| "b 90f\n" |
| "89:" // Height 3: Full writeback |
| "str q16, [x26, #0x0]\n" |
| "add x26, x26, #0x10\n" |
| "str q20, [x21, #0x0]\n" |
| "str q24, [x20, #0x0]\n" |
| "90:" // Height 3: Writeback done |
| "subs x9, x9, #0x10\n" |
| "bgt 62b\n" |
| "b 122f\n" |
| "91:" // Height 4 |
| "movi v11.4s, #0x0\n" |
| "ldr x9, [%x[args_ptr], %[offsetof_N]]\n" |
| "mov x27, %x[col_bias]\n" |
| "movi v12.4s, #0x0\n" |
| "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n" |
| "bic %x[flags], %x[flags], #0x80000000\n" |
| "movi v13.4s, #0x0\n" |
| "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" |
| "mov x26, %x[output_ptr]\n" |
| "movi v14.4s, #0x0\n" |
| "mov x19, #0x4\n" |
| "movi v15.16b, #0x1\n" |
| "madd %x[output_ptr], x20, x19, %x[output_ptr]\n" |
| "92:" // 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" |
| "93:" // Height 4: setup done |
| "mov x25, #0x0\n" |
| "94:" // Height 4: String loop |
| "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" |
| "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" |
| "ldr w24, [x20, x25, LSL #0x2]\n" |
| "tbz %x[flags], #3, 95f\n" |
| "ldr x20, [%x[input_ptr], x25, LSL #0x3]\n" |
| "add x20, x20, x19, LSL #3\n" |
| "ldr x23, [x20, #0x0]\n" |
| "ldr x22, [x20, #0x8]\n" |
| "ldr x21, [x20, #0x10]\n" |
| "ldr x20, [x20, #0x18]\n" |
| "cbnz x25, 96f\n" |
| "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" |
| "add x23, x23, x19\n" |
| "add x22, x22, x19\n" |
| "add x21, x21, x19\n" |
| "add x20, x20, x19\n" |
| "b 96f\n" |
| "95:" // Height 4: setup direct input |
| "mov x23, %x[input_ptr]\n" |
| "add x22, x23, x19\n" |
| "add x21, x22, x19\n" |
| "add x20, x21, x19\n" |
| "96:" // Height 4: input setup done |
| "cmp x24, #0x10\n" |
| "blt 101f\n" |
| "ldr q0, [x23, #0x0]\n" |
| "ldr q1, [x22, #0x0]\n" |
| "cmp x24, #0x20\n" |
| "ldr q2, [x21, #0x0]\n" |
| "ldr q3, [x20, #0x0]\n" |
| "ldr q4, [x28, #0x0]\n" |
| "blt 99f\n" |
| "97:" // Height 4: Multiply loop: Main loop head |
| ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" |
| "ldr q5, [x28, #0x10]\n" |
| "add x23, x23, #0x10\n" |
| ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" |
| "ldr q6, [x28, #0x20]\n" |
| "add x22, x22, #0x10\n" |
| ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" |
| "ldr q7, [x28, #0x30]\n" |
| "add x21, x21, #0x10\n" |
| ".inst 0x6f83e09c // udot v28.4s, v4.16b, v3.4b[0]\n" |
| "ldr q8, [x28, #0x40]\n" |
| "add x20, x20, #0x10\n" |
| ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" |
| "ldr q9, [x28, #0x50]\n" |
| ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" |
| "ldr q10, [x28, #0x60]\n" |
| ".inst 0x6f82e0b9 // udot v25.4s, v5.16b, v2.4b[0]\n" |
| "ldr q4, [x28, #0x70]\n" |
| ".inst 0x6f83e0bd // udot v29.4s, v5.16b, v3.4b[0]\n" |
| "ldr q5, [x28, #0x80]\n" |
| ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0da // udot v26.4s, v6.16b, v2.4b[0]\n" |
| ".inst 0x6f83e0de // udot v30.4s, v6.16b, v3.4b[0]\n" |
| "ldr q6, [x28, #0x90]\n" |
| ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0fb // udot v27.4s, v7.16b, v2.4b[0]\n" |
| ".inst 0x6f83e0ff // udot v31.4s, v7.16b, v3.4b[0]\n" |
| "ldr q7, [x28, #0xa0]\n" |
| ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e118 // udot v24.4s, v8.16b, v2.4b[1]\n" |
| ".inst 0x6fa3e11c // udot v28.4s, v8.16b, v3.4b[1]\n" |
| "ldr q8, [x28, #0xb0]\n" |
| ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e139 // udot v25.4s, v9.16b, v2.4b[1]\n" |
| ".inst 0x6fa3e13d // udot v29.4s, v9.16b, v3.4b[1]\n" |
| "ldr q9, [x28, #0xc0]\n" |
| ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e15a // udot v26.4s, v10.16b, v2.4b[1]\n" |
| ".inst 0x6fa3e15e // udot v30.4s, v10.16b, v3.4b[1]\n" |
| "ldr q10, [x28, #0xd0]\n" |
| ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e09b // udot v27.4s, v4.16b, v2.4b[1]\n" |
| ".inst 0x6fa3e09f // udot v31.4s, v4.16b, v3.4b[1]\n" |
| "ldr q4, [x28, #0xe0]\n" |
| ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8b8 // udot v24.4s, v5.16b, v2.4b[2]\n" |
| ".inst 0x6f83e8bc // udot v28.4s, v5.16b, v3.4b[2]\n" |
| "ldr q5, [x28, #0xf0]\n" |
| "add x28, x28, #0x100\n" |
| ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8d9 // udot v25.4s, v6.16b, v2.4b[2]\n" |
| ".inst 0x6f83e8dd // udot v29.4s, v6.16b, v3.4b[2]\n" |
| ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8fa // udot v26.4s, v7.16b, v2.4b[2]\n" |
| ".inst 0x6f83e8fe // udot v30.4s, v7.16b, v3.4b[2]\n" |
| ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" |
| ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" |
| ".inst 0x6f82e91b // udot v27.4s, v8.16b, v2.4b[2]\n" |
| ".inst 0x6f83e91f // udot v31.4s, v8.16b, v3.4b[2]\n" |
| ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e938 // udot v24.4s, v9.16b, v2.4b[3]\n" |
| ".inst 0x6fa3e93c // udot v28.4s, v9.16b, v3.4b[3]\n" |
| ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e959 // udot v25.4s, v10.16b, v2.4b[3]\n" |
| ".inst 0x6fa3e95d // udot v29.4s, v10.16b, v3.4b[3]\n" |
| ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e89a // udot v26.4s, v4.16b, v2.4b[3]\n" |
| ".inst 0x6fa3e89e // udot v30.4s, v4.16b, v3.4b[3]\n" |
| ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e8b7 // udot v23.4s, v5.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e8bb // udot v27.4s, v5.16b, v2.4b[3]\n" |
| ".inst 0x6fa3e8bf // udot v31.4s, v5.16b, v3.4b[3]\n" |
| "tbnz %x[flags], #31, 98f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" |
| ".inst 0x6e8f946e // udot v14.4s, v3.16b, v15.16b\n" |
| "98:" // Height 4: Multiply loop: unique 13: skip row sum |
| "prfm pldl1keep, [x23, #0x80]\n" |
| "sub x24, x24, #0x10\n" |
| "prfm pldl1keep, [x22, #0x80]\n" |
| "cmp x24, #0x20\n" |
| "prfm pldl1keep, [x21, #0x80]\n" |
| "prfm pldl1keep, [x20, #0x80]\n" |
| "ldr q0, [x23, #0x0]\n" |
| "ldr q1, [x22, #0x0]\n" |
| "ldr q2, [x21, #0x0]\n" |
| "ldr q3, [x20, #0x0]\n" |
| "ldr q4, [x28, #0x0]\n" |
| "bge 97b\n" |
| "99:" // Height 4: Multiply loop: Single iteration only |
| ".inst 0x6f80e090 // udot v16.4s, v4.16b, v0.4b[0]\n" |
| "ldr q5, [x28, #0x10]\n" |
| "sub x24, x24, #0x10\n" |
| ".inst 0x6f81e094 // udot v20.4s, v4.16b, v1.4b[0]\n" |
| "ldr q6, [x28, #0x20]\n" |
| "add x23, x23, #0x10\n" |
| ".inst 0x6f82e098 // udot v24.4s, v4.16b, v2.4b[0]\n" |
| "ldr q7, [x28, #0x30]\n" |
| "add x22, x22, #0x10\n" |
| ".inst 0x6f83e09c // udot v28.4s, v4.16b, v3.4b[0]\n" |
| "ldr q8, [x28, #0x40]\n" |
| "add x21, x21, #0x10\n" |
| ".inst 0x6f80e0b1 // udot v17.4s, v5.16b, v0.4b[0]\n" |
| "ldr q9, [x28, #0x50]\n" |
| "add x20, x20, #0x10\n" |
| ".inst 0x6f81e0b5 // udot v21.4s, v5.16b, v1.4b[0]\n" |
| "ldr q10, [x28, #0x60]\n" |
| ".inst 0x6f82e0b9 // udot v25.4s, v5.16b, v2.4b[0]\n" |
| "ldr q4, [x28, #0x70]\n" |
| ".inst 0x6f83e0bd // udot v29.4s, v5.16b, v3.4b[0]\n" |
| "ldr q5, [x28, #0x80]\n" |
| ".inst 0x6f80e0d2 // udot v18.4s, v6.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0d6 // udot v22.4s, v6.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0da // udot v26.4s, v6.16b, v2.4b[0]\n" |
| ".inst 0x6f83e0de // udot v30.4s, v6.16b, v3.4b[0]\n" |
| "ldr q6, [x28, #0x90]\n" |
| ".inst 0x6f80e0f3 // udot v19.4s, v7.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0f7 // udot v23.4s, v7.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0fb // udot v27.4s, v7.16b, v2.4b[0]\n" |
| ".inst 0x6f83e0ff // udot v31.4s, v7.16b, v3.4b[0]\n" |
| "ldr q7, [x28, #0xa0]\n" |
| ".inst 0x6fa0e110 // udot v16.4s, v8.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e114 // udot v20.4s, v8.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e118 // udot v24.4s, v8.16b, v2.4b[1]\n" |
| ".inst 0x6fa3e11c // udot v28.4s, v8.16b, v3.4b[1]\n" |
| "ldr q8, [x28, #0xb0]\n" |
| ".inst 0x6fa0e131 // udot v17.4s, v9.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e135 // udot v21.4s, v9.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e139 // udot v25.4s, v9.16b, v2.4b[1]\n" |
| ".inst 0x6fa3e13d // udot v29.4s, v9.16b, v3.4b[1]\n" |
| "ldr q9, [x28, #0xc0]\n" |
| ".inst 0x6fa0e152 // udot v18.4s, v10.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e156 // udot v22.4s, v10.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e15a // udot v26.4s, v10.16b, v2.4b[1]\n" |
| ".inst 0x6fa3e15e // udot v30.4s, v10.16b, v3.4b[1]\n" |
| "ldr q10, [x28, #0xd0]\n" |
| ".inst 0x6fa0e093 // udot v19.4s, v4.16b, v0.4b[1]\n" |
| ".inst 0x6fa1e097 // udot v23.4s, v4.16b, v1.4b[1]\n" |
| ".inst 0x6fa2e09b // udot v27.4s, v4.16b, v2.4b[1]\n" |
| ".inst 0x6fa3e09f // udot v31.4s, v4.16b, v3.4b[1]\n" |
| "ldr q4, [x28, #0xe0]\n" |
| ".inst 0x6f80e8b0 // udot v16.4s, v5.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8b4 // udot v20.4s, v5.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8b8 // udot v24.4s, v5.16b, v2.4b[2]\n" |
| ".inst 0x6f83e8bc // udot v28.4s, v5.16b, v3.4b[2]\n" |
| "ldr q5, [x28, #0xf0]\n" |
| "add x28, x28, #0x100\n" |
| ".inst 0x6f80e8d1 // udot v17.4s, v6.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8d5 // udot v21.4s, v6.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8d9 // udot v25.4s, v6.16b, v2.4b[2]\n" |
| ".inst 0x6f83e8dd // udot v29.4s, v6.16b, v3.4b[2]\n" |
| ".inst 0x6f80e8f2 // udot v18.4s, v7.16b, v0.4b[2]\n" |
| ".inst 0x6f81e8f6 // udot v22.4s, v7.16b, v1.4b[2]\n" |
| ".inst 0x6f82e8fa // udot v26.4s, v7.16b, v2.4b[2]\n" |
| ".inst 0x6f83e8fe // udot v30.4s, v7.16b, v3.4b[2]\n" |
| ".inst 0x6f80e913 // udot v19.4s, v8.16b, v0.4b[2]\n" |
| ".inst 0x6f81e917 // udot v23.4s, v8.16b, v1.4b[2]\n" |
| ".inst 0x6f82e91b // udot v27.4s, v8.16b, v2.4b[2]\n" |
| ".inst 0x6f83e91f // udot v31.4s, v8.16b, v3.4b[2]\n" |
| ".inst 0x6fa0e930 // udot v16.4s, v9.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e934 // udot v20.4s, v9.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e938 // udot v24.4s, v9.16b, v2.4b[3]\n" |
| ".inst 0x6fa3e93c // udot v28.4s, v9.16b, v3.4b[3]\n" |
| ".inst 0x6fa0e951 // udot v17.4s, v10.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e955 // udot v21.4s, v10.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e959 // udot v25.4s, v10.16b, v2.4b[3]\n" |
| ".inst 0x6fa3e95d // udot v29.4s, v10.16b, v3.4b[3]\n" |
| ".inst 0x6fa0e892 // udot v18.4s, v4.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e896 // udot v22.4s, v4.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e89a // udot v26.4s, v4.16b, v2.4b[3]\n" |
| ".inst 0x6fa3e89e // udot v30.4s, v4.16b, v3.4b[3]\n" |
| ".inst 0x6fa0e8b3 // udot v19.4s, v5.16b, v0.4b[3]\n" |
| ".inst 0x6fa1e8b7 // udot v23.4s, v5.16b, v1.4b[3]\n" |
| ".inst 0x6fa2e8bb // udot v27.4s, v5.16b, v2.4b[3]\n" |
| ".inst 0x6fa3e8bf // udot v31.4s, v5.16b, v3.4b[3]\n" |
| "tbnz %x[flags], #31, 100f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" |
| ".inst 0x6e8f946e // udot v14.4s, v3.16b, v15.16b\n" |
| "100:" // Height 4: Multiply loop: unique 14: skip row sum |
| "prfm pldl1keep, [x23, #0x80]\n" |
| "prfm pldl1keep, [x22, #0x80]\n" |
| "prfm pldl1keep, [x21, #0x80]\n" |
| "prfm pldl1keep, [x20, #0x80]\n" |
| "101:" // Height 4: Multiply loop: Main loop skip |
| "cbz x24, 108f\n" |
| "cmp x24, #0x4\n" |
| "blt 104f\n" |
| "102:" // Height 4: Multiply loop: Odd block loop |
| "ldr s0, [x23], #0x4\n" |
| "ldr s1, [x22], #0x4\n" |
| "ldr s2, [x21], #0x4\n" |
| "ldr s3, [x20], #0x4\n" |
| "tbnz %x[flags], #31, 103f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" |
| ".inst 0x6e8f946e // udot v14.4s, v3.16b, v15.16b\n" |
| "103:" // Height 4: Multiply loop: unique 15: skip row sum |
| "ldr q6, [x28, #0x0]\n" |
| ".inst 0x6f80e0d0 // udot v16.4s, v6.16b, v0.4b[0]\n" |
| "ldr q7, [x28, #0x10]\n" |
| "sub x24, x24, #0x4\n" |
| ".inst 0x6f81e0d4 // udot v20.4s, v6.16b, v1.4b[0]\n" |
| "ldr q8, [x28, #0x20]\n" |
| "cmp x24, #0x4\n" |
| ".inst 0x6f82e0d8 // udot v24.4s, v6.16b, v2.4b[0]\n" |
| "ldr q9, [x28, #0x30]\n" |
| "add x28, x28, #0x40\n" |
| ".inst 0x6f83e0dc // udot v28.4s, v6.16b, v3.4b[0]\n" |
| ".inst 0x6f80e0f1 // udot v17.4s, v7.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0f5 // udot v21.4s, v7.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0f9 // udot v25.4s, v7.16b, v2.4b[0]\n" |
| ".inst 0x6f83e0fd // udot v29.4s, v7.16b, v3.4b[0]\n" |
| ".inst 0x6f80e112 // udot v18.4s, v8.16b, v0.4b[0]\n" |
| ".inst 0x6f81e116 // udot v22.4s, v8.16b, v1.4b[0]\n" |
| ".inst 0x6f82e11a // udot v26.4s, v8.16b, v2.4b[0]\n" |
| ".inst 0x6f83e11e // udot v30.4s, v8.16b, v3.4b[0]\n" |
| ".inst 0x6f80e133 // udot v19.4s, v9.16b, v0.4b[0]\n" |
| ".inst 0x6f81e137 // udot v23.4s, v9.16b, v1.4b[0]\n" |
| ".inst 0x6f82e13b // udot v27.4s, v9.16b, v2.4b[0]\n" |
| ".inst 0x6f83e13f // udot v31.4s, v9.16b, v3.4b[0]\n" |
| "bge 102b\n" |
| "cbz x24, 108f\n" |
| "104:" // Height 4: Multiply loop: Skip odd blocks |
| "tbz x24, #1, 105f\n" |
| "ldr h0, [x23], #0x2\n" |
| "ldr h1, [x22], #0x2\n" |
| "ldr h2, [x21], #0x2\n" |
| "ldr h3, [x20], #0x2\n" |
| "tbz x24, #0, 106f\n" |
| "ld1 { v0.b }[2], [x23]\n" |
| "ld1 { v1.b }[2], [x22]\n" |
| "ld1 { v2.b }[2], [x21]\n" |
| "ld1 { v3.b }[2], [x20]\n" |
| "b 106f\n" |
| "105:" // Height 4: Multiply loop: Ragged operand read: partial_1_0 |
| "ldr b0, [x23, #0x0]\n" |
| "ldr b1, [x22, #0x0]\n" |
| "ldr b2, [x21, #0x0]\n" |
| "ldr b3, [x20, #0x0]\n" |
| "106:" // Height 4: Multiply loop: Ragged operand read: Done |
| "tbnz %x[flags], #31, 107f\n" |
| ".inst 0x6e8f940b // udot v11.4s, v0.16b, v15.16b\n" |
| ".inst 0x6e8f942c // udot v12.4s, v1.16b, v15.16b\n" |
| ".inst 0x6e8f944d // udot v13.4s, v2.16b, v15.16b\n" |
| ".inst 0x6e8f946e // udot v14.4s, v3.16b, v15.16b\n" |
| "107:" // Height 4: Multiply loop: unique 16: skip row sum |
| "ldr q10, [x28, #0x0]\n" |
| ".inst 0x6f80e150 // udot v16.4s, v10.16b, v0.4b[0]\n" |
| "ldr q4, [x28, #0x10]\n" |
| ".inst 0x6f81e154 // udot v20.4s, v10.16b, v1.4b[0]\n" |
| "ldr q5, [x28, #0x20]\n" |
| ".inst 0x6f82e158 // udot v24.4s, v10.16b, v2.4b[0]\n" |
| "ldr q6, [x28, #0x30]\n" |
| "add x28, x28, #0x40\n" |
| ".inst 0x6f83e15c // udot v28.4s, v10.16b, v3.4b[0]\n" |
| ".inst 0x6f80e091 // udot v17.4s, v4.16b, v0.4b[0]\n" |
| ".inst 0x6f81e095 // udot v21.4s, v4.16b, v1.4b[0]\n" |
| ".inst 0x6f82e099 // udot v25.4s, v4.16b, v2.4b[0]\n" |
| ".inst 0x6f83e09d // udot v29.4s, v4.16b, v3.4b[0]\n" |
| ".inst 0x6f80e0b2 // udot v18.4s, v5.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0b6 // udot v22.4s, v5.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0ba // udot v26.4s, v5.16b, v2.4b[0]\n" |
| ".inst 0x6f83e0be // udot v30.4s, v5.16b, v3.4b[0]\n" |
| ".inst 0x6f80e0d3 // udot v19.4s, v6.16b, v0.4b[0]\n" |
| ".inst 0x6f81e0d7 // udot v23.4s, v6.16b, v1.4b[0]\n" |
| ".inst 0x6f82e0db // udot v27.4s, v6.16b, v2.4b[0]\n" |
| ".inst 0x6f83e0df // udot v31.4s, v6.16b, v3.4b[0]\n" |
| "108:" // Height 4: Multiply loop: No odd multiplies |
| "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" |
| "add x25, x25, #0x1\n" |
| "cmp x25, x19\n" |
| "bne 94b\n" |
| "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" |
| "prfm pstl1keep, [x26, #0x0]\n" |
| "add x21, x26, x19\n" |
| "prfm pstl1keep, [x21, #0x0]\n" |
| "add x20, x21, x19\n" |
| "prfm pstl1keep, [x20, #0x0]\n" |
| "add x19, x20, x19\n" |
| "prfm pstl1keep, [x19, #0x0]\n" |
| "tbnz %x[flags], #31, 109f\n" |
| "addp v11.4s, v11.4s, v11.4s\n" |
| "add x22, %x[qp], %[b_offset]\n" |
| "ld1r { v4.4s }, [x22]\n" |
| "addp v12.4s, v12.4s, v12.4s\n" |
| "addp v13.4s, v13.4s, v13.4s\n" |
| "addp v14.4s, v14.4s, v14.4s\n" |
| "addp v11.4s, v11.4s, v11.4s\n" |
| "neg v4.4s, v4.4s\n" |
| "addp v12.4s, v12.4s, v12.4s\n" |
| "addp v13.4s, v13.4s, v13.4s\n" |
| "addp v14.4s, v14.4s, v14.4s\n" |
| "mul v11.4s, v11.4s, v4.4s\n" |
| "mul v12.4s, v12.4s, v4.4s\n" |
| "mul v13.4s, v13.4s, v4.4s\n" |
| "mul v14.4s, v14.4s, v4.4s\n" |
| "109:" // Height 4: skip row sum fixup |
| "add v16.4s, v16.4s, v11.4s\n" |
| "ldr q0, [x27, #0x0]\n" |
| "orr %x[flags], %x[flags], #0x80000000\n" |
| "add v17.4s, v17.4s, v11.4s\n" |
| "ldr q1, [x27, #0x10]\n" |
| "add x23, %x[qp], %[per_layer_right_shift]\n" |
| "add v18.4s, v18.4s, v11.4s\n" |
| "ldr q2, [x27, #0x20]\n" |
| "add x22, %x[qp], %[per_layer_mul]\n" |
| "add v19.4s, v19.4s, v11.4s\n" |
| "ldr q3, [x27, #0x30]\n" |
| "add x27, x27, #0x40\n" |
| "add v20.4s, v20.4s, v12.4s\n" |
| "ld1r { v4.4s }, [x22]\n" |
| "add v21.4s, v21.4s, v12.4s\n" |
| "add v22.4s, v22.4s, v12.4s\n" |
| "add v23.4s, v23.4s, v12.4s\n" |
| "add v24.4s, v24.4s, v13.4s\n" |
| "add v25.4s, v25.4s, v13.4s\n" |
| "add v26.4s, v26.4s, v13.4s\n" |
| "add v27.4s, v27.4s, v13.4s\n" |
| "add v28.4s, v28.4s, v14.4s\n" |
| "add v29.4s, v29.4s, v14.4s\n" |
| "add v30.4s, v30.4s, v14.4s\n" |
| "add v31.4s, v31.4s, v14.4s\n" |
| "add v16.4s, v16.4s, v0.4s\n" |
| "add v17.4s, v17.4s, v1.4s\n" |
| "add v18.4s, v18.4s, v2.4s\n" |
| "add v19.4s, v19.4s, v3.4s\n" |
| "add v20.4s, v20.4s, v0.4s\n" |
| "add v21.4s, v21.4s, v1.4s\n" |
| "add v22.4s, v22.4s, v2.4s\n" |
| "add v23.4s, v23.4s, v3.4s\n" |
| "add v24.4s, v24.4s, v0.4s\n" |
| "add v25.4s, v25.4s, v1.4s\n" |
| "add v26.4s, v26.4s, v2.4s\n" |
| "add v27.4s, v27.4s, v3.4s\n" |
| "add v28.4s, v28.4s, v0.4s\n" |
| "ld1r { v0.4s }, [x23]\n" |
| "add v29.4s, v29.4s, v1.4s\n" |
| "add v30.4s, v30.4s, v2.4s\n" |
| "add v31.4s, v31.4s, v3.4s\n" |
| "sqrdmulh v16.4s, v16.4s, v4.4s\n" |
| "sqrdmulh v17.4s, v17.4s, v4.4s\n" |
| "sqrdmulh v18.4s, v18.4s, v4.4s\n" |
| "sqrdmulh v19.4s, v19.4s, v4.4s\n" |
| "sqrdmulh v20.4s, v20.4s, v4.4s\n" |
| "sqrdmulh v21.4s, v21.4s, v4.4s\n" |
| "sqrdmulh v22.4s, v22.4s, v4.4s\n" |
| "sqrdmulh v23.4s, v23.4s, v4.4s\n" |
| "sqrdmulh v24.4s, v24.4s, v4.4s\n" |
| "sqrdmulh v25.4s, v25.4s, v4.4s\n" |
| "sqrdmulh v26.4s, v26.4s, v4.4s\n" |
| "sqrdmulh v27.4s, v27.4s, v4.4s\n" |
| "sqrdmulh v28.4s, v28.4s, v4.4s\n" |
| "sqrdmulh v29.4s, v29.4s, v4.4s\n" |
| "sqrdmulh v30.4s, v30.4s, v4.4s\n" |
| "sqrdmulh v31.4s, v31.4s, v4.4s\n" |
| "tbz %x[flags], #5, 110f\n" |
| "and v4.16b, v16.16b, v0.16b\n" |
| "and v5.16b, v17.16b, v0.16b\n" |
| "and v6.16b, v18.16b, v0.16b\n" |
| "sshr v4.4s, v4.4s, #0x1f\n" |
| "sshr v5.4s, v5.4s, #0x1f\n" |
| "sshr v6.4s, v6.4s, #0x1f\n" |
| "sqadd v16.4s, v16.4s, v4.4s\n" |
| "sqadd v17.4s, v17.4s, v5.4s\n" |
| "sqadd v18.4s, v18.4s, v6.4s\n" |
| "and v7.16b, v19.16b, v0.16b\n" |
| "and v8.16b, v20.16b, v0.16b\n" |
| "and v9.16b, v21.16b, v0.16b\n" |
| "sshr v7.4s, v7.4s, #0x1f\n" |
| "sshr v8.4s, v8.4s, #0x1f\n" |
| "sshr v9.4s, v9.4s, #0x1f\n" |
| "sqadd v19.4s, v19.4s, v7.4s\n" |
| "sqadd v20.4s, v20.4s, v8.4s\n" |
| "sqadd v21.4s, v21.4s, v9.4s\n" |
| "and v10.16b, v22.16b, v0.16b\n" |
| "and v4.16b, v23.16b, v0.16b\n" |
| "and v5.16b, v24.16b, v0.16b\n" |
| "sshr v10.4s, v10.4s, #0x1f\n" |
| "sshr v4.4s, v4.4s, #0x1f\n" |
| "sshr v5.4s, v5.4s, #0x1f\n" |
| "sqadd v22.4s, v22.4s, v10.4s\n" |
| "sqadd v23.4s, v23.4s, v4.4s\n" |
| "sqadd v24.4s, v24.4s, v5.4s\n" |
| "and v6.16b, v25.16b, v0.16b\n" |
| "and v7.16b, v26.16b, v0.16b\n" |
| "and v8.16b, v27.16b, v0.16b\n" |
| "sshr v6.4s, v6.4s, #0x1f\n" |
| "sshr v7.4s, v7.4s, #0x1f\n" |
| "sshr v8.4s, v8.4s, #0x1f\n" |
| "sqadd v25.4s, v25.4s, v6.4s\n" |
| "sqadd v26.4s, v26.4s, v7.4s\n" |
| "sqadd v27.4s, v27.4s, v8.4s\n" |
| "and v9.16b, v28.16b, v0.16b\n" |
| "and v10.16b, v29.16b, v0.16b\n" |
| "and v4.16b, v30.16b, v0.16b\n" |
| "sshr v9.4s, v9.4s, #0x1f\n" |
| "sshr v10.4s, v10.4s, #0x1f\n" |
| "sshr v4.4s, v4.4s, #0x1f\n" |
| "sqadd v28.4s, v28.4s, v9.4s\n" |
| "sqadd v29.4s, v29.4s, v10.4s\n" |
| "sqadd v30.4s, v30.4s, v4.4s\n" |
| "and v5.16b, v31.16b, v0.16b\n" |
| "sshr v5.4s, v5.4s, #0x1f\n" |
| "sqadd v31.4s, v31.4s, v5.4s\n" |
| "110:" // Height 4: no shift correction |
| "srshl v16.4s, v16.4s, v0.4s\n" |
| "add x22, %x[qp], %[c_offset]\n" |
| "ld1r { v4.4s }, [x22]\n" |
| "srshl v17.4s, v17.4s, v0.4s\n" |
| "add x22, %x[qp], %[minval]\n" |
| "srshl v18.4s, v18.4s, v0.4s\n" |
| "ld1r { v5.4s }, [x22]\n" |
| "add x22, %x[qp], %[maxval]\n" |
| "srshl v19.4s, v19.4s, v0.4s\n" |
| "ld1r { v6.4s }, [x22]\n" |
| "cmp x9, #0x10\n" |
| "srshl v20.4s, v20.4s, v0.4s\n" |
| "srshl v21.4s, v21.4s, v0.4s\n" |
| "add v16.4s, v16.4s, v4.4s\n" |
| "add v17.4s, v17.4s, v4.4s\n" |
| "add v18.4s, v18.4s, v4.4s\n" |
| "smin v16.4s, v16.4s, v6.4s\n" |
| "smin v17.4s, v17.4s, v6.4s\n" |
| "smin v18.4s, v18.4s, v6.4s\n" |
| "smax v16.4s, v16.4s, v5.4s\n" |
| "smax v17.4s, v17.4s, v5.4s\n" |
| "smax v18.4s, v18.4s, v5.4s\n" |
| "add v19.4s, v19.4s, v4.4s\n" |
| "add v20.4s, v20.4s, v4.4s\n" |
| "add v21.4s, v21.4s, v4.4s\n" |
| "smin v19.4s, v19.4s, v6.4s\n" |
| "smin v20.4s, v20.4s, v6.4s\n" |
| "smin v21.4s, v21.4s, v6.4s\n" |
| "smax v19.4s, v19.4s, v5.4s\n" |
| "smax v20.4s, v20.4s, v5.4s\n" |
| "smax v21.4s, v21.4s, v5.4s\n" |
| "srshl v22.4s, v22.4s, v0.4s\n" |
| "srshl v23.4s, v23.4s, v0.4s\n" |
| "srshl v24.4s, v24.4s, v0.4s\n" |
| "srshl v25.4s, v25.4s, v0.4s\n" |
| "add v22.4s, v22.4s, v4.4s\n" |
| "add v23.4s, v23.4s, v4.4s\n" |
| "add v24.4s, v24.4s, v4.4s\n" |
| "smin v22.4s, v22.4s, v6.4s\n" |
| "smin v23.4s, v23.4s, v6.4s\n" |
| "smin v24.4s, v24.4s, v6.4s\n" |
| "smax v22.4s, v22.4s, v5.4s\n" |
| "smax v23.4s, v23.4s, v5.4s\n" |
| "smax v24.4s, v24.4s, v5.4s\n" |
| "add v25.4s, v25.4s, v4.4s\n" |
| "srshl v26.4s, v26.4s, v0.4s\n" |
| "srshl v27.4s, v27.4s, v0.4s\n" |
| "smin v25.4s, v25.4s, v6.4s\n" |
| "srshl v28.4s, v28.4s, v0.4s\n" |
| "add v26.4s, v26.4s, v4.4s\n" |
| "smax v25.4s, v25.4s, v5.4s\n" |
| "add v27.4s, v27.4s, v4.4s\n" |
| "smin v26.4s, v26.4s, v6.4s\n" |
| "add v28.4s, v28.4s, v4.4s\n" |
| "smin v27.4s, v27.4s, v6.4s\n" |
| "smax v26.4s, v26.4s, v5.4s\n" |
| "smin v28.4s, v28.4s, v6.4s\n" |
| "smax v27.4s, v27.4s, v5.4s\n" |
| "srshl v29.4s, v29.4s, v0.4s\n" |
| "smax v28.4s, v28.4s, v5.4s\n" |
| "srshl v30.4s, v30.4s, v0.4s\n" |
| "srshl v31.4s, v31.4s, v0.4s\n" |
| "add v29.4s, v29.4s, v4.4s\n" |
| "uzp1 v16.8h, v16.8h, v17.8h\n" |
| "add v30.4s, v30.4s, v4.4s\n" |
| "smin v29.4s, v29.4s, v6.4s\n" |
| "add v31.4s, v31.4s, v4.4s\n" |
| "smin v30.4s, v30.4s, v6.4s\n" |
| "smax v29.4s, v29.4s, v5.4s\n" |
| "smin v31.4s, v31.4s, v6.4s\n" |
| "smax v30.4s, v30.4s, v5.4s\n" |
| "uzp1 v17.8h, v18.8h, v19.8h\n" |
| "smax v31.4s, v31.4s, v5.4s\n" |
| "uzp1 v20.8h, v20.8h, v21.8h\n" |
| "uzp1 v21.8h, v22.8h, v23.8h\n" |
| "uzp1 v24.8h, v24.8h, v25.8h\n" |
| "uzp1 v25.8h, v26.8h, v27.8h\n" |
| "uzp1 v28.8h, v28.8h, v29.8h\n" |
| "uzp1 v29.8h, v30.8h, v31.8h\n" |
| "uzp1 v16.16b, v16.16b, v17.16b\n" |
| "uzp1 v20.16b, v20.16b, v21.16b\n" |
| "uzp1 v24.16b, v24.16b, v25.16b\n" |
| "uzp1 v28.16b, v28.16b, v29.16b\n" |
| "bge 119f\n" |
| "tbz x9, #3, 114f\n" |
| "str d16, [x26], #0x8\n" |
| "str d20, [x21], #0x8\n" |
| "str d24, [x20], #0x8\n" |
| "str d28, [x19], #0x8\n" |
| "tbz x9, #2, 112f\n" |
| "st1 { v16.s }[2], [x26], #0x4\n" |
| "st1 { v20.s }[2], [x21], #0x4\n" |
| "st1 { v24.s }[2], [x20], #0x4\n" |
| "st1 { v28.s }[2], [x19], #0x4\n" |
| "tbz x9, #1, 111f\n" |
| "st1 { v16.h }[6], [x26], #0x2\n" |
| "st1 { v20.h }[6], [x21], #0x2\n" |
| "st1 { v24.h }[6], [x20], #0x2\n" |
| "st1 { v28.h }[6], [x19], #0x2\n" |
| "tbz x9, #0, 118f\n" |
| "st1 { v16.b }[14], [x26]\n" |
| "st1 { v20.b }[14], [x21]\n" |
| "st1 { v24.b }[14], [x20]\n" |
| "st1 { v28.b }[14], [x19]\n" |
| "b 118f\n" |
| "111:" // Height 4: Partial direct writeback: partial_1_12 |
| "tbz x9, #0, 118f\n" |
| "st1 { v16.b }[12], [x26]\n" |
| "st1 { v20.b }[12], [x21]\n" |
| "st1 { v24.b }[12], [x20]\n" |
| "st1 { v28.b }[12], [x19]\n" |
| "b 118f\n" |
| "112:" // Height 4: Partial direct writeback: partial_2_8 |
| "tbz x9, #1, 113f\n" |
| "st1 { v16.h }[4], [x26], #0x2\n" |
| "st1 { v20.h }[4], [x21], #0x2\n" |
| "st1 { v24.h }[4], [x20], #0x2\n" |
| "st1 { v28.h }[4], [x19], #0x2\n" |
| "tbz x9, #0, 118f\n" |
| "st1 { v16.b }[10], [x26]\n" |
| "st1 { v20.b }[10], [x21]\n" |
| "st1 { v24.b }[10], [x20]\n" |
| "st1 { v28.b }[10], [x19]\n" |
| "b 118f\n" |
| "113:" // Height 4: Partial direct writeback: partial_1_8 |
| "tbz x9, #0, 118f\n" |
| "st1 { v16.b }[8], [x26]\n" |
| "st1 { v20.b }[8], [x21]\n" |
| "st1 { v24.b }[8], [x20]\n" |
| "st1 { v28.b }[8], [x19]\n" |
| "b 118f\n" |
| "114:" // Height 4: Partial direct writeback: partial_4_0 |
| "tbz x9, #2, 116f\n" |
| "str s16, [x26], #0x4\n" |
| "str s20, [x21], #0x4\n" |
| "str s24, [x20], #0x4\n" |
| "str s28, [x19], #0x4\n" |
| "tbz x9, #1, 115f\n" |
| "st1 { v16.h }[2], [x26], #0x2\n" |
| "st1 { v20.h }[2], [x21], #0x2\n" |
| "st1 { v24.h }[2], [x20], #0x2\n" |
| "st1 { v28.h }[2], [x19], #0x2\n" |
| "tbz x9, #0, 118f\n" |
| "st1 { v16.b }[6], [x26]\n" |
| "st1 { v20.b }[6], [x21]\n" |
| "st1 { v24.b }[6], [x20]\n" |
| "st1 { v28.b }[6], [x19]\n" |
| "b 118f\n" |
| "115:" // Height 4: Partial direct writeback: partial_1_4 |
| "tbz x9, #0, 118f\n" |
| "st1 { v16.b }[4], [x26]\n" |
| "st1 { v20.b }[4], [x21]\n" |
| "st1 { v24.b }[4], [x20]\n" |
| "st1 { v28.b }[4], [x19]\n" |
| "b 118f\n" |
| "116:" // Height 4: Partial direct writeback: partial_2_0 |
| "tbz x9, #1, 117f\n" |
| "str h16, [x26], #0x2\n" |
| "str h20, [x21], #0x2\n" |
| "str h24, [x20], #0x2\n" |
| "str h28, [x19], #0x2\n" |
| "tbz x9, #0, 118f\n" |
| "st1 { v16.b }[2], [x26]\n" |
| "st1 { v20.b }[2], [x21]\n" |
| "st1 { v24.b }[2], [x20]\n" |
| "st1 { v28.b }[2], [x19]\n" |
| "b 118f\n" |
| "117:" // Height 4: Partial direct writeback: partial_1_0 |
| "str b16, [x26, #0x0]\n" |
| "str b20, [x21, #0x0]\n" |
| "str b24, [x20, #0x0]\n" |
| "str b28, [x19, #0x0]\n" |
| "118:" // Height 4: Partial direct writeback: Done |
| "b 120f\n" |
| "119:" // Height 4: Full writeback |
| "str q16, [x26, #0x0]\n" |
| "add x26, x26, #0x10\n" |
| "str q20, [x21, #0x0]\n" |
| "str q24, [x20, #0x0]\n" |
| "str q28, [x19, #0x0]\n" |
| "120:" // Height 4: Writeback done |
| "subs x9, x9, #0x10\n" |
| "bgt 92b\n" |
| "subs %x[M], %x[M], #0x4\n" |
| "beq 122f\n" |
| "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" |
| "tbz %x[flags], #3, 121f\n" |
| "add x20, x20, #0x4\n" |
| "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n" |
| "b 1b\n" |
| "121:" // Update direct input |
| "mov x19, #0x4\n" |
| "madd %x[input_ptr], x19, x20, %x[input_ptr]\n" |
| "b 1b\n" |
| "122:" // 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", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" |
| ); |
| } |
| |
| } // namespace arm_gemm |
| #endif // __aarch64__ |