From c0b6f76561580414f08633a804fc548ccad65659 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 2 Nov 2020 01:37:17 +0000 Subject: COMPMID-3776: Indirect GEMM Signed-off-by: Georgios Pinitas Change-Id: I51a1b0f098bc3a8c408c50c92221e4df3061e12c Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4343 Tested-by: Arm Jenkins Reviewed-by: Sang-Hoon Park Reviewed-by: Michele Di Giorgio Comments-Addressed: Arm Jenkins --- .../kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp | 1904 ++++++++++++++++++++ 1 file changed, 1904 insertions(+) create mode 100644 src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp new file mode 100644 index 0000000000..fc8ce636dd --- /dev/null +++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8u32_dot_6x4VL/generic.cpp @@ -0,0 +1,1904 @@ +/* + * Copyright (c) 2019-2020 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ +#ifdef __ARM_FEATURE_SVE + +#include "arm_gemm.hpp" +#include "../../utils.hpp" + +#include + +namespace arm_gemm { + +void sve_hybrid_u8u32_dot_6x4VL ( + unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg A_arg, + size_t M, size_t N, const uint8_t *B_ptr, IndirectOutputArg output_arg, + const uint32_t *, Activation, bool accumulate +) +{ + 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; + } + if (accumulate) { + flags |= 0x1; + } + ka.num_strings = num_strings; + ka.string_lengths = string_lengths; + ka.N = N; + ka.B_ptr = B_ptr; + __asm__ __volatile__( + "ptrue p5.b\n" + "1:" // Row loop + "cmp %x[M], #0x6\n" + "bge 61f\n" + "cmp %x[M], #0x4\n" + "bgt 49f\n" + "beq 37f\n" + "cmp %x[M], #0x2\n" + "bgt 25f\n" + "beq 13f\n" + "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "tbz %x[flags], #2, 2f\n" + "ldr x13, [%x[output_ptr], #0x0]\n" + "add x13, x13, x19, LSL #2\n" + "b 3f\n" + "2:" // Height 1: setup direct output + "mov x13, %x[output_ptr]\n" + "3:" // Height 1: Column loop + "mov x19, #0x0\n" + "whilelt p4.s, x19, x15\n" + "incw x19\n" + "whilelt p3.s, x19, x15\n" + "incw x19\n" + "whilelt p2.s, x19, x15\n" + "incw x19\n" + "whilelt p1.s, x19, x15\n" + "tbz %x[flags], #0, 4f\n" + "ld1w { z8.s }, p4/Z, [x13]\n" + "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" + "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" + "b 5f\n" + "4:" // Height 1: no accumulate + "mov z8.s, #0x0\n" + "mov z9.s, #0x0\n" + "mov z10.s, #0x0\n" + "mov z11.s, #0x0\n" + "5:" // Height 1: setup done + "mov x12, #0x0\n" + "6:" // Height 1: String loop + "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w11, [x20, x12, LSL #0x2]\n" + "tbz %x[flags], #3, 7f\n" + "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "add x20, x20, x19, LSL #3\n" + "ldr x10, [x20, #0x0]\n" + "cbnz x12, 8f\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x10, x10, x19\n" + "b 8f\n" + "7:" // Height 1: setup direct input + "mov x10, %x[input_ptr]\n" + "8:" // Height 1: input setup done + "cmp x11, #0x10\n" + "ble 10f\n" + "9:" // Height 1: Multiply loop: Main loop head + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "sub x11, x11, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "cmp x11, #0x10\n" + "udot z10.s, z6.b, z0.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #4, MUL VL]\n" + "prfm pldl1keep, [x10, #0x80]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #5, MUL VL]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #7, MUL VL]\n" + "addvl x14, x14, #16\n" + "udot z10.s, z6.b, z0.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #-8, MUL VL]\n" + "udot z11.s, z7.b, z0.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #-7, MUL VL]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-5, MUL VL]\n" + "udot z10.s, z6.b, z0.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-4, MUL VL]\n" + "udot z11.s, z7.b, z0.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-3, MUL VL]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "ld1b { z6.b }, p5/Z, [x14, #-2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #-1, MUL VL]\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "bgt 9b\n" + "10:" // Height 1: Multiply loop: Single iteration only + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[0]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "ble 11f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z9.s, z7.b, z0.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "udot z10.s, z6.b, z0.b[1]\n" + "addvl x14, x14, #4\n" + "udot z11.s, z7.b, z0.b[1]\n" + "ble 11f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z9.s, z7.b, z0.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "udot z10.s, z6.b, z0.b[2]\n" + "addvl x14, x14, #4\n" + "udot z11.s, z7.b, z0.b[2]\n" + "ble 11f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "11:" // Height 1: Multiply loop: multiply skip + "prfm pldl1keep, [x10, #0x80]\n" + "add x12, x12, #0x1\n" + "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "cmp x12, x19\n" + "bne 6b\n" + "prfm pstl1keep, [x13, #0x0]\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "12:" // Height 1: Writeback done + "mov x19, #0x0\n" + "incw x19, ALL, MUL #4\n" + "subs x15, x15, x19\n" + "bgt 3b\n" + "b 74f\n" + "13:" // Height 2 + "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "tbz %x[flags], #2, 14f\n" + "ldr x13, [%x[output_ptr], #0x0]\n" + "add x13, x13, x19, LSL #2\n" + "ldr x9, [%x[output_ptr], #0x8]\n" + "add x9, x9, x19, LSL #2\n" + "b 15f\n" + "14:" // Height 2: setup direct output + "mov x13, %x[output_ptr]\n" + "add x9, x13, x19, LSL #2\n" + "15:" // Height 2: Column loop + "mov x19, #0x0\n" + "whilelt p4.s, x19, x15\n" + "incw x19\n" + "whilelt p3.s, x19, x15\n" + "incw x19\n" + "whilelt p2.s, x19, x15\n" + "incw x19\n" + "whilelt p1.s, x19, x15\n" + "tbz %x[flags], #0, 16f\n" + "ld1w { z8.s }, p4/Z, [x13]\n" + "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" + "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x9]\n" + "ld1w { z13.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x9, #3, MUL VL]\n" + "b 17f\n" + "16:" // Height 2: no accumulate + "mov z8.s, #0x0\n" + "mov z9.s, #0x0\n" + "mov z10.s, #0x0\n" + "mov z11.s, #0x0\n" + "mov z12.s, #0x0\n" + "mov z13.s, #0x0\n" + "mov z14.s, #0x0\n" + "mov z15.s, #0x0\n" + "17:" // Height 2: setup done + "mov x12, #0x0\n" + "18:" // Height 2: String loop + "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w11, [x20, x12, LSL #0x2]\n" + "tbz %x[flags], #3, 19f\n" + "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "add x20, x20, x19, LSL #3\n" + "ldr x10, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "cbnz x12, 20f\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x10, x10, x19\n" + "add x28, x28, x19\n" + "b 20f\n" + "19:" // Height 2: setup direct input + "mov x10, %x[input_ptr]\n" + "add x28, x10, x19\n" + "20:" // Height 2: input setup done + "cmp x11, #0x10\n" + "ble 22f\n" + "21:" // Height 2: Multiply loop: Main loop head + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "sub x11, x11, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1rqb { z1.b }, p0/Z, [x28]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "prfm pldl1keep, [x10, #0x80]\n" + "add x28, x28, #0x10\n" + "udot z12.s, z6.b, z1.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "cmp x11, #0x10\n" + "udot z13.s, z7.b, z1.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "udot z10.s, z6.b, z0.b[0]\n" + "udot z14.s, z6.b, z1.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #4, MUL VL]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "udot z15.s, z7.b, z1.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #5, MUL VL]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "udot z12.s, z6.b, z1.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[1]\n" + "udot z13.s, z7.b, z1.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #7, MUL VL]\n" + "addvl x14, x14, #16\n" + "udot z10.s, z6.b, z0.b[1]\n" + "udot z14.s, z6.b, z1.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #-8, MUL VL]\n" + "udot z11.s, z7.b, z0.b[1]\n" + "udot z15.s, z7.b, z1.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #-7, MUL VL]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "udot z12.s, z6.b, z1.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[2]\n" + "udot z13.s, z7.b, z1.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-5, MUL VL]\n" + "udot z10.s, z6.b, z0.b[2]\n" + "udot z14.s, z6.b, z1.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-4, MUL VL]\n" + "udot z11.s, z7.b, z0.b[2]\n" + "udot z15.s, z7.b, z1.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-3, MUL VL]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "udot z12.s, z6.b, z1.b[3]\n" + "ld1b { z6.b }, p5/Z, [x14, #-2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "udot z13.s, z7.b, z1.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #-1, MUL VL]\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z14.s, z6.b, z1.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "udot z15.s, z7.b, z1.b[3]\n" + "bgt 21b\n" + "22:" // Height 2: Multiply loop: Single iteration only + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1rqb { z1.b }, p0/Z, [x28]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "add x28, x28, #0x10\n" + "udot z12.s, z6.b, z1.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z13.s, z7.b, z1.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[0]\n" + "udot z14.s, z6.b, z1.b[0]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "udot z15.s, z7.b, z1.b[0]\n" + "ble 23f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z12.s, z6.b, z1.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[1]\n" + "udot z13.s, z7.b, z1.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[1]\n" + "udot z14.s, z6.b, z1.b[1]\n" + "udot z11.s, z7.b, z0.b[1]\n" + "udot z15.s, z7.b, z1.b[1]\n" + "ble 23f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z12.s, z6.b, z1.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[2]\n" + "udot z13.s, z7.b, z1.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[2]\n" + "udot z14.s, z6.b, z1.b[2]\n" + "udot z11.s, z7.b, z0.b[2]\n" + "udot z15.s, z7.b, z1.b[2]\n" + "ble 23f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "udot z12.s, z6.b, z1.b[3]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "udot z13.s, z7.b, z1.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z14.s, z6.b, z1.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "udot z15.s, z7.b, z1.b[3]\n" + "23:" // Height 2: Multiply loop: multiply skip + "prfm pldl1keep, [x10, #0x80]\n" + "add x12, x12, #0x1\n" + "prfm pldl1keep, [x28, #0x80]\n" + "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "cmp x12, x19\n" + "bne 18b\n" + "prfm pstl1keep, [x13, #0x0]\n" + "prfm pstl1keep, [x9, #0x0]\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "st1w { z12.s }, p4, [x9]\n" + "st1w { z13.s }, p3, [x9, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z15.s }, p1, [x9, #3, MUL VL]\n" + "addvl x9, x9, #4\n" + "24:" // Height 2: Writeback done + "mov x19, #0x0\n" + "incw x19, ALL, MUL #4\n" + "subs x15, x15, x19\n" + "bgt 15b\n" + "b 74f\n" + "25:" // Height 3 + "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "tbz %x[flags], #2, 26f\n" + "ldr x13, [%x[output_ptr], #0x0]\n" + "add x13, x13, x19, LSL #2\n" + "ldr x9, [%x[output_ptr], #0x8]\n" + "ldr x27, [%x[output_ptr], #0x10]\n" + "add x9, x9, x19, LSL #2\n" + "add x27, x27, x19, LSL #2\n" + "b 27f\n" + "26:" // Height 3: setup direct output + "mov x13, %x[output_ptr]\n" + "add x9, x13, x19, LSL #2\n" + "add x27, x9, x19, LSL #2\n" + "27:" // Height 3: Column loop + "mov x19, #0x0\n" + "whilelt p4.s, x19, x15\n" + "incw x19\n" + "whilelt p3.s, x19, x15\n" + "incw x19\n" + "whilelt p2.s, x19, x15\n" + "incw x19\n" + "whilelt p1.s, x19, x15\n" + "tbz %x[flags], #0, 28f\n" + "ld1w { z8.s }, p4/Z, [x13]\n" + "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" + "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x9]\n" + "ld1w { z13.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x9, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x27]\n" + "ld1w { z17.s }, p3/Z, [x27, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x27, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x27, #3, MUL VL]\n" + "b 29f\n" + "28:" // Height 3: no accumulate + "mov z8.s, #0x0\n" + "mov z9.s, #0x0\n" + "mov z10.s, #0x0\n" + "mov z11.s, #0x0\n" + "mov z12.s, #0x0\n" + "mov z13.s, #0x0\n" + "mov z14.s, #0x0\n" + "mov z15.s, #0x0\n" + "mov z16.s, #0x0\n" + "mov z17.s, #0x0\n" + "mov z18.s, #0x0\n" + "mov z19.s, #0x0\n" + "29:" // Height 3: setup done + "mov x12, #0x0\n" + "30:" // Height 3: String loop + "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w11, [x20, x12, LSL #0x2]\n" + "tbz %x[flags], #3, 31f\n" + "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "add x20, x20, x19, LSL #3\n" + "ldr x10, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "cbnz x12, 32f\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x10, x10, x19\n" + "add x28, x28, x19\n" + "add x26, x26, x19\n" + "b 32f\n" + "31:" // Height 3: setup direct input + "mov x10, %x[input_ptr]\n" + "add x28, x10, x19\n" + "add x26, x28, x19\n" + "32:" // Height 3: input setup done + "cmp x11, #0x10\n" + "ble 34f\n" + "33:" // Height 3: Multiply loop: Main loop head + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "sub x11, x11, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1rqb { z1.b }, p0/Z, [x28]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "ld1rqb { z2.b }, p0/Z, [x26]\n" + "add x28, x28, #0x10\n" + "udot z12.s, z6.b, z1.b[0]\n" + "prfm pldl1keep, [x10, #0x80]\n" + "add x26, x26, #0x10\n" + "udot z16.s, z6.b, z2.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "cmp x11, #0x10\n" + "udot z13.s, z7.b, z1.b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "udot z17.s, z7.b, z2.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "udot z10.s, z6.b, z0.b[0]\n" + "udot z14.s, z6.b, z1.b[0]\n" + "udot z18.s, z6.b, z2.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #4, MUL VL]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "udot z15.s, z7.b, z1.b[0]\n" + "udot z19.s, z7.b, z2.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #5, MUL VL]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "udot z12.s, z6.b, z1.b[1]\n" + "udot z16.s, z6.b, z2.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[1]\n" + "udot z13.s, z7.b, z1.b[1]\n" + "udot z17.s, z7.b, z2.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #7, MUL VL]\n" + "addvl x14, x14, #16\n" + "udot z10.s, z6.b, z0.b[1]\n" + "udot z14.s, z6.b, z1.b[1]\n" + "udot z18.s, z6.b, z2.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #-8, MUL VL]\n" + "udot z11.s, z7.b, z0.b[1]\n" + "udot z15.s, z7.b, z1.b[1]\n" + "udot z19.s, z7.b, z2.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #-7, MUL VL]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "udot z12.s, z6.b, z1.b[2]\n" + "udot z16.s, z6.b, z2.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[2]\n" + "udot z13.s, z7.b, z1.b[2]\n" + "udot z17.s, z7.b, z2.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-5, MUL VL]\n" + "udot z10.s, z6.b, z0.b[2]\n" + "udot z14.s, z6.b, z1.b[2]\n" + "udot z18.s, z6.b, z2.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-4, MUL VL]\n" + "udot z11.s, z7.b, z0.b[2]\n" + "udot z15.s, z7.b, z1.b[2]\n" + "udot z19.s, z7.b, z2.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-3, MUL VL]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "udot z12.s, z6.b, z1.b[3]\n" + "udot z16.s, z6.b, z2.b[3]\n" + "ld1b { z6.b }, p5/Z, [x14, #-2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "udot z13.s, z7.b, z1.b[3]\n" + "udot z17.s, z7.b, z2.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #-1, MUL VL]\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z14.s, z6.b, z1.b[3]\n" + "udot z18.s, z6.b, z2.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "udot z15.s, z7.b, z1.b[3]\n" + "udot z19.s, z7.b, z2.b[3]\n" + "bgt 33b\n" + "34:" // Height 3: Multiply loop: Single iteration only + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1rqb { z1.b }, p0/Z, [x28]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "ld1rqb { z2.b }, p0/Z, [x26]\n" + "add x28, x28, #0x10\n" + "udot z12.s, z6.b, z1.b[0]\n" + "add x26, x26, #0x10\n" + "udot z13.s, z7.b, z1.b[0]\n" + "udot z16.s, z6.b, z2.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z17.s, z7.b, z2.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[0]\n" + "udot z14.s, z6.b, z1.b[0]\n" + "udot z18.s, z6.b, z2.b[0]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "udot z15.s, z7.b, z1.b[0]\n" + "udot z19.s, z7.b, z2.b[0]\n" + "ble 35f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z12.s, z6.b, z1.b[1]\n" + "udot z16.s, z6.b, z2.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[1]\n" + "udot z13.s, z7.b, z1.b[1]\n" + "udot z17.s, z7.b, z2.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[1]\n" + "udot z14.s, z6.b, z1.b[1]\n" + "udot z18.s, z6.b, z2.b[1]\n" + "udot z11.s, z7.b, z0.b[1]\n" + "udot z15.s, z7.b, z1.b[1]\n" + "udot z19.s, z7.b, z2.b[1]\n" + "ble 35f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z12.s, z6.b, z1.b[2]\n" + "udot z16.s, z6.b, z2.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[2]\n" + "udot z13.s, z7.b, z1.b[2]\n" + "udot z17.s, z7.b, z2.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[2]\n" + "udot z14.s, z6.b, z1.b[2]\n" + "udot z18.s, z6.b, z2.b[2]\n" + "udot z11.s, z7.b, z0.b[2]\n" + "udot z15.s, z7.b, z1.b[2]\n" + "udot z19.s, z7.b, z2.b[2]\n" + "ble 35f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "udot z12.s, z6.b, z1.b[3]\n" + "udot z16.s, z6.b, z2.b[3]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "udot z13.s, z7.b, z1.b[3]\n" + "udot z17.s, z7.b, z2.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z14.s, z6.b, z1.b[3]\n" + "udot z18.s, z6.b, z2.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "udot z15.s, z7.b, z1.b[3]\n" + "udot z19.s, z7.b, z2.b[3]\n" + "35:" // Height 3: Multiply loop: multiply skip + "prfm pldl1keep, [x10, #0x80]\n" + "add x12, x12, #0x1\n" + "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "cmp x12, x19\n" + "bne 30b\n" + "prfm pstl1keep, [x13, #0x0]\n" + "prfm pstl1keep, [x9, #0x0]\n" + "prfm pstl1keep, [x27, #0x0]\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "st1w { z12.s }, p4, [x9]\n" + "st1w { z13.s }, p3, [x9, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z15.s }, p1, [x9, #3, MUL VL]\n" + "addvl x9, x9, #4\n" + "st1w { z16.s }, p4, [x27]\n" + "st1w { z17.s }, p3, [x27, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x27, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x27, #3, MUL VL]\n" + "addvl x27, x27, #4\n" + "36:" // Height 3: Writeback done + "mov x19, #0x0\n" + "incw x19, ALL, MUL #4\n" + "subs x15, x15, x19\n" + "bgt 27b\n" + "b 74f\n" + "37:" // Height 4 + "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "tbz %x[flags], #2, 38f\n" + "ldr x13, [%x[output_ptr], #0x0]\n" + "add x13, x13, x19, LSL #2\n" + "ldr x9, [%x[output_ptr], #0x8]\n" + "ldr x27, [%x[output_ptr], #0x10]\n" + "add x9, x9, x19, LSL #2\n" + "ldr x25, [%x[output_ptr], #0x18]\n" + "add x27, x27, x19, LSL #2\n" + "add x25, x25, x19, LSL #2\n" + "b 39f\n" + "38:" // Height 4: setup direct output + "mov x13, %x[output_ptr]\n" + "add x9, x13, x19, LSL #2\n" + "add x27, x9, x19, LSL #2\n" + "add x25, x27, x19, LSL #2\n" + "39:" // Height 4: Column loop + "mov x19, #0x0\n" + "whilelt p4.s, x19, x15\n" + "incw x19\n" + "whilelt p3.s, x19, x15\n" + "incw x19\n" + "whilelt p2.s, x19, x15\n" + "incw x19\n" + "whilelt p1.s, x19, x15\n" + "tbz %x[flags], #0, 40f\n" + "ld1w { z8.s }, p4/Z, [x13]\n" + "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" + "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x9]\n" + "ld1w { z13.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x9, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x27]\n" + "ld1w { z17.s }, p3/Z, [x27, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x27, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x27, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x25]\n" + "ld1w { z21.s }, p3/Z, [x25, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x25, #2, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x25, #3, MUL VL]\n" + "b 41f\n" + "40:" // Height 4: no accumulate + "mov z8.s, #0x0\n" + "mov z9.s, #0x0\n" + "mov z10.s, #0x0\n" + "mov z11.s, #0x0\n" + "mov z12.s, #0x0\n" + "mov z13.s, #0x0\n" + "mov z14.s, #0x0\n" + "mov z15.s, #0x0\n" + "mov z16.s, #0x0\n" + "mov z17.s, #0x0\n" + "mov z18.s, #0x0\n" + "mov z19.s, #0x0\n" + "mov z20.s, #0x0\n" + "mov z21.s, #0x0\n" + "mov z22.s, #0x0\n" + "mov z23.s, #0x0\n" + "41:" // Height 4: setup done + "mov x12, #0x0\n" + "42:" // Height 4: String loop + "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w11, [x20, x12, LSL #0x2]\n" + "tbz %x[flags], #3, 43f\n" + "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "add x20, x20, x19, LSL #3\n" + "ldr x10, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" + "cbnz x12, 44f\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x10, x10, x19\n" + "add x28, x28, x19\n" + "add x26, x26, x19\n" + "add x24, x24, x19\n" + "b 44f\n" + "43:" // Height 4: setup direct input + "mov x10, %x[input_ptr]\n" + "add x28, x10, x19\n" + "add x26, x28, x19\n" + "add x24, x26, x19\n" + "44:" // Height 4: input setup done + "cmp x11, #0x10\n" + "ble 46f\n" + "45:" // Height 4: Multiply loop: Main loop head + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "sub x11, x11, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1rqb { z1.b }, p0/Z, [x28]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "ld1rqb { z2.b }, p0/Z, [x26]\n" + "add x28, x28, #0x10\n" + "udot z12.s, z6.b, z1.b[0]\n" + "ld1rqb { z3.b }, p0/Z, [x24]\n" + "add x26, x26, #0x10\n" + "udot z16.s, z6.b, z2.b[0]\n" + "prfm pldl1keep, [x10, #0x80]\n" + "add x24, x24, #0x10\n" + "udot z13.s, z7.b, z1.b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "cmp x11, #0x10\n" + "udot z20.s, z6.b, z3.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z17.s, z7.b, z2.b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "udot z21.s, z7.b, z3.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "udot z10.s, z6.b, z0.b[0]\n" + "udot z14.s, z6.b, z1.b[0]\n" + "udot z18.s, z6.b, z2.b[0]\n" + "udot z22.s, z6.b, z3.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #4, MUL VL]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "udot z15.s, z7.b, z1.b[0]\n" + "udot z19.s, z7.b, z2.b[0]\n" + "udot z23.s, z7.b, z3.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #5, MUL VL]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "udot z12.s, z6.b, z1.b[1]\n" + "udot z16.s, z6.b, z2.b[1]\n" + "udot z20.s, z6.b, z3.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[1]\n" + "udot z13.s, z7.b, z1.b[1]\n" + "udot z17.s, z7.b, z2.b[1]\n" + "udot z21.s, z7.b, z3.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #7, MUL VL]\n" + "addvl x14, x14, #16\n" + "udot z10.s, z6.b, z0.b[1]\n" + "udot z14.s, z6.b, z1.b[1]\n" + "udot z18.s, z6.b, z2.b[1]\n" + "udot z22.s, z6.b, z3.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #-8, MUL VL]\n" + "udot z11.s, z7.b, z0.b[1]\n" + "udot z15.s, z7.b, z1.b[1]\n" + "udot z19.s, z7.b, z2.b[1]\n" + "udot z23.s, z7.b, z3.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #-7, MUL VL]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "udot z12.s, z6.b, z1.b[2]\n" + "udot z16.s, z6.b, z2.b[2]\n" + "udot z20.s, z6.b, z3.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[2]\n" + "udot z13.s, z7.b, z1.b[2]\n" + "udot z17.s, z7.b, z2.b[2]\n" + "udot z21.s, z7.b, z3.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-5, MUL VL]\n" + "udot z10.s, z6.b, z0.b[2]\n" + "udot z14.s, z6.b, z1.b[2]\n" + "udot z18.s, z6.b, z2.b[2]\n" + "udot z22.s, z6.b, z3.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-4, MUL VL]\n" + "udot z11.s, z7.b, z0.b[2]\n" + "udot z15.s, z7.b, z1.b[2]\n" + "udot z19.s, z7.b, z2.b[2]\n" + "udot z23.s, z7.b, z3.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-3, MUL VL]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "udot z12.s, z6.b, z1.b[3]\n" + "udot z16.s, z6.b, z2.b[3]\n" + "udot z20.s, z6.b, z3.b[3]\n" + "ld1b { z6.b }, p5/Z, [x14, #-2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "udot z13.s, z7.b, z1.b[3]\n" + "udot z17.s, z7.b, z2.b[3]\n" + "udot z21.s, z7.b, z3.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #-1, MUL VL]\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z14.s, z6.b, z1.b[3]\n" + "udot z18.s, z6.b, z2.b[3]\n" + "udot z22.s, z6.b, z3.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "udot z15.s, z7.b, z1.b[3]\n" + "udot z19.s, z7.b, z2.b[3]\n" + "udot z23.s, z7.b, z3.b[3]\n" + "bgt 45b\n" + "46:" // Height 4: Multiply loop: Single iteration only + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1rqb { z1.b }, p0/Z, [x28]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "ld1rqb { z2.b }, p0/Z, [x26]\n" + "add x28, x28, #0x10\n" + "udot z12.s, z6.b, z1.b[0]\n" + "ld1rqb { z3.b }, p0/Z, [x24]\n" + "add x26, x26, #0x10\n" + "udot z16.s, z6.b, z2.b[0]\n" + "add x24, x24, #0x10\n" + "udot z13.s, z7.b, z1.b[0]\n" + "udot z17.s, z7.b, z2.b[0]\n" + "udot z20.s, z6.b, z3.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z21.s, z7.b, z3.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[0]\n" + "udot z14.s, z6.b, z1.b[0]\n" + "udot z18.s, z6.b, z2.b[0]\n" + "udot z22.s, z6.b, z3.b[0]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "udot z15.s, z7.b, z1.b[0]\n" + "udot z19.s, z7.b, z2.b[0]\n" + "udot z23.s, z7.b, z3.b[0]\n" + "ble 47f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z12.s, z6.b, z1.b[1]\n" + "udot z16.s, z6.b, z2.b[1]\n" + "udot z20.s, z6.b, z3.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[1]\n" + "udot z13.s, z7.b, z1.b[1]\n" + "udot z17.s, z7.b, z2.b[1]\n" + "udot z21.s, z7.b, z3.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[1]\n" + "udot z14.s, z6.b, z1.b[1]\n" + "udot z18.s, z6.b, z2.b[1]\n" + "udot z22.s, z6.b, z3.b[1]\n" + "udot z11.s, z7.b, z0.b[1]\n" + "udot z15.s, z7.b, z1.b[1]\n" + "udot z19.s, z7.b, z2.b[1]\n" + "udot z23.s, z7.b, z3.b[1]\n" + "ble 47f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z12.s, z6.b, z1.b[2]\n" + "udot z16.s, z6.b, z2.b[2]\n" + "udot z20.s, z6.b, z3.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[2]\n" + "udot z13.s, z7.b, z1.b[2]\n" + "udot z17.s, z7.b, z2.b[2]\n" + "udot z21.s, z7.b, z3.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[2]\n" + "udot z14.s, z6.b, z1.b[2]\n" + "udot z18.s, z6.b, z2.b[2]\n" + "udot z22.s, z6.b, z3.b[2]\n" + "udot z11.s, z7.b, z0.b[2]\n" + "udot z15.s, z7.b, z1.b[2]\n" + "udot z19.s, z7.b, z2.b[2]\n" + "udot z23.s, z7.b, z3.b[2]\n" + "ble 47f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "udot z12.s, z6.b, z1.b[3]\n" + "udot z16.s, z6.b, z2.b[3]\n" + "udot z20.s, z6.b, z3.b[3]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "udot z13.s, z7.b, z1.b[3]\n" + "udot z17.s, z7.b, z2.b[3]\n" + "udot z21.s, z7.b, z3.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z14.s, z6.b, z1.b[3]\n" + "udot z18.s, z6.b, z2.b[3]\n" + "udot z22.s, z6.b, z3.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "udot z15.s, z7.b, z1.b[3]\n" + "udot z19.s, z7.b, z2.b[3]\n" + "udot z23.s, z7.b, z3.b[3]\n" + "47:" // Height 4: Multiply loop: multiply skip + "prfm pldl1keep, [x10, #0x80]\n" + "add x12, x12, #0x1\n" + "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "cmp x12, x19\n" + "bne 42b\n" + "prfm pstl1keep, [x13, #0x0]\n" + "prfm pstl1keep, [x9, #0x0]\n" + "prfm pstl1keep, [x27, #0x0]\n" + "prfm pstl1keep, [x25, #0x0]\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "st1w { z12.s }, p4, [x9]\n" + "st1w { z13.s }, p3, [x9, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z15.s }, p1, [x9, #3, MUL VL]\n" + "addvl x9, x9, #4\n" + "st1w { z16.s }, p4, [x27]\n" + "st1w { z17.s }, p3, [x27, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x27, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x27, #3, MUL VL]\n" + "addvl x27, x27, #4\n" + "st1w { z20.s }, p4, [x25]\n" + "st1w { z21.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z22.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z23.s }, p1, [x25, #3, MUL VL]\n" + "addvl x25, x25, #4\n" + "48:" // Height 4: Writeback done + "mov x19, #0x0\n" + "incw x19, ALL, MUL #4\n" + "subs x15, x15, x19\n" + "bgt 39b\n" + "b 74f\n" + "49:" // Height 5 + "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "tbz %x[flags], #2, 50f\n" + "ldr x13, [%x[output_ptr], #0x0]\n" + "add x13, x13, x19, LSL #2\n" + "ldr x9, [%x[output_ptr], #0x8]\n" + "ldr x27, [%x[output_ptr], #0x10]\n" + "add x9, x9, x19, LSL #2\n" + "ldr x25, [%x[output_ptr], #0x18]\n" + "ldr x23, [%x[output_ptr], #0x20]\n" + "add x27, x27, x19, LSL #2\n" + "add x25, x25, x19, LSL #2\n" + "add x23, x23, x19, LSL #2\n" + "b 51f\n" + "50:" // Height 5: setup direct output + "mov x13, %x[output_ptr]\n" + "add x9, x13, x19, LSL #2\n" + "add x27, x9, x19, LSL #2\n" + "add x25, x27, x19, LSL #2\n" + "add x23, x25, x19, LSL #2\n" + "51:" // Height 5: Column loop + "mov x19, #0x0\n" + "whilelt p4.s, x19, x15\n" + "incw x19\n" + "whilelt p3.s, x19, x15\n" + "incw x19\n" + "whilelt p2.s, x19, x15\n" + "incw x19\n" + "whilelt p1.s, x19, x15\n" + "tbz %x[flags], #0, 52f\n" + "ld1w { z8.s }, p4/Z, [x13]\n" + "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" + "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x9]\n" + "ld1w { z13.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x9, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x27]\n" + "ld1w { z17.s }, p3/Z, [x27, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x27, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x27, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x25]\n" + "ld1w { z21.s }, p3/Z, [x25, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x25, #2, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x25, #3, MUL VL]\n" + "ld1w { z24.s }, p4/Z, [x23]\n" + "ld1w { z25.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z26.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z27.s }, p1/Z, [x23, #3, MUL VL]\n" + "b 53f\n" + "52:" // Height 5: no accumulate + "mov z8.s, #0x0\n" + "mov z9.s, #0x0\n" + "mov z10.s, #0x0\n" + "mov z11.s, #0x0\n" + "mov z12.s, #0x0\n" + "mov z13.s, #0x0\n" + "mov z14.s, #0x0\n" + "mov z15.s, #0x0\n" + "mov z16.s, #0x0\n" + "mov z17.s, #0x0\n" + "mov z18.s, #0x0\n" + "mov z19.s, #0x0\n" + "mov z20.s, #0x0\n" + "mov z21.s, #0x0\n" + "mov z22.s, #0x0\n" + "mov z23.s, #0x0\n" + "mov z24.s, #0x0\n" + "mov z25.s, #0x0\n" + "mov z26.s, #0x0\n" + "mov z27.s, #0x0\n" + "53:" // Height 5: setup done + "mov x12, #0x0\n" + "54:" // Height 5: String loop + "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w11, [x20, x12, LSL #0x2]\n" + "tbz %x[flags], #3, 55f\n" + "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "add x20, x20, x19, LSL #3\n" + "ldr x10, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" + "cbnz x12, 56f\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x10, x10, x19\n" + "add x28, x28, x19\n" + "add x26, x26, x19\n" + "add x24, x24, x19\n" + "add x22, x22, x19\n" + "b 56f\n" + "55:" // Height 5: setup direct input + "mov x10, %x[input_ptr]\n" + "add x28, x10, x19\n" + "add x26, x28, x19\n" + "add x24, x26, x19\n" + "add x22, x24, x19\n" + "56:" // Height 5: input setup done + "cmp x11, #0x10\n" + "ble 58f\n" + "57:" // Height 5: Multiply loop: Main loop head + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "sub x11, x11, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1rqb { z1.b }, p0/Z, [x28]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "ld1rqb { z2.b }, p0/Z, [x26]\n" + "add x28, x28, #0x10\n" + "udot z12.s, z6.b, z1.b[0]\n" + "ld1rqb { z3.b }, p0/Z, [x24]\n" + "add x26, x26, #0x10\n" + "udot z16.s, z6.b, z2.b[0]\n" + "ld1rqb { z4.b }, p0/Z, [x22]\n" + "add x24, x24, #0x10\n" + "udot z13.s, z7.b, z1.b[0]\n" + "prfm pldl1keep, [x10, #0x80]\n" + "add x22, x22, #0x10\n" + "udot z20.s, z6.b, z3.b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "cmp x11, #0x10\n" + "udot z24.s, z6.b, z4.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z17.s, z7.b, z2.b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "udot z21.s, z7.b, z3.b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "udot z25.s, z7.b, z4.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "udot z10.s, z6.b, z0.b[0]\n" + "prfm pldl1keep, [x22, #0x80]\n" + "udot z14.s, z6.b, z1.b[0]\n" + "udot z18.s, z6.b, z2.b[0]\n" + "udot z22.s, z6.b, z3.b[0]\n" + "udot z26.s, z6.b, z4.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #4, MUL VL]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "udot z15.s, z7.b, z1.b[0]\n" + "udot z19.s, z7.b, z2.b[0]\n" + "udot z23.s, z7.b, z3.b[0]\n" + "udot z27.s, z7.b, z4.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #5, MUL VL]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "udot z12.s, z6.b, z1.b[1]\n" + "udot z16.s, z6.b, z2.b[1]\n" + "udot z20.s, z6.b, z3.b[1]\n" + "udot z24.s, z6.b, z4.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[1]\n" + "udot z13.s, z7.b, z1.b[1]\n" + "udot z17.s, z7.b, z2.b[1]\n" + "udot z21.s, z7.b, z3.b[1]\n" + "udot z25.s, z7.b, z4.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #7, MUL VL]\n" + "addvl x14, x14, #16\n" + "udot z10.s, z6.b, z0.b[1]\n" + "udot z14.s, z6.b, z1.b[1]\n" + "udot z18.s, z6.b, z2.b[1]\n" + "udot z22.s, z6.b, z3.b[1]\n" + "udot z26.s, z6.b, z4.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #-8, MUL VL]\n" + "udot z11.s, z7.b, z0.b[1]\n" + "udot z15.s, z7.b, z1.b[1]\n" + "udot z19.s, z7.b, z2.b[1]\n" + "udot z23.s, z7.b, z3.b[1]\n" + "udot z27.s, z7.b, z4.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #-7, MUL VL]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "udot z12.s, z6.b, z1.b[2]\n" + "udot z16.s, z6.b, z2.b[2]\n" + "udot z20.s, z6.b, z3.b[2]\n" + "udot z24.s, z6.b, z4.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[2]\n" + "udot z13.s, z7.b, z1.b[2]\n" + "udot z17.s, z7.b, z2.b[2]\n" + "udot z21.s, z7.b, z3.b[2]\n" + "udot z25.s, z7.b, z4.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-5, MUL VL]\n" + "udot z10.s, z6.b, z0.b[2]\n" + "udot z14.s, z6.b, z1.b[2]\n" + "udot z18.s, z6.b, z2.b[2]\n" + "udot z22.s, z6.b, z3.b[2]\n" + "udot z26.s, z6.b, z4.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-4, MUL VL]\n" + "udot z11.s, z7.b, z0.b[2]\n" + "udot z15.s, z7.b, z1.b[2]\n" + "udot z19.s, z7.b, z2.b[2]\n" + "udot z23.s, z7.b, z3.b[2]\n" + "udot z27.s, z7.b, z4.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-3, MUL VL]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "udot z12.s, z6.b, z1.b[3]\n" + "udot z16.s, z6.b, z2.b[3]\n" + "udot z20.s, z6.b, z3.b[3]\n" + "udot z24.s, z6.b, z4.b[3]\n" + "ld1b { z6.b }, p5/Z, [x14, #-2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "udot z13.s, z7.b, z1.b[3]\n" + "udot z17.s, z7.b, z2.b[3]\n" + "udot z21.s, z7.b, z3.b[3]\n" + "udot z25.s, z7.b, z4.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #-1, MUL VL]\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z14.s, z6.b, z1.b[3]\n" + "udot z18.s, z6.b, z2.b[3]\n" + "udot z22.s, z6.b, z3.b[3]\n" + "udot z26.s, z6.b, z4.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "udot z15.s, z7.b, z1.b[3]\n" + "udot z19.s, z7.b, z2.b[3]\n" + "udot z23.s, z7.b, z3.b[3]\n" + "udot z27.s, z7.b, z4.b[3]\n" + "bgt 57b\n" + "58:" // Height 5: Multiply loop: Single iteration only + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1rqb { z1.b }, p0/Z, [x28]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "ld1rqb { z2.b }, p0/Z, [x26]\n" + "add x28, x28, #0x10\n" + "udot z12.s, z6.b, z1.b[0]\n" + "ld1rqb { z3.b }, p0/Z, [x24]\n" + "add x26, x26, #0x10\n" + "udot z16.s, z6.b, z2.b[0]\n" + "ld1rqb { z4.b }, p0/Z, [x22]\n" + "add x24, x24, #0x10\n" + "udot z13.s, z7.b, z1.b[0]\n" + "add x22, x22, #0x10\n" + "udot z17.s, z7.b, z2.b[0]\n" + "udot z20.s, z6.b, z3.b[0]\n" + "udot z24.s, z6.b, z4.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z21.s, z7.b, z3.b[0]\n" + "udot z25.s, z7.b, z4.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[0]\n" + "udot z14.s, z6.b, z1.b[0]\n" + "udot z18.s, z6.b, z2.b[0]\n" + "udot z22.s, z6.b, z3.b[0]\n" + "udot z26.s, z6.b, z4.b[0]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "udot z15.s, z7.b, z1.b[0]\n" + "udot z19.s, z7.b, z2.b[0]\n" + "udot z23.s, z7.b, z3.b[0]\n" + "udot z27.s, z7.b, z4.b[0]\n" + "ble 59f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z12.s, z6.b, z1.b[1]\n" + "udot z16.s, z6.b, z2.b[1]\n" + "udot z20.s, z6.b, z3.b[1]\n" + "udot z24.s, z6.b, z4.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[1]\n" + "udot z13.s, z7.b, z1.b[1]\n" + "udot z17.s, z7.b, z2.b[1]\n" + "udot z21.s, z7.b, z3.b[1]\n" + "udot z25.s, z7.b, z4.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[1]\n" + "udot z14.s, z6.b, z1.b[1]\n" + "udot z18.s, z6.b, z2.b[1]\n" + "udot z22.s, z6.b, z3.b[1]\n" + "udot z26.s, z6.b, z4.b[1]\n" + "udot z11.s, z7.b, z0.b[1]\n" + "udot z15.s, z7.b, z1.b[1]\n" + "udot z19.s, z7.b, z2.b[1]\n" + "udot z23.s, z7.b, z3.b[1]\n" + "udot z27.s, z7.b, z4.b[1]\n" + "ble 59f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z12.s, z6.b, z1.b[2]\n" + "udot z16.s, z6.b, z2.b[2]\n" + "udot z20.s, z6.b, z3.b[2]\n" + "udot z24.s, z6.b, z4.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[2]\n" + "udot z13.s, z7.b, z1.b[2]\n" + "udot z17.s, z7.b, z2.b[2]\n" + "udot z21.s, z7.b, z3.b[2]\n" + "udot z25.s, z7.b, z4.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[2]\n" + "udot z14.s, z6.b, z1.b[2]\n" + "udot z18.s, z6.b, z2.b[2]\n" + "udot z22.s, z6.b, z3.b[2]\n" + "udot z26.s, z6.b, z4.b[2]\n" + "udot z11.s, z7.b, z0.b[2]\n" + "udot z15.s, z7.b, z1.b[2]\n" + "udot z19.s, z7.b, z2.b[2]\n" + "udot z23.s, z7.b, z3.b[2]\n" + "udot z27.s, z7.b, z4.b[2]\n" + "ble 59f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "udot z12.s, z6.b, z1.b[3]\n" + "udot z16.s, z6.b, z2.b[3]\n" + "udot z20.s, z6.b, z3.b[3]\n" + "udot z24.s, z6.b, z4.b[3]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "udot z13.s, z7.b, z1.b[3]\n" + "udot z17.s, z7.b, z2.b[3]\n" + "udot z21.s, z7.b, z3.b[3]\n" + "udot z25.s, z7.b, z4.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z14.s, z6.b, z1.b[3]\n" + "udot z18.s, z6.b, z2.b[3]\n" + "udot z22.s, z6.b, z3.b[3]\n" + "udot z26.s, z6.b, z4.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "udot z15.s, z7.b, z1.b[3]\n" + "udot z19.s, z7.b, z2.b[3]\n" + "udot z23.s, z7.b, z3.b[3]\n" + "udot z27.s, z7.b, z4.b[3]\n" + "59:" // Height 5: Multiply loop: multiply skip + "prfm pldl1keep, [x10, #0x80]\n" + "add x12, x12, #0x1\n" + "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" + "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "cmp x12, x19\n" + "bne 54b\n" + "prfm pstl1keep, [x13, #0x0]\n" + "prfm pstl1keep, [x9, #0x0]\n" + "prfm pstl1keep, [x27, #0x0]\n" + "prfm pstl1keep, [x25, #0x0]\n" + "prfm pstl1keep, [x23, #0x0]\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "st1w { z12.s }, p4, [x9]\n" + "st1w { z13.s }, p3, [x9, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z15.s }, p1, [x9, #3, MUL VL]\n" + "addvl x9, x9, #4\n" + "st1w { z16.s }, p4, [x27]\n" + "st1w { z17.s }, p3, [x27, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x27, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x27, #3, MUL VL]\n" + "addvl x27, x27, #4\n" + "st1w { z20.s }, p4, [x25]\n" + "st1w { z21.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z22.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z23.s }, p1, [x25, #3, MUL VL]\n" + "addvl x25, x25, #4\n" + "st1w { z24.s }, p4, [x23]\n" + "st1w { z25.s }, p3, [x23, #1, MUL VL]\n" + "st1w { z26.s }, p2, [x23, #2, MUL VL]\n" + "st1w { z27.s }, p1, [x23, #3, MUL VL]\n" + "addvl x23, x23, #4\n" + "60:" // Height 5: Writeback done + "mov x19, #0x0\n" + "incw x19, ALL, MUL #4\n" + "subs x15, x15, x19\n" + "bgt 51b\n" + "b 74f\n" + "61:" // Height 6 + "ldr x15, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x14, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n" + "tbz %x[flags], #2, 62f\n" + "ldr x13, [%x[output_ptr], #0x0]\n" + "add x13, x13, x19, LSL #2\n" + "ldr x9, [%x[output_ptr], #0x8]\n" + "ldr x27, [%x[output_ptr], #0x10]\n" + "add x9, x9, x19, LSL #2\n" + "ldr x25, [%x[output_ptr], #0x18]\n" + "ldr x23, [%x[output_ptr], #0x20]\n" + "add x27, x27, x19, LSL #2\n" + "ldr x21, [%x[output_ptr], #0x28]\n" + "add %x[output_ptr], %x[output_ptr], #0x30\n" + "add x25, x25, x19, LSL #2\n" + "add x23, x23, x19, LSL #2\n" + "add x21, x21, x19, LSL #2\n" + "b 63f\n" + "62:" // Height 6: setup direct output + "mov x13, %x[output_ptr]\n" + "add x9, x13, x19, LSL #2\n" + "add x27, x9, x19, LSL #2\n" + "add x25, x27, x19, LSL #2\n" + "add x23, x25, x19, LSL #2\n" + "add x21, x23, x19, LSL #2\n" + "add %x[output_ptr], x21, x19, LSL #2\n" + "63:" // Height 6: Column loop + "mov x19, #0x0\n" + "whilelt p4.s, x19, x15\n" + "incw x19\n" + "whilelt p3.s, x19, x15\n" + "incw x19\n" + "whilelt p2.s, x19, x15\n" + "incw x19\n" + "whilelt p1.s, x19, x15\n" + "tbz %x[flags], #0, 64f\n" + "ld1w { z8.s }, p4/Z, [x13]\n" + "ld1w { z9.s }, p3/Z, [x13, #1, MUL VL]\n" + "ld1w { z10.s }, p2/Z, [x13, #2, MUL VL]\n" + "ld1w { z11.s }, p1/Z, [x13, #3, MUL VL]\n" + "ld1w { z12.s }, p4/Z, [x9]\n" + "ld1w { z13.s }, p3/Z, [x9, #1, MUL VL]\n" + "ld1w { z14.s }, p2/Z, [x9, #2, MUL VL]\n" + "ld1w { z15.s }, p1/Z, [x9, #3, MUL VL]\n" + "ld1w { z16.s }, p4/Z, [x27]\n" + "ld1w { z17.s }, p3/Z, [x27, #1, MUL VL]\n" + "ld1w { z18.s }, p2/Z, [x27, #2, MUL VL]\n" + "ld1w { z19.s }, p1/Z, [x27, #3, MUL VL]\n" + "ld1w { z20.s }, p4/Z, [x25]\n" + "ld1w { z21.s }, p3/Z, [x25, #1, MUL VL]\n" + "ld1w { z22.s }, p2/Z, [x25, #2, MUL VL]\n" + "ld1w { z23.s }, p1/Z, [x25, #3, MUL VL]\n" + "ld1w { z24.s }, p4/Z, [x23]\n" + "ld1w { z25.s }, p3/Z, [x23, #1, MUL VL]\n" + "ld1w { z26.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z27.s }, p1/Z, [x23, #3, MUL VL]\n" + "ld1w { z28.s }, p4/Z, [x21]\n" + "ld1w { z29.s }, p3/Z, [x21, #1, MUL VL]\n" + "ld1w { z30.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z31.s }, p1/Z, [x21, #3, MUL VL]\n" + "b 65f\n" + "64:" // Height 6: no accumulate + "mov z8.s, #0x0\n" + "mov z9.s, #0x0\n" + "mov z10.s, #0x0\n" + "mov z11.s, #0x0\n" + "mov z12.s, #0x0\n" + "mov z13.s, #0x0\n" + "mov z14.s, #0x0\n" + "mov z15.s, #0x0\n" + "mov z16.s, #0x0\n" + "mov z17.s, #0x0\n" + "mov z18.s, #0x0\n" + "mov z19.s, #0x0\n" + "mov z20.s, #0x0\n" + "mov z21.s, #0x0\n" + "mov z22.s, #0x0\n" + "mov z23.s, #0x0\n" + "mov z24.s, #0x0\n" + "mov z25.s, #0x0\n" + "mov z26.s, #0x0\n" + "mov z27.s, #0x0\n" + "mov z28.s, #0x0\n" + "mov z29.s, #0x0\n" + "mov z30.s, #0x0\n" + "mov z31.s, #0x0\n" + "65:" // Height 6: setup done + "mov x12, #0x0\n" + "66:" // Height 6: String loop + "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w11, [x20, x12, LSL #0x2]\n" + "tbz %x[flags], #3, 67f\n" + "ldr x20, [%x[input_ptr], x12, LSL #0x3]\n" + "add x20, x20, x19, LSL #3\n" + "ldr x10, [x20, #0x0]\n" + "ldr x28, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x24, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" + "ldr x20, [x20, #0x28]\n" + "cbnz x12, 68f\n" + "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x10, x10, x19\n" + "add x28, x28, x19\n" + "add x26, x26, x19\n" + "add x24, x24, x19\n" + "add x22, x22, x19\n" + "add x20, x20, x19\n" + "b 68f\n" + "67:" // Height 6: setup direct input + "mov x10, %x[input_ptr]\n" + "add x28, x10, x19\n" + "add x26, x28, x19\n" + "add x24, x26, x19\n" + "add x22, x24, x19\n" + "add x20, x22, x19\n" + "68:" // Height 6: input setup done + "cmp x11, #0x10\n" + "ble 70f\n" + "69:" // Height 6: Multiply loop: Main loop head + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "sub x11, x11, #0x10\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1rqb { z1.b }, p0/Z, [x28]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "ld1rqb { z2.b }, p0/Z, [x26]\n" + "add x28, x28, #0x10\n" + "udot z12.s, z6.b, z1.b[0]\n" + "ld1rqb { z3.b }, p0/Z, [x24]\n" + "add x26, x26, #0x10\n" + "udot z16.s, z6.b, z2.b[0]\n" + "ld1rqb { z4.b }, p0/Z, [x22]\n" + "add x24, x24, #0x10\n" + "udot z13.s, z7.b, z1.b[0]\n" + "ld1rqb { z5.b }, p0/Z, [x20]\n" + "add x22, x22, #0x10\n" + "udot z20.s, z6.b, z3.b[0]\n" + "prfm pldl1keep, [x10, #0x80]\n" + "add x20, x20, #0x10\n" + "udot z24.s, z6.b, z4.b[0]\n" + "prfm pldl1keep, [x28, #0x80]\n" + "cmp x11, #0x10\n" + "udot z28.s, z6.b, z5.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z17.s, z7.b, z2.b[0]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "udot z21.s, z7.b, z3.b[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "udot z25.s, z7.b, z4.b[0]\n" + "prfm pldl1keep, [x22, #0x80]\n" + "udot z29.s, z7.b, z5.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "udot z10.s, z6.b, z0.b[0]\n" + "prfm pldl1keep, [x20, #0x80]\n" + "udot z14.s, z6.b, z1.b[0]\n" + "udot z18.s, z6.b, z2.b[0]\n" + "udot z22.s, z6.b, z3.b[0]\n" + "udot z26.s, z6.b, z4.b[0]\n" + "udot z30.s, z6.b, z5.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #4, MUL VL]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "udot z15.s, z7.b, z1.b[0]\n" + "udot z19.s, z7.b, z2.b[0]\n" + "udot z23.s, z7.b, z3.b[0]\n" + "udot z27.s, z7.b, z4.b[0]\n" + "udot z31.s, z7.b, z5.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #5, MUL VL]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "udot z12.s, z6.b, z1.b[1]\n" + "udot z16.s, z6.b, z2.b[1]\n" + "udot z20.s, z6.b, z3.b[1]\n" + "udot z24.s, z6.b, z4.b[1]\n" + "udot z28.s, z6.b, z5.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[1]\n" + "udot z13.s, z7.b, z1.b[1]\n" + "udot z17.s, z7.b, z2.b[1]\n" + "udot z21.s, z7.b, z3.b[1]\n" + "udot z25.s, z7.b, z4.b[1]\n" + "udot z29.s, z7.b, z5.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #7, MUL VL]\n" + "addvl x14, x14, #16\n" + "udot z10.s, z6.b, z0.b[1]\n" + "udot z14.s, z6.b, z1.b[1]\n" + "udot z18.s, z6.b, z2.b[1]\n" + "udot z22.s, z6.b, z3.b[1]\n" + "udot z26.s, z6.b, z4.b[1]\n" + "udot z30.s, z6.b, z5.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #-8, MUL VL]\n" + "udot z11.s, z7.b, z0.b[1]\n" + "udot z15.s, z7.b, z1.b[1]\n" + "udot z19.s, z7.b, z2.b[1]\n" + "udot z23.s, z7.b, z3.b[1]\n" + "udot z27.s, z7.b, z4.b[1]\n" + "udot z31.s, z7.b, z5.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #-7, MUL VL]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "udot z12.s, z6.b, z1.b[2]\n" + "udot z16.s, z6.b, z2.b[2]\n" + "udot z20.s, z6.b, z3.b[2]\n" + "udot z24.s, z6.b, z4.b[2]\n" + "udot z28.s, z6.b, z5.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-6, MUL VL]\n" + "udot z9.s, z7.b, z0.b[2]\n" + "udot z13.s, z7.b, z1.b[2]\n" + "udot z17.s, z7.b, z2.b[2]\n" + "udot z21.s, z7.b, z3.b[2]\n" + "udot z25.s, z7.b, z4.b[2]\n" + "udot z29.s, z7.b, z5.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-5, MUL VL]\n" + "udot z10.s, z6.b, z0.b[2]\n" + "udot z14.s, z6.b, z1.b[2]\n" + "udot z18.s, z6.b, z2.b[2]\n" + "udot z22.s, z6.b, z3.b[2]\n" + "udot z26.s, z6.b, z4.b[2]\n" + "udot z30.s, z6.b, z5.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #-4, MUL VL]\n" + "udot z11.s, z7.b, z0.b[2]\n" + "udot z15.s, z7.b, z1.b[2]\n" + "udot z19.s, z7.b, z2.b[2]\n" + "udot z23.s, z7.b, z3.b[2]\n" + "udot z27.s, z7.b, z4.b[2]\n" + "udot z31.s, z7.b, z5.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #-3, MUL VL]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "udot z12.s, z6.b, z1.b[3]\n" + "udot z16.s, z6.b, z2.b[3]\n" + "udot z20.s, z6.b, z3.b[3]\n" + "udot z24.s, z6.b, z4.b[3]\n" + "udot z28.s, z6.b, z5.b[3]\n" + "ld1b { z6.b }, p5/Z, [x14, #-2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "udot z13.s, z7.b, z1.b[3]\n" + "udot z17.s, z7.b, z2.b[3]\n" + "udot z21.s, z7.b, z3.b[3]\n" + "udot z25.s, z7.b, z4.b[3]\n" + "udot z29.s, z7.b, z5.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #-1, MUL VL]\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z14.s, z6.b, z1.b[3]\n" + "udot z18.s, z6.b, z2.b[3]\n" + "udot z22.s, z6.b, z3.b[3]\n" + "udot z26.s, z6.b, z4.b[3]\n" + "udot z30.s, z6.b, z5.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "udot z15.s, z7.b, z1.b[3]\n" + "udot z19.s, z7.b, z2.b[3]\n" + "udot z23.s, z7.b, z3.b[3]\n" + "udot z27.s, z7.b, z4.b[3]\n" + "udot z31.s, z7.b, z5.b[3]\n" + "bgt 69b\n" + "70:" // Height 6: Multiply loop: Single iteration only + "ld1b { z6.b }, p5/Z, [x14]\n" + "whilelt p0.b, XZR, x11\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "ld1rqb { z0.b }, p0/Z, [x10]\n" + "udot z8.s, z6.b, z0.b[0]\n" + "ld1rqb { z1.b }, p0/Z, [x28]\n" + "add x10, x10, #0x10\n" + "udot z9.s, z7.b, z0.b[0]\n" + "ld1rqb { z2.b }, p0/Z, [x26]\n" + "add x28, x28, #0x10\n" + "udot z12.s, z6.b, z1.b[0]\n" + "ld1rqb { z3.b }, p0/Z, [x24]\n" + "add x26, x26, #0x10\n" + "udot z16.s, z6.b, z2.b[0]\n" + "ld1rqb { z4.b }, p0/Z, [x22]\n" + "add x24, x24, #0x10\n" + "udot z13.s, z7.b, z1.b[0]\n" + "ld1rqb { z5.b }, p0/Z, [x20]\n" + "add x22, x22, #0x10\n" + "udot z20.s, z6.b, z3.b[0]\n" + "add x20, x20, #0x10\n" + "udot z17.s, z7.b, z2.b[0]\n" + "udot z24.s, z6.b, z4.b[0]\n" + "udot z28.s, z6.b, z5.b[0]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z21.s, z7.b, z3.b[0]\n" + "udot z25.s, z7.b, z4.b[0]\n" + "udot z29.s, z7.b, z5.b[0]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[0]\n" + "udot z14.s, z6.b, z1.b[0]\n" + "udot z18.s, z6.b, z2.b[0]\n" + "udot z22.s, z6.b, z3.b[0]\n" + "udot z26.s, z6.b, z4.b[0]\n" + "udot z30.s, z6.b, z5.b[0]\n" + "udot z11.s, z7.b, z0.b[0]\n" + "udot z15.s, z7.b, z1.b[0]\n" + "udot z19.s, z7.b, z2.b[0]\n" + "udot z23.s, z7.b, z3.b[0]\n" + "udot z27.s, z7.b, z4.b[0]\n" + "udot z31.s, z7.b, z5.b[0]\n" + "ble 71f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z12.s, z6.b, z1.b[1]\n" + "udot z16.s, z6.b, z2.b[1]\n" + "udot z20.s, z6.b, z3.b[1]\n" + "udot z24.s, z6.b, z4.b[1]\n" + "udot z28.s, z6.b, z5.b[1]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[1]\n" + "udot z13.s, z7.b, z1.b[1]\n" + "udot z17.s, z7.b, z2.b[1]\n" + "udot z21.s, z7.b, z3.b[1]\n" + "udot z25.s, z7.b, z4.b[1]\n" + "udot z29.s, z7.b, z5.b[1]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[1]\n" + "udot z14.s, z6.b, z1.b[1]\n" + "udot z18.s, z6.b, z2.b[1]\n" + "udot z22.s, z6.b, z3.b[1]\n" + "udot z26.s, z6.b, z4.b[1]\n" + "udot z30.s, z6.b, z5.b[1]\n" + "udot z11.s, z7.b, z0.b[1]\n" + "udot z15.s, z7.b, z1.b[1]\n" + "udot z19.s, z7.b, z2.b[1]\n" + "udot z23.s, z7.b, z3.b[1]\n" + "udot z27.s, z7.b, z4.b[1]\n" + "udot z31.s, z7.b, z5.b[1]\n" + "ble 71f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "subs x11, x11, #0x4\n" + "udot z12.s, z6.b, z1.b[2]\n" + "udot z16.s, z6.b, z2.b[2]\n" + "udot z20.s, z6.b, z3.b[2]\n" + "udot z24.s, z6.b, z4.b[2]\n" + "udot z28.s, z6.b, z5.b[2]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[2]\n" + "udot z13.s, z7.b, z1.b[2]\n" + "udot z17.s, z7.b, z2.b[2]\n" + "udot z21.s, z7.b, z3.b[2]\n" + "udot z25.s, z7.b, z4.b[2]\n" + "udot z29.s, z7.b, z5.b[2]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[2]\n" + "udot z14.s, z6.b, z1.b[2]\n" + "udot z18.s, z6.b, z2.b[2]\n" + "udot z22.s, z6.b, z3.b[2]\n" + "udot z26.s, z6.b, z4.b[2]\n" + "udot z30.s, z6.b, z5.b[2]\n" + "udot z11.s, z7.b, z0.b[2]\n" + "udot z15.s, z7.b, z1.b[2]\n" + "udot z19.s, z7.b, z2.b[2]\n" + "udot z23.s, z7.b, z3.b[2]\n" + "udot z27.s, z7.b, z4.b[2]\n" + "udot z31.s, z7.b, z5.b[2]\n" + "ble 71f\n" + "ld1b { z6.b }, p5/Z, [x14]\n" + "udot z8.s, z6.b, z0.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #1, MUL VL]\n" + "udot z12.s, z6.b, z1.b[3]\n" + "udot z16.s, z6.b, z2.b[3]\n" + "udot z20.s, z6.b, z3.b[3]\n" + "udot z24.s, z6.b, z4.b[3]\n" + "udot z28.s, z6.b, z5.b[3]\n" + "ld1b { z6.b }, p5/Z, [x14, #2, MUL VL]\n" + "udot z9.s, z7.b, z0.b[3]\n" + "udot z13.s, z7.b, z1.b[3]\n" + "udot z17.s, z7.b, z2.b[3]\n" + "udot z21.s, z7.b, z3.b[3]\n" + "udot z25.s, z7.b, z4.b[3]\n" + "udot z29.s, z7.b, z5.b[3]\n" + "ld1b { z7.b }, p5/Z, [x14, #3, MUL VL]\n" + "addvl x14, x14, #4\n" + "udot z10.s, z6.b, z0.b[3]\n" + "udot z14.s, z6.b, z1.b[3]\n" + "udot z18.s, z6.b, z2.b[3]\n" + "udot z22.s, z6.b, z3.b[3]\n" + "udot z26.s, z6.b, z4.b[3]\n" + "udot z30.s, z6.b, z5.b[3]\n" + "udot z11.s, z7.b, z0.b[3]\n" + "udot z15.s, z7.b, z1.b[3]\n" + "udot z19.s, z7.b, z2.b[3]\n" + "udot z23.s, z7.b, z3.b[3]\n" + "udot z27.s, z7.b, z4.b[3]\n" + "udot z31.s, z7.b, z5.b[3]\n" + "71:" // Height 6: Multiply loop: multiply skip + "prfm pldl1keep, [x10, #0x80]\n" + "add x12, x12, #0x1\n" + "prfm pldl1keep, [x28, #0x80]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" + "prfm pldl1keep, [x20, #0x80]\n" + "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n" + "cmp x12, x19\n" + "bne 66b\n" + "prfm pstl1keep, [x13, #0x0]\n" + "prfm pstl1keep, [x9, #0x0]\n" + "prfm pstl1keep, [x27, #0x0]\n" + "prfm pstl1keep, [x25, #0x0]\n" + "prfm pstl1keep, [x23, #0x0]\n" + "prfm pstl1keep, [x21, #0x0]\n" + "st1w { z8.s }, p4, [x13]\n" + "st1w { z9.s }, p3, [x13, #1, MUL VL]\n" + "st1w { z10.s }, p2, [x13, #2, MUL VL]\n" + "st1w { z11.s }, p1, [x13, #3, MUL VL]\n" + "addvl x13, x13, #4\n" + "st1w { z12.s }, p4, [x9]\n" + "st1w { z13.s }, p3, [x9, #1, MUL VL]\n" + "st1w { z14.s }, p2, [x9, #2, MUL VL]\n" + "st1w { z15.s }, p1, [x9, #3, MUL VL]\n" + "addvl x9, x9, #4\n" + "st1w { z16.s }, p4, [x27]\n" + "st1w { z17.s }, p3, [x27, #1, MUL VL]\n" + "st1w { z18.s }, p2, [x27, #2, MUL VL]\n" + "st1w { z19.s }, p1, [x27, #3, MUL VL]\n" + "addvl x27, x27, #4\n" + "st1w { z20.s }, p4, [x25]\n" + "st1w { z21.s }, p3, [x25, #1, MUL VL]\n" + "st1w { z22.s }, p2, [x25, #2, MUL VL]\n" + "st1w { z23.s }, p1, [x25, #3, MUL VL]\n" + "addvl x25, x25, #4\n" + "st1w { z24.s }, p4, [x23]\n" + "st1w { z25.s }, p3, [x23, #1, MUL VL]\n" + "st1w { z26.s }, p2, [x23, #2, MUL VL]\n" + "st1w { z27.s }, p1, [x23, #3, MUL VL]\n" + "addvl x23, x23, #4\n" + "st1w { z28.s }, p4, [x21]\n" + "st1w { z29.s }, p3, [x21, #1, MUL VL]\n" + "st1w { z30.s }, p2, [x21, #2, MUL VL]\n" + "st1w { z31.s }, p1, [x21, #3, MUL VL]\n" + "addvl x21, x21, #4\n" + "72:" // Height 6: Writeback done + "mov x19, #0x0\n" + "incw x19, ALL, MUL #4\n" + "subs x15, x15, x19\n" + "bgt 63b\n" + "subs %x[M], %x[M], #0x6\n" + "beq 74f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "tbz %x[flags], #3, 73f\n" + "add x20, x20, #0x6\n" + "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "b 1b\n" + "73:" // Update direct input + "mov x19, #0x6\n" + "madd %x[input_ptr], x19, x20, %x[input_ptr]\n" + "b 1b\n" + "74:" // Exit + + : [M] "+r" (M), [input_ptr] "+r" (input_ptr), [output_ptr] "+r" (output_ptr) + : [args_ptr] "r" (&ka), [flags] "r" (flags), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) + : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31" + ); +} + +} // namespace arm_gemm +#endif // __ARM_FEATURE_SVE -- cgit v1.2.1