From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 12 Apr 2023 17:43:17 +0100 Subject: Update CPU kernel implementations and guard directives Resolves COMPMID-6023 Change-Id: I868975d14c4f98af6716726feda22405a6a4c891 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686 Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../kernels/a64_hybrid_fp32_mla_8x4/a55.cpp | 2786 ++++++++++---------- .../kernels/a64_hybrid_fp32_mla_8x4/generic.cpp | 338 ++- 2 files changed, 1560 insertions(+), 1564 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp index 67e0c1e8cc..236865315e 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/a55.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021, 2023 Arm Limited. + * Copyright (c) 2022-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -92,7 +92,6 @@ void a64_hybrid_fp32_mla_8x4_a55 ( break; } __asm__ __volatile__( - "1:" // Row loop "cmp %x[M], #0x8\n" "bge 148f\n" @@ -105,563 +104,563 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "cmp %x[M], #0x2\n" "bgt 43f\n" "beq 22f\n" - "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x15, %x[bias]\n" - "mov x14, %x[output_ptr]\n" + "mov x3, %x[bias]\n" + "ldr x4, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x6, %x[output_ptr]\n" "2:" // Height 1: Column loop - "cbz x15, 3f\n" - "ldr q24, [x15, #0x0]\n" - "add x15, x15, #0x10\n" + "cbz x3, 3f\n" + "ldr q24, [x3, #0x0]\n" + "add x3, x3, #0x10\n" "b 8f\n" "3:" // Height 1: no bias "tbz %x[flags], #0, 7f\n" - "cmp x17, #0x4\n" + "cmp x4, #0x4\n" "bge 6f\n" - "tbz x17, #1, 4f\n" - "ldr d24, [x14], #0x8\n" - "mov x8, #0x8\n" - "tbz x17, #0, 5f\n" - "ld1 { v24.s }[2], [x14]\n" + "tbz x4, #1, 4f\n" + "ldr d24, [x6], #0x8\n" + "mov x26, #0x8\n" + "tbz x4, #0, 5f\n" + "ld1 { v24.s }[2], [x6]\n" "b 5f\n" "4:" // Height 1: Partial accumulate: partial_1_0 - "ldr s24, [x14, #0x0]\n" - "mov x8, #0x0\n" + "ldr s24, [x6, #0x0]\n" + "mov x26, #0x0\n" "5:" // Height 1: Partial accumulate: Done - "sub x14, x14, x8\n" + "sub x6, x6, x26\n" "b 8f\n" "6:" // Height 1: full accumulate - "ldr q24, [x14, #0x0]\n" + "ldr q24, [x6, #0x0]\n" "b 8f\n" "7:" // Height 1: no accumulate "movi v24.16b, #0x0\n" "8:" // Height 1: setup done - "mov x13, #0x0\n" + "mov x7, #0x0\n" "9:" // Height 1: String loop - "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w12, [x20, x13, LSL #0x2]\n" + "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr w8, [x26, x7, LSL #0x2]\n" + "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 10f\n" - "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n" - "add x20, x20, x8, LSL #3\n" - "ldr x11, [x20, #0x0]\n" - "cbnz x13, 11f\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x11, x11, x8, LSL #2\n" + "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n" + "add x26, x26, x27, LSL #3\n" + "ldr x17, [x26, #0x0]\n" + "cbnz x7, 11f\n" + "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x17, x17, x26, LSL #2\n" "b 11f\n" "10:" // Height 1: setup direct input - "mov x11, %x[input_ptr]\n" + "mov x17, %x[input_ptr]\n" "11:" // Height 1: input setup done - "cmp x12, #0x4\n" + "cmp x8, #0x4\n" "blt 14f\n" - "ldr q0, [x11, #0x0]\n" - "ldr q8, [x16, #0x0]\n" - "cmp x12, #0x8\n" + "ldr q0, [x17, #0x0]\n" + "cmp x8, #0x8\n" + "ldr q8, [x5, #0x0]\n" + "ldr q9, [x5, #0x10]\n" + "ldr q10, [x5, #0x20]\n" + "ldr q11, [x5, #0x30]\n" "blt 13f\n" "12:" // Height 1: Multiply loop: Main loop head "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr d9, [x16, #0x10]\n" - "ldr x8, [x16, #0x18]\n" - "add x11, x11, #0x10\n" - "ldr d10, [x16, #0x20]\n" - "sub x12, x12, #0x4\n" - "ldr x21, [x16, #0x28]\n" - "cmp x12, #0x8\n" - "mov v9.d[1], x8\n" - "ldr d11, [x16, #0x30]\n" - "ldr x8, [x16, #0x38]\n" - "add x16, x16, #0x40\n" + "add x17, x17, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "mov v10.d[1], x21\n" - "prfm pldl1keep, [x11, #0x80]\n" - "mov v11.d[1], x8\n" - "ldr d8, [x16, #0x0]\n" - "ldr x26, [x16, #0x8]\n" + "add x5, x5, #0x40\n" + "ldr d8, [x5, #0x0]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "ldr x10, [x11, #0x8]\n" - "mov v8.d[1], x26\n" + "ldr d9, [x5, #0x10]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" - "ldr d0, [x11, #0x0]\n" - "mov v0.d[1], x10\n" + "ldr d0, [x17, #0x0]\n" + "sub x8, x8, #0x4\n" + "ldr d10, [x5, #0x20]\n" + "cmp x8, #0x8\n" + "ldr d11, [x5, #0x30]\n" + "ldr x26, [x5, #0x8]\n" + "mov v8.d[1], x26\n" + "ldr x26, [x5, #0x18]\n" + "mov v9.d[1], x26\n" + "ldr x26, [x17, #0x8]\n" + "mov v0.d[1], x26\n" + "ldr x26, [x5, #0x28]\n" + "mov v10.d[1], x26\n" + "ldr x26, [x5, #0x38]\n" + "mov v11.d[1], x26\n" + "prfm pldl1keep, [x17, #0x80]\n" "bge 12b\n" "13:" // Height 1: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr q9, [x16, #0x10]\n" - "ldr q10, [x16, #0x20]\n" - "sub x12, x12, #0x4\n" - "ldr q11, [x16, #0x30]\n" - "add x11, x11, #0x10\n" - "prfm pldl1keep, [x11, #0x80]\n" - "add x16, x16, #0x40\n" + "add x17, x17, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" + "sub x8, x8, #0x4\n" "fmla v24.4s, v10.4s, v0.s[2]\n" + "prfm pldl1keep, [x17, #0x80]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" + "add x5, x5, #0x40\n" "14:" // Height 1: Multiply loop: Main loop skip - "cbz x12, 16f\n" + "cbz x8, 16f\n" "15:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x11], #0x4\n" - "sub x12, x12, #0x1\n" - "ldr q12, [x16, #0x0]\n" - "add x16, x16, #0x10\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "cbnz x12, 15b\n" + "ldr s17, [x17], #0x4\n" + "sub x8, x8, #0x1\n" + "ldr q16, [x5, #0x0]\n" + "fmla v24.4s, v16.4s, v17.s[0]\n" + "add x5, x5, #0x10\n" + "cbnz x8, 15b\n" "16:" // Height 1: Multiply loop: No odd multiplies - "ldr w8, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x13, x13, #0x1\n" - "cmp x13, x8\n" + "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x7, x7, #0x1\n" + "cmp x7, x26\n" "bne 9b\n" - "prfm pstl1keep, [x14, #0x0]\n" + "prfm pstl1keep, [x6, #0x0]\n" "tbz %x[flags], #1, 17f\n" - "add x8, %x[args_ptr], %[offset_min]\n" - "ld1r { v17.4s }, [x8]\n" - "add x8, %x[args_ptr], %[offset_max]\n" - "ld1r { v16.4s }, [x8]\n" + "add x26, %x[args_ptr], %[offset_max]\n" + "ld1r { v16.4s }, [x26]\n" "fmin v24.4s, v24.4s, v16.4s\n" - "fmax v24.4s, v24.4s, v17.4s\n" + "add x26, %x[args_ptr], %[offset_min]\n" + "ld1r { v16.4s }, [x26]\n" + "fmax v24.4s, v24.4s, v16.4s\n" "17:" // Height 1: No activation - "cmp x17, #0x4\n" + "cmp x4, #0x4\n" "bge 20f\n" - "tbz x17, #1, 18f\n" - "str d24, [x14], #0x8\n" - "tbz x17, #0, 19f\n" - "st1 { v24.s }[2], [x14]\n" + "tbz x4, #1, 18f\n" + "str d24, [x6], #0x8\n" + "tbz x4, #0, 19f\n" + "st1 { v24.s }[2], [x6]\n" "b 19f\n" "18:" // Height 1: Partial direct writeback: partial_1_0 - "str s24, [x14, #0x0]\n" + "str s24, [x6, #0x0]\n" "19:" // Height 1: Partial direct writeback: Done "b 21f\n" "20:" // Height 1: Full writeback - "str q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" + "str q24, [x6, #0x0]\n" + "add x6, x6, #0x10\n" "21:" // Height 1: Writeback done - "subs x17, x17, #0x4\n" + "subs x4, x4, #0x4\n" "bgt 2b\n" "b 170f\n" "22:" // Height 2 - "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" - "mov x15, %x[bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[output_ptr]\n" + "mov x3, %x[bias]\n" + "ldr x4, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x6, %x[output_ptr]\n" "23:" // Height 2: Column loop - "cbz x15, 24f\n" - "ldr q24, [x15, #0x0]\n" - "add x15, x15, #0x10\n" + "cbz x3, 24f\n" + "ldr q24, [x3, #0x0]\n" "mov v25.16b, v24.16b\n" + "add x3, x3, #0x10\n" "b 29f\n" "24:" // Height 2: no bias "tbz %x[flags], #0, 28f\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x17, #0x4\n" - "add x27, x14, x8, LSL #2\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "cmp x4, #0x4\n" + "add x13, x6, x26, LSL #2\n" "bge 27f\n" - "tbz x17, #1, 25f\n" - "ldr d24, [x14], #0x8\n" - "ldr d25, [x27], #0x8\n" - "mov x8, #0x8\n" - "tbz x17, #0, 26f\n" - "ld1 { v24.s }[2], [x14]\n" - "ld1 { v25.s }[2], [x27]\n" + "tbz x4, #1, 25f\n" + "ldr d24, [x6], #0x8\n" + "mov x26, #0x8\n" + "ldr d25, [x13], #0x8\n" + "tbz x4, #0, 26f\n" + "ld1 { v24.s }[2], [x6]\n" + "ld1 { v25.s }[2], [x13]\n" "b 26f\n" "25:" // Height 2: Partial accumulate: partial_1_0 - "ldr s24, [x14, #0x0]\n" - "mov x8, #0x0\n" - "ldr s25, [x27, #0x0]\n" + "ldr s24, [x6, #0x0]\n" + "mov x26, #0x0\n" + "ldr s25, [x13, #0x0]\n" "26:" // Height 2: Partial accumulate: Done - "sub x14, x14, x8\n" + "sub x6, x6, x26\n" "b 29f\n" "27:" // Height 2: full accumulate - "ldr q24, [x14, #0x0]\n" - "ldr q25, [x27, #0x0]\n" + "ldr q24, [x6, #0x0]\n" + "ldr q25, [x13, #0x0]\n" "b 29f\n" "28:" // Height 2: no accumulate "movi v24.16b, #0x0\n" "movi v25.16b, #0x0\n" "29:" // Height 2: setup done - "mov x13, #0x0\n" + "mov x7, #0x0\n" "30:" // Height 2: String loop - "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w12, [x20, x13, LSL #0x2]\n" + "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr w8, [x26, x7, LSL #0x2]\n" + "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 31f\n" - "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n" - "add x20, x20, x8, LSL #3\n" - "ldr x11, [x20, #0x0]\n" - "ldr x9, [x20, #0x8]\n" - "cbnz x13, 32f\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x11, x11, x8, LSL #2\n" - "add x9, x9, x8, LSL #2\n" + "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n" + "add x26, x26, x27, LSL #3\n" + "ldr x17, [x26, #0x0]\n" + "ldr x16, [x26, #0x8]\n" + "cbnz x7, 32f\n" + "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x17, x17, x26, LSL #2\n" + "add x16, x16, x26, LSL #2\n" "b 32f\n" "31:" // Height 2: setup direct input - "mov x11, %x[input_ptr]\n" - "add x9, x11, x8, LSL #2\n" + "mov x17, %x[input_ptr]\n" + "add x16, x17, x27, LSL #2\n" "32:" // Height 2: input setup done - "cmp x12, #0x4\n" + "cmp x8, #0x4\n" "blt 35f\n" - "ldr q0, [x11, #0x0]\n" - "ldr q1, [x9, #0x0]\n" - "cmp x12, #0x8\n" - "ldr q8, [x16, #0x0]\n" + "ldr q0, [x17, #0x0]\n" + "cmp x8, #0x8\n" + "ldr q1, [x16, #0x0]\n" + "ldr q8, [x5, #0x0]\n" + "ldr q9, [x5, #0x10]\n" + "ldr q10, [x5, #0x20]\n" + "ldr q11, [x5, #0x30]\n" "blt 34f\n" "33:" // Height 2: Multiply loop: Main loop head "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr d9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr x8, [x16, #0x18]\n" - "ldr d10, [x16, #0x20]\n" - "add x11, x11, #0x10\n" - "ldr x21, [x16, #0x28]\n" - "add x9, x9, #0x10\n" - "mov v9.d[1], x8\n" - "ldr d11, [x16, #0x30]\n" - "ldr x8, [x16, #0x38]\n" - "sub x12, x12, #0x4\n" + "add x16, x16, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "mov v10.d[1], x21\n" + "add x5, x5, #0x40\n" + "ldr d8, [x5, #0x0]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "prfm pldl1keep, [x11, #0x80]\n" - "mov v11.d[1], x8\n" - "prfm pldl1keep, [x9, #0x80]\n" - "ldr x10, [x11, #0x8]\n" - "cmp x12, #0x8\n" + "ldr d9, [x5, #0x10]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "ldr x28, [x9, #0x8]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "add x16, x16, #0x40\n" - "ldr d8, [x16, #0x0]\n" - "ldr x26, [x16, #0x8]\n" + "ldr d10, [x5, #0x20]\n" + "ldr x27, [x5, #0x8]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" - "ldr d0, [x11, #0x0]\n" + "ldr d0, [x17, #0x0]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" - "ldr d1, [x9, #0x0]\n" - "mov v8.d[1], x26\n" - "mov v0.d[1], x10\n" - "mov v1.d[1], x28\n" + "ldr d1, [x16, #0x0]\n" + "sub x8, x8, #0x4\n" + "ldr d11, [x5, #0x30]\n" + "cmp x8, #0x8\n" + "ldr x26, [x5, #0x18]\n" + "mov v8.d[1], x27\n" + "ldr x27, [x17, #0x8]\n" + "mov v9.d[1], x26\n" + "ldr x26, [x16, #0x8]\n" + "mov v0.d[1], x27\n" + "ldr x27, [x5, #0x28]\n" + "mov v1.d[1], x26\n" + "ldr x26, [x5, #0x38]\n" + "mov v10.d[1], x27\n" + "mov v11.d[1], x26\n" + "prfm pldl1keep, [x17, #0x80]\n" + "prfm pldl1keep, [x16, #0x80]\n" "bge 33b\n" "34:" // Height 2: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr q9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr q10, [x16, #0x20]\n" - "ldr q11, [x16, #0x30]\n" - "sub x12, x12, #0x4\n" - "add x11, x11, #0x10\n" - "add x9, x9, #0x10\n" + "add x16, x16, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "prfm pldl1keep, [x11, #0x80]\n" + "sub x8, x8, #0x4\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x16, x16, #0x40\n" + "prfm pldl1keep, [x17, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" + "prfm pldl1keep, [x16, #0x80]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" + "add x5, x5, #0x40\n" "fmla v24.4s, v11.4s, v0.s[3]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" "35:" // Height 2: Multiply loop: Main loop skip - "cbz x12, 37f\n" + "cbz x8, 37f\n" "36:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x11], #0x4\n" - "sub x12, x12, #0x1\n" - "ldr s1, [x9], #0x4\n" - "ldr q12, [x16, #0x0]\n" - "add x16, x16, #0x10\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "cbnz x12, 36b\n" + "ldr s18, [x17], #0x4\n" + "sub x8, x8, #0x1\n" + "ldr s17, [x16], #0x4\n" + "ldr q16, [x5, #0x0]\n" + "fmla v24.4s, v16.4s, v18.s[0]\n" + "fmla v25.4s, v16.4s, v17.s[0]\n" + "add x5, x5, #0x10\n" + "cbnz x8, 36b\n" "37:" // Height 2: Multiply loop: No odd multiplies - "ldr w8, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x13, x13, #0x1\n" - "cmp x13, x8\n" + "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x7, x7, #0x1\n" + "cmp x7, x26\n" "bne 30b\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x14, #0x0]\n" - "add x27, x14, x8, LSL #2\n" - "prfm pstl1keep, [x27, #0x0]\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "prfm pstl1keep, [x6, #0x0]\n" + "prfm pstl1keep, [x13, #0x0]\n" "tbz %x[flags], #1, 38f\n" - "add x20, %x[args_ptr], %[offset_min]\n" - "add x8, %x[args_ptr], %[offset_max]\n" - "ld1r { v17.4s }, [x20]\n" - "ld1r { v16.4s }, [x8]\n" + "add x26, %x[args_ptr], %[offset_max]\n" + "ld1r { v16.4s }, [x26]\n" "fmin v24.4s, v24.4s, v16.4s\n" "fmin v25.4s, v25.4s, v16.4s\n" - "fmax v24.4s, v24.4s, v17.4s\n" - "fmax v25.4s, v25.4s, v17.4s\n" + "add x26, %x[args_ptr], %[offset_min]\n" + "ld1r { v16.4s }, [x26]\n" + "fmax v24.4s, v24.4s, v16.4s\n" + "fmax v25.4s, v25.4s, v16.4s\n" "38:" // Height 2: No activation - "cmp x17, #0x4\n" + "cmp x4, #0x4\n" "bge 41f\n" - "tbz x17, #1, 39f\n" - "str d24, [x14], #0x8\n" - "str d25, [x27], #0x8\n" - "tbz x17, #0, 40f\n" - "st1 { v24.s }[2], [x14]\n" - "st1 { v25.s }[2], [x27]\n" + "tbz x4, #1, 39f\n" + "str d24, [x6], #0x8\n" + "str d25, [x13], #0x8\n" + "tbz x4, #0, 40f\n" + "st1 { v24.s }[2], [x6]\n" + "st1 { v25.s }[2], [x13]\n" "b 40f\n" "39:" // Height 2: Partial direct writeback: partial_1_0 - "str s24, [x14, #0x0]\n" - "str s25, [x27, #0x0]\n" + "str s24, [x6, #0x0]\n" + "str s25, [x13, #0x0]\n" "40:" // Height 2: Partial direct writeback: Done "b 42f\n" "41:" // Height 2: Full writeback - "str q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" - "str q25, [x27, #0x0]\n" + "str q24, [x6, #0x0]\n" + "add x6, x6, #0x10\n" + "str q25, [x13, #0x0]\n" "42:" // Height 2: Writeback done - "subs x17, x17, #0x4\n" + "subs x4, x4, #0x4\n" "bgt 23b\n" "b 170f\n" "43:" // Height 3 - "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" - "mov x15, %x[bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[output_ptr]\n" + "mov x3, %x[bias]\n" + "ldr x4, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x6, %x[output_ptr]\n" "44:" // Height 3: Column loop - "cbz x15, 45f\n" - "ldr q24, [x15, #0x0]\n" - "add x15, x15, #0x10\n" + "cbz x3, 45f\n" + "ldr q24, [x3, #0x0]\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x3, x3, #0x10\n" "b 50f\n" "45:" // Height 3: no bias "tbz %x[flags], #0, 49f\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x17, #0x4\n" - "add x27, x14, x8, LSL #2\n" - "add x26, x27, x8, LSL #2\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "cmp x4, #0x4\n" + "add x12, x13, x26, LSL #2\n" "bge 48f\n" - "tbz x17, #1, 46f\n" - "ldr d24, [x14], #0x8\n" - "ldr d25, [x27], #0x8\n" - "mov x8, #0x8\n" - "ldr d26, [x26], #0x8\n" - "tbz x17, #0, 47f\n" - "ld1 { v24.s }[2], [x14]\n" - "ld1 { v25.s }[2], [x27]\n" - "ld1 { v26.s }[2], [x26]\n" + "tbz x4, #1, 46f\n" + "ldr d24, [x6], #0x8\n" + "mov x26, #0x8\n" + "ldr d25, [x13], #0x8\n" + "ldr d26, [x12], #0x8\n" + "tbz x4, #0, 47f\n" + "ld1 { v24.s }[2], [x6]\n" + "ld1 { v25.s }[2], [x13]\n" + "ld1 { v26.s }[2], [x12]\n" "b 47f\n" "46:" // Height 3: Partial accumulate: partial_1_0 - "ldr s24, [x14, #0x0]\n" - "mov x8, #0x0\n" - "ldr s25, [x27, #0x0]\n" - "ldr s26, [x26, #0x0]\n" + "ldr s24, [x6, #0x0]\n" + "mov x26, #0x0\n" + "ldr s25, [x13, #0x0]\n" + "ldr s26, [x12, #0x0]\n" "47:" // Height 3: Partial accumulate: Done - "sub x14, x14, x8\n" + "sub x6, x6, x26\n" "b 50f\n" "48:" // Height 3: full accumulate - "ldr q24, [x14, #0x0]\n" - "ldr q25, [x27, #0x0]\n" - "ldr q26, [x26, #0x0]\n" + "ldr q24, [x6, #0x0]\n" + "ldr q25, [x13, #0x0]\n" + "ldr q26, [x12, #0x0]\n" "b 50f\n" "49:" // Height 3: no accumulate "movi v24.16b, #0x0\n" "movi v25.16b, #0x0\n" "movi v26.16b, #0x0\n" "50:" // Height 3: setup done - "mov x13, #0x0\n" + "mov x7, #0x0\n" "51:" // Height 3: String loop - "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w12, [x20, x13, LSL #0x2]\n" + "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr w8, [x26, x7, LSL #0x2]\n" + "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 52f\n" - "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n" - "add x20, x20, x8, LSL #3\n" - "ldr x11, [x20, #0x0]\n" - "ldr x9, [x20, #0x8]\n" - "ldr x27, [x20, #0x10]\n" - "cbnz x13, 53f\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x11, x11, x8, LSL #2\n" - "add x9, x9, x8, LSL #2\n" - "add x27, x27, x8, LSL #2\n" + "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n" + "add x26, x26, x27, LSL #3\n" + "ldr x17, [x26, #0x0]\n" + "ldr x16, [x26, #0x8]\n" + "ldr x15, [x26, #0x10]\n" + "cbnz x7, 53f\n" + "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x17, x17, x26, LSL #2\n" + "add x16, x16, x26, LSL #2\n" + "add x15, x15, x26, LSL #2\n" "b 53f\n" "52:" // Height 3: setup direct input - "mov x11, %x[input_ptr]\n" - "add x9, x11, x8, LSL #2\n" - "add x27, x9, x8, LSL #2\n" + "mov x17, %x[input_ptr]\n" + "add x16, x17, x27, LSL #2\n" + "add x15, x16, x27, LSL #2\n" "53:" // Height 3: input setup done - "cmp x12, #0x4\n" + "cmp x8, #0x4\n" "blt 56f\n" - "ldr q0, [x11, #0x0]\n" - "ldr q1, [x9, #0x0]\n" - "cmp x12, #0x8\n" - "ldr q2, [x27, #0x0]\n" - "ldr q8, [x16, #0x0]\n" + "ldr q0, [x17, #0x0]\n" + "cmp x8, #0x8\n" + "ldr q1, [x16, #0x0]\n" + "ldr q2, [x15, #0x0]\n" + "ldr q8, [x5, #0x0]\n" + "ldr q9, [x5, #0x10]\n" + "ldr q10, [x5, #0x20]\n" + "ldr q11, [x5, #0x30]\n" "blt 55f\n" "54:" // Height 3: Multiply loop: Main loop head "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr d9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr x8, [x16, #0x18]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr d10, [x16, #0x20]\n" - "ldr x21, [x16, #0x28]\n" - "add x11, x11, #0x10\n" - "mov v9.d[1], x8\n" - "ldr d11, [x16, #0x30]\n" - "ldr x8, [x16, #0x38]\n" - "add x9, x9, #0x10\n" + "add x15, x15, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "mov v10.d[1], x21\n" + "add x5, x5, #0x40\n" + "ldr d8, [x5, #0x0]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "prfm pldl1keep, [x11, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "mov v11.d[1], x8\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x27, x27, #0x10\n" + "ldr d9, [x5, #0x10]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "ldr x28, [x5, #0x8]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "ldr x10, [x11, #0x8]\n" + "ldr x27, [x5, #0x18]\n" "fmla v26.4s, v10.4s, v2.s[2]\n" - "ldr x28, [x9, #0x8]\n" - "ldr x26, [x27, #0x8]\n" - "sub x12, x12, #0x4\n" + "ldr d10, [x5, #0x20]\n" + "ldr x26, [x5, #0x28]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" - "ldr d0, [x11, #0x0]\n" + "ldr d0, [x17, #0x0]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" - "ldr d1, [x9, #0x0]\n" + "ldr d1, [x16, #0x0]\n" "fmla v26.4s, v11.4s, v2.s[3]\n" - "ldr d2, [x27, #0x0]\n" - "mov v0.d[1], x10\n" - "cmp x12, #0x8\n" + "ldr d2, [x15, #0x0]\n" + "sub x8, x8, #0x4\n" + "ldr d11, [x5, #0x30]\n" + "cmp x8, #0x8\n" + "ldr x9, [x17, #0x8]\n" + "mov v8.d[1], x28\n" + "ldr x28, [x16, #0x8]\n" + "mov v9.d[1], x27\n" + "ldr x27, [x15, #0x8]\n" + "mov v10.d[1], x26\n" + "ldr x26, [x5, #0x38]\n" + "mov v0.d[1], x9\n" "mov v1.d[1], x28\n" - "add x16, x16, #0x40\n" - "mov v2.d[1], x26\n" - "ldr d8, [x16, #0x0]\n" - "ldr x26, [x16, #0x8]\n" - "mov v8.d[1], x26\n" + "prfm pldl1keep, [x17, #0x80]\n" + "mov v2.d[1], x27\n" + "prfm pldl1keep, [x16, #0x80]\n" + "mov v11.d[1], x26\n" + "prfm pldl1keep, [x15, #0x80]\n" "bge 54b\n" "55:" // Height 3: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr q9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr q10, [x16, #0x20]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr q11, [x16, #0x30]\n" - "sub x12, x12, #0x4\n" - "add x11, x11, #0x10\n" + "add x15, x15, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "prfm pldl1keep, [x11, #0x80]\n" + "sub x8, x8, #0x4\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "add x9, x9, #0x10\n" + "prfm pldl1keep, [x17, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "prfm pldl1keep, [x9, #0x80]\n" - "add x27, x27, #0x10\n" - "add x16, x16, #0x40\n" + "prfm pldl1keep, [x16, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x15, #0x80]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" + "add x5, x5, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" "fmla v26.4s, v11.4s, v2.s[3]\n" "56:" // Height 3: Multiply loop: Main loop skip - "cbz x12, 58f\n" + "cbz x8, 58f\n" "57:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x11], #0x4\n" - "sub x12, x12, #0x1\n" - "ldr s1, [x9], #0x4\n" - "ldr s2, [x27], #0x4\n" - "ldr q12, [x16, #0x0]\n" - "add x16, x16, #0x10\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "cbnz x12, 57b\n" + "ldr s19, [x17], #0x4\n" + "sub x8, x8, #0x1\n" + "ldr s18, [x16], #0x4\n" + "ldr s17, [x15], #0x4\n" + "ldr q16, [x5, #0x0]\n" + "fmla v24.4s, v16.4s, v19.s[0]\n" + "fmla v25.4s, v16.4s, v18.s[0]\n" + "add x5, x5, #0x10\n" + "fmla v26.4s, v16.4s, v17.s[0]\n" + "cbnz x8, 57b\n" "58:" // Height 3: Multiply loop: No odd multiplies - "ldr w8, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x13, x13, #0x1\n" - "cmp x13, x8\n" + "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x7, x7, #0x1\n" + "cmp x7, x26\n" "bne 51b\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x14, #0x0]\n" - "add x27, x14, x8, LSL #2\n" - "prfm pstl1keep, [x27, #0x0]\n" - "add x26, x27, x8, LSL #2\n" - "prfm pstl1keep, [x26, #0x0]\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "add x12, x13, x26, LSL #2\n" + "prfm pstl1keep, [x6, #0x0]\n" + "prfm pstl1keep, [x13, #0x0]\n" + "prfm pstl1keep, [x12, #0x0]\n" "tbz %x[flags], #1, 59f\n" - "add x20, %x[args_ptr], %[offset_min]\n" - "add x8, %x[args_ptr], %[offset_max]\n" - "ld1r { v17.4s }, [x20]\n" - "ld1r { v16.4s }, [x8]\n" + "add x26, %x[args_ptr], %[offset_max]\n" + "ld1r { v16.4s }, [x26]\n" "fmin v24.4s, v24.4s, v16.4s\n" "fmin v25.4s, v25.4s, v16.4s\n" "fmin v26.4s, v26.4s, v16.4s\n" - "fmax v24.4s, v24.4s, v17.4s\n" - "fmax v25.4s, v25.4s, v17.4s\n" - "fmax v26.4s, v26.4s, v17.4s\n" + "add x26, %x[args_ptr], %[offset_min]\n" + "ld1r { v16.4s }, [x26]\n" + "fmax v24.4s, v24.4s, v16.4s\n" + "fmax v25.4s, v25.4s, v16.4s\n" + "fmax v26.4s, v26.4s, v16.4s\n" "59:" // Height 3: No activation - "cmp x17, #0x4\n" + "cmp x4, #0x4\n" "bge 62f\n" - "tbz x17, #1, 60f\n" - "str d24, [x14], #0x8\n" - "str d25, [x27], #0x8\n" - "str d26, [x26], #0x8\n" - "tbz x17, #0, 61f\n" - "st1 { v24.s }[2], [x14]\n" - "st1 { v25.s }[2], [x27]\n" - "st1 { v26.s }[2], [x26]\n" + "tbz x4, #1, 60f\n" + "str d24, [x6], #0x8\n" + "str d25, [x13], #0x8\n" + "str d26, [x12], #0x8\n" + "tbz x4, #0, 61f\n" + "st1 { v24.s }[2], [x6]\n" + "st1 { v25.s }[2], [x13]\n" + "st1 { v26.s }[2], [x12]\n" "b 61f\n" "60:" // Height 3: Partial direct writeback: partial_1_0 - "str s24, [x14, #0x0]\n" - "str s25, [x27, #0x0]\n" - "str s26, [x26, #0x0]\n" + "str s24, [x6, #0x0]\n" + "str s25, [x13, #0x0]\n" + "str s26, [x12, #0x0]\n" "61:" // Height 3: Partial direct writeback: Done "b 63f\n" "62:" // Height 3: Full writeback - "str q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" - "str q25, [x27, #0x0]\n" - "str q26, [x26, #0x0]\n" + "str q24, [x6, #0x0]\n" + "add x6, x6, #0x10\n" + "str q25, [x13, #0x0]\n" + "str q26, [x12, #0x0]\n" "63:" // Height 3: Writeback done - "subs x17, x17, #0x4\n" + "subs x4, x4, #0x4\n" "bgt 44b\n" "b 170f\n" "64:" // Height 4 - "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" - "mov x15, %x[bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[output_ptr]\n" + "mov x3, %x[bias]\n" + "ldr x4, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x6, %x[output_ptr]\n" "65:" // Height 4: Column loop - "cbz x15, 66f\n" - "ldr q24, [x15, #0x0]\n" - "add x15, x15, #0x10\n" + "cbz x3, 66f\n" + "ldr q24, [x3, #0x0]\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x3, x3, #0x10\n" "mov v27.16b, v24.16b\n" "b 71f\n" "66:" // Height 4: no bias "tbz %x[flags], #0, 70f\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x17, #0x4\n" - "add x27, x14, x8, LSL #2\n" - "add x26, x27, x8, LSL #2\n" - "add x25, x26, x8, LSL #2\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "add x12, x13, x26, LSL #2\n" + "cmp x4, #0x4\n" + "add x11, x12, x26, LSL #2\n" "bge 69f\n" - "tbz x17, #1, 67f\n" - "ldr d24, [x14], #0x8\n" - "ldr d25, [x27], #0x8\n" - "mov x8, #0x8\n" - "ldr d26, [x26], #0x8\n" - "ldr d27, [x25], #0x8\n" - "tbz x17, #0, 68f\n" - "ld1 { v24.s }[2], [x14]\n" - "ld1 { v25.s }[2], [x27]\n" - "ld1 { v26.s }[2], [x26]\n" - "ld1 { v27.s }[2], [x25]\n" + "tbz x4, #1, 67f\n" + "ldr d24, [x6], #0x8\n" + "mov x26, #0x8\n" + "ldr d25, [x13], #0x8\n" + "ldr d26, [x12], #0x8\n" + "ldr d27, [x11], #0x8\n" + "tbz x4, #0, 68f\n" + "ld1 { v24.s }[2], [x6]\n" + "ld1 { v25.s }[2], [x13]\n" + "ld1 { v26.s }[2], [x12]\n" + "ld1 { v27.s }[2], [x11]\n" "b 68f\n" "67:" // Height 4: Partial accumulate: partial_1_0 - "ldr s24, [x14, #0x0]\n" - "mov x8, #0x0\n" - "ldr s25, [x27, #0x0]\n" - "ldr s26, [x26, #0x0]\n" - "ldr s27, [x25, #0x0]\n" + "ldr s24, [x6, #0x0]\n" + "mov x26, #0x0\n" + "ldr s25, [x13, #0x0]\n" + "ldr s26, [x12, #0x0]\n" + "ldr s27, [x11, #0x0]\n" "68:" // Height 4: Partial accumulate: Done - "sub x14, x14, x8\n" + "sub x6, x6, x26\n" "b 71f\n" "69:" // Height 4: full accumulate - "ldr q24, [x14, #0x0]\n" - "ldr q25, [x27, #0x0]\n" - "ldr q26, [x26, #0x0]\n" - "ldr q27, [x25, #0x0]\n" + "ldr q24, [x6, #0x0]\n" + "ldr q25, [x13, #0x0]\n" + "ldr q26, [x12, #0x0]\n" + "ldr q27, [x11, #0x0]\n" "b 71f\n" "70:" // Height 4: no accumulate "movi v24.16b, #0x0\n" @@ -669,248 +668,248 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "movi v26.16b, #0x0\n" "movi v27.16b, #0x0\n" "71:" // Height 4: setup done - "mov x13, #0x0\n" + "mov x7, #0x0\n" "72:" // Height 4: String loop - "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w12, [x20, x13, LSL #0x2]\n" + "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr w8, [x26, x7, LSL #0x2]\n" + "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 73f\n" - "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n" - "add x20, x20, x8, LSL #3\n" - "ldr x11, [x20, #0x0]\n" - "ldr x9, [x20, #0x8]\n" - "ldr x27, [x20, #0x10]\n" - "ldr x25, [x20, #0x18]\n" - "cbnz x13, 74f\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x11, x11, x8, LSL #2\n" - "add x9, x9, x8, LSL #2\n" - "add x27, x27, x8, LSL #2\n" - "add x25, x25, x8, LSL #2\n" + "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n" + "add x26, x26, x27, LSL #3\n" + "ldr x17, [x26, #0x0]\n" + "ldr x16, [x26, #0x8]\n" + "ldr x15, [x26, #0x10]\n" + "ldr x14, [x26, #0x18]\n" + "cbnz x7, 74f\n" + "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x17, x17, x26, LSL #2\n" + "add x16, x16, x26, LSL #2\n" + "add x15, x15, x26, LSL #2\n" + "add x14, x14, x26, LSL #2\n" "b 74f\n" "73:" // Height 4: setup direct input - "mov x11, %x[input_ptr]\n" - "add x9, x11, x8, LSL #2\n" - "add x27, x9, x8, LSL #2\n" - "add x25, x27, x8, LSL #2\n" + "mov x17, %x[input_ptr]\n" + "add x16, x17, x27, LSL #2\n" + "add x15, x16, x27, LSL #2\n" + "add x14, x15, x27, LSL #2\n" "74:" // Height 4: input setup done - "cmp x12, #0x4\n" + "cmp x8, #0x4\n" "blt 77f\n" - "ldr q0, [x11, #0x0]\n" - "ldr q1, [x9, #0x0]\n" - "cmp x12, #0x8\n" - "ldr q2, [x27, #0x0]\n" - "ldr q3, [x25, #0x0]\n" - "ldr q8, [x16, #0x0]\n" + "ldr q0, [x17, #0x0]\n" + "cmp x8, #0x8\n" + "ldr q1, [x16, #0x0]\n" + "ldr q2, [x15, #0x0]\n" + "ldr q3, [x14, #0x0]\n" + "ldr q8, [x5, #0x0]\n" + "ldr q9, [x5, #0x10]\n" + "ldr q10, [x5, #0x20]\n" + "ldr q11, [x5, #0x30]\n" "blt 76f\n" "75:" // Height 4: Multiply loop: Main loop head "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr d9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr x8, [x16, #0x18]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr d10, [x16, #0x20]\n" + "add x15, x15, #0x10\n" "fmla v27.4s, v8.4s, v3.s[0]\n" - "ldr x21, [x16, #0x28]\n" - "mov v9.d[1], x8\n" - "ldr d11, [x16, #0x30]\n" - "ldr x8, [x16, #0x38]\n" - "add x11, x11, #0x10\n" + "add x14, x14, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "mov v10.d[1], x21\n" + "add x5, x5, #0x40\n" + "ldr d8, [x5, #0x0]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "prfm pldl1keep, [x11, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "mov v11.d[1], x8\n" + "ldr x27, [x5, #0x8]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "ldr x10, [x11, #0x8]\n" + "ldr d9, [x5, #0x10]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "add x9, x9, #0x10\n" + "ldr x26, [x5, #0x18]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "ldr x11, [x5, #0x28]\n" "fmla v26.4s, v10.4s, v2.s[2]\n" - "ldr x28, [x9, #0x8]\n" + "ldr x10, [x17, #0x8]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" - "add x27, x27, #0x10\n" + "ldr d10, [x5, #0x20]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "ldr d0, [x17, #0x0]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" - "ldr d0, [x11, #0x0]\n" + "ldr d1, [x16, #0x0]\n" + "ldr x9, [x16, #0x8]\n" "fmla v26.4s, v11.4s, v2.s[3]\n" - "ldr d1, [x9, #0x0]\n" + "ldr d2, [x15, #0x0]\n" "fmla v27.4s, v11.4s, v3.s[3]\n" - "ldr d2, [x27, #0x0]\n" + "ldr d3, [x14, #0x0]\n" + "sub x8, x8, #0x4\n" + "ldr d11, [x5, #0x30]\n" + "cmp x8, #0x8\n" + "ldr x28, [x15, #0x8]\n" + "mov v8.d[1], x27\n" + "ldr x27, [x14, #0x8]\n" + "mov v9.d[1], x26\n" + "ldr x26, [x5, #0x38]\n" + "mov v10.d[1], x11\n" + "prfm pldl1keep, [x17, #0x80]\n" "mov v0.d[1], x10\n" - "ldr x26, [x27, #0x8]\n" - "mov v1.d[1], x28\n" - "add x25, x25, #0x10\n" - "prfm pldl1keep, [x25, #0x80]\n" - "sub x12, x12, #0x4\n" - "mov v2.d[1], x26\n" - "ldr d3, [x25, #0x0]\n" - "ldr x8, [x25, #0x8]\n" - "cmp x12, #0x8\n" - "add x16, x16, #0x40\n" - "ldr d8, [x16, #0x0]\n" - "mov v3.d[1], x8\n" - "ldr x26, [x16, #0x8]\n" - "mov v8.d[1], x26\n" + "prfm pldl1keep, [x16, #0x80]\n" + "mov v1.d[1], x9\n" + "mov v2.d[1], x28\n" + "prfm pldl1keep, [x15, #0x80]\n" + "mov v3.d[1], x27\n" + "prfm pldl1keep, [x14, #0x80]\n" + "mov v11.d[1], x26\n" "bge 75b\n" "76:" // Height 4: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr q9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr q10, [x16, #0x20]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr q11, [x16, #0x30]\n" + "add x15, x15, #0x10\n" "fmla v27.4s, v8.4s, v3.s[0]\n" - "sub x12, x12, #0x4\n" + "add x14, x14, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "add x11, x11, #0x10\n" + "sub x8, x8, #0x4\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "prfm pldl1keep, [x11, #0x80]\n" + "prfm pldl1keep, [x17, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "add x9, x9, #0x10\n" + "prfm pldl1keep, [x16, #0x80]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "prfm pldl1keep, [x15, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "add x27, x27, #0x10\n" + "prfm pldl1keep, [x14, #0x80]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "add x5, x5, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" - "add x25, x25, #0x10\n" "fmla v27.4s, v10.4s, v3.s[2]\n" - "prfm pldl1keep, [x25, #0x80]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" - "add x16, x16, #0x40\n" "fmla v25.4s, v11.4s, v1.s[3]\n" "fmla v26.4s, v11.4s, v2.s[3]\n" "fmla v27.4s, v11.4s, v3.s[3]\n" "77:" // Height 4: Multiply loop: Main loop skip - "cbz x12, 79f\n" + "cbz x8, 79f\n" "78:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x11], #0x4\n" - "sub x12, x12, #0x1\n" - "ldr s1, [x9], #0x4\n" - "ldr s2, [x27], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr q12, [x16, #0x0]\n" - "add x16, x16, #0x10\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" - "cbnz x12, 78b\n" + "ldr s20, [x17], #0x4\n" + "sub x8, x8, #0x1\n" + "ldr s19, [x16], #0x4\n" + "ldr s18, [x15], #0x4\n" + "ldr s17, [x14], #0x4\n" + "ldr q16, [x5, #0x0]\n" + "fmla v24.4s, v16.4s, v20.s[0]\n" + "fmla v25.4s, v16.4s, v19.s[0]\n" + "add x5, x5, #0x10\n" + "fmla v26.4s, v16.4s, v18.s[0]\n" + "fmla v27.4s, v16.4s, v17.s[0]\n" + "cbnz x8, 78b\n" "79:" // Height 4: Multiply loop: No odd multiplies - "ldr w8, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x13, x13, #0x1\n" - "cmp x13, x8\n" + "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x7, x7, #0x1\n" + "cmp x7, x26\n" "bne 72b\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x14, #0x0]\n" - "add x27, x14, x8, LSL #2\n" - "prfm pstl1keep, [x27, #0x0]\n" - "add x26, x27, x8, LSL #2\n" - "prfm pstl1keep, [x26, #0x0]\n" - "add x25, x26, x8, LSL #2\n" - "prfm pstl1keep, [x25, #0x0]\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "add x12, x13, x26, LSL #2\n" + "add x11, x12, x26, LSL #2\n" + "prfm pstl1keep, [x6, #0x0]\n" + "prfm pstl1keep, [x13, #0x0]\n" + "prfm pstl1keep, [x12, #0x0]\n" + "prfm pstl1keep, [x11, #0x0]\n" "tbz %x[flags], #1, 80f\n" - "add x20, %x[args_ptr], %[offset_min]\n" - "add x8, %x[args_ptr], %[offset_max]\n" - "ld1r { v17.4s }, [x20]\n" - "ld1r { v16.4s }, [x8]\n" + "add x26, %x[args_ptr], %[offset_max]\n" + "ld1r { v16.4s }, [x26]\n" "fmin v24.4s, v24.4s, v16.4s\n" "fmin v25.4s, v25.4s, v16.4s\n" "fmin v26.4s, v26.4s, v16.4s\n" "fmin v27.4s, v27.4s, v16.4s\n" - "fmax v24.4s, v24.4s, v17.4s\n" - "fmax v25.4s, v25.4s, v17.4s\n" - "fmax v26.4s, v26.4s, v17.4s\n" - "fmax v27.4s, v27.4s, v17.4s\n" + "add x26, %x[args_ptr], %[offset_min]\n" + "ld1r { v16.4s }, [x26]\n" + "fmax v24.4s, v24.4s, v16.4s\n" + "fmax v25.4s, v25.4s, v16.4s\n" + "fmax v26.4s, v26.4s, v16.4s\n" + "fmax v27.4s, v27.4s, v16.4s\n" "80:" // Height 4: No activation - "cmp x17, #0x4\n" + "cmp x4, #0x4\n" "bge 83f\n" - "tbz x17, #1, 81f\n" - "str d24, [x14], #0x8\n" - "str d25, [x27], #0x8\n" - "str d26, [x26], #0x8\n" - "str d27, [x25], #0x8\n" - "tbz x17, #0, 82f\n" - "st1 { v24.s }[2], [x14]\n" - "st1 { v25.s }[2], [x27]\n" - "st1 { v26.s }[2], [x26]\n" - "st1 { v27.s }[2], [x25]\n" + "tbz x4, #1, 81f\n" + "str d24, [x6], #0x8\n" + "str d25, [x13], #0x8\n" + "str d26, [x12], #0x8\n" + "str d27, [x11], #0x8\n" + "tbz x4, #0, 82f\n" + "st1 { v24.s }[2], [x6]\n" + "st1 { v25.s }[2], [x13]\n" + "st1 { v26.s }[2], [x12]\n" + "st1 { v27.s }[2], [x11]\n" "b 82f\n" "81:" // Height 4: Partial direct writeback: partial_1_0 - "str s24, [x14, #0x0]\n" - "str s25, [x27, #0x0]\n" - "str s26, [x26, #0x0]\n" - "str s27, [x25, #0x0]\n" + "str s24, [x6, #0x0]\n" + "str s25, [x13, #0x0]\n" + "str s26, [x12, #0x0]\n" + "str s27, [x11, #0x0]\n" "82:" // Height 4: Partial direct writeback: Done "b 84f\n" "83:" // Height 4: Full writeback - "str q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" - "str q25, [x27, #0x0]\n" - "str q26, [x26, #0x0]\n" - "str q27, [x25, #0x0]\n" + "str q24, [x6, #0x0]\n" + "add x6, x6, #0x10\n" + "str q25, [x13, #0x0]\n" + "str q26, [x12, #0x0]\n" + "str q27, [x11, #0x0]\n" "84:" // Height 4: Writeback done - "subs x17, x17, #0x4\n" + "subs x4, x4, #0x4\n" "bgt 65b\n" "b 170f\n" "85:" // Height 5 - "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" - "mov x15, %x[bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[output_ptr]\n" + "mov x3, %x[bias]\n" + "ldr x4, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x6, %x[output_ptr]\n" "86:" // Height 5: Column loop - "cbz x15, 87f\n" - "ldr q24, [x15, #0x0]\n" - "add x15, x15, #0x10\n" + "cbz x3, 87f\n" + "ldr q24, [x3, #0x0]\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x3, x3, #0x10\n" "mov v27.16b, v24.16b\n" "mov v28.16b, v24.16b\n" "b 92f\n" "87:" // Height 5: no bias "tbz %x[flags], #0, 91f\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x17, #0x4\n" - "add x27, x14, x8, LSL #2\n" - "add x26, x27, x8, LSL #2\n" - "add x25, x26, x8, LSL #2\n" - "add x24, x25, x8, LSL #2\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "add x12, x13, x26, LSL #2\n" + "add x11, x12, x26, LSL #2\n" + "cmp x4, #0x4\n" + "add x10, x11, x26, LSL #2\n" "bge 90f\n" - "tbz x17, #1, 88f\n" - "ldr d24, [x14], #0x8\n" - "ldr d25, [x27], #0x8\n" - "mov x8, #0x8\n" - "ldr d26, [x26], #0x8\n" - "ldr d27, [x25], #0x8\n" - "ldr d28, [x24], #0x8\n" - "tbz x17, #0, 89f\n" - "ld1 { v24.s }[2], [x14]\n" - "ld1 { v25.s }[2], [x27]\n" - "ld1 { v26.s }[2], [x26]\n" - "ld1 { v27.s }[2], [x25]\n" - "ld1 { v28.s }[2], [x24]\n" + "tbz x4, #1, 88f\n" + "ldr d24, [x6], #0x8\n" + "mov x26, #0x8\n" + "ldr d25, [x13], #0x8\n" + "ldr d26, [x12], #0x8\n" + "ldr d27, [x11], #0x8\n" + "ldr d28, [x10], #0x8\n" + "tbz x4, #0, 89f\n" + "ld1 { v24.s }[2], [x6]\n" + "ld1 { v25.s }[2], [x13]\n" + "ld1 { v26.s }[2], [x12]\n" + "ld1 { v27.s }[2], [x11]\n" + "ld1 { v28.s }[2], [x10]\n" "b 89f\n" "88:" // Height 5: Partial accumulate: partial_1_0 - "ldr s24, [x14, #0x0]\n" - "mov x8, #0x0\n" - "ldr s25, [x27, #0x0]\n" - "ldr s26, [x26, #0x0]\n" - "ldr s27, [x25, #0x0]\n" - "ldr s28, [x24, #0x0]\n" + "ldr s24, [x6, #0x0]\n" + "mov x26, #0x0\n" + "ldr s25, [x13, #0x0]\n" + "ldr s26, [x12, #0x0]\n" + "ldr s27, [x11, #0x0]\n" + "ldr s28, [x10, #0x0]\n" "89:" // Height 5: Partial accumulate: Done - "sub x14, x14, x8\n" + "sub x6, x6, x26\n" "b 92f\n" "90:" // Height 5: full accumulate - "ldr q24, [x14, #0x0]\n" - "ldr q25, [x27, #0x0]\n" - "ldr q26, [x26, #0x0]\n" - "ldr q27, [x25, #0x0]\n" - "ldr q28, [x24, #0x0]\n" + "ldr q24, [x6, #0x0]\n" + "ldr q25, [x13, #0x0]\n" + "ldr q26, [x12, #0x0]\n" + "ldr q27, [x11, #0x0]\n" + "ldr q28, [x10, #0x0]\n" "b 92f\n" "91:" // Height 5: no accumulate "movi v24.16b, #0x0\n" @@ -919,283 +918,283 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "movi v27.16b, #0x0\n" "movi v28.16b, #0x0\n" "92:" // Height 5: setup done - "mov x13, #0x0\n" + "mov x7, #0x0\n" "93:" // Height 5: String loop - "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w12, [x20, x13, LSL #0x2]\n" + "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr w8, [x26, x7, LSL #0x2]\n" + "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 94f\n" - "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n" - "add x20, x20, x8, LSL #3\n" - "ldr x11, [x20, #0x0]\n" - "ldr x9, [x20, #0x8]\n" - "ldr x27, [x20, #0x10]\n" - "ldr x25, [x20, #0x18]\n" - "ldr x24, [x20, #0x20]\n" - "cbnz x13, 95f\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x11, x11, x8, LSL #2\n" - "add x9, x9, x8, LSL #2\n" - "add x27, x27, x8, LSL #2\n" - "add x25, x25, x8, LSL #2\n" - "add x24, x24, x8, LSL #2\n" + "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n" + "add x26, x26, x27, LSL #3\n" + "ldr x17, [x26, #0x0]\n" + "ldr x16, [x26, #0x8]\n" + "ldr x15, [x26, #0x10]\n" + "ldr x14, [x26, #0x18]\n" + "ldr x13, [x26, #0x20]\n" + "cbnz x7, 95f\n" + "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x17, x17, x26, LSL #2\n" + "add x16, x16, x26, LSL #2\n" + "add x15, x15, x26, LSL #2\n" + "add x14, x14, x26, LSL #2\n" + "add x13, x13, x26, LSL #2\n" "b 95f\n" "94:" // Height 5: setup direct input - "mov x11, %x[input_ptr]\n" - "add x9, x11, x8, LSL #2\n" - "add x27, x9, x8, LSL #2\n" - "add x25, x27, x8, LSL #2\n" - "add x24, x25, x8, LSL #2\n" + "mov x17, %x[input_ptr]\n" + "add x16, x17, x27, LSL #2\n" + "add x15, x16, x27, LSL #2\n" + "add x14, x15, x27, LSL #2\n" + "add x13, x14, x27, LSL #2\n" "95:" // Height 5: input setup done - "cmp x12, #0x4\n" + "cmp x8, #0x4\n" "blt 98f\n" - "ldr q0, [x11, #0x0]\n" - "ldr q1, [x9, #0x0]\n" - "cmp x12, #0x8\n" - "ldr q2, [x27, #0x0]\n" - "ldr q3, [x25, #0x0]\n" - "ldr q4, [x24, #0x0]\n" - "ldr q8, [x16, #0x0]\n" + "ldr q0, [x17, #0x0]\n" + "cmp x8, #0x8\n" + "ldr q1, [x16, #0x0]\n" + "ldr q2, [x15, #0x0]\n" + "ldr q3, [x14, #0x0]\n" + "ldr q4, [x13, #0x0]\n" + "ldr q8, [x5, #0x0]\n" + "ldr q9, [x5, #0x10]\n" + "ldr q10, [x5, #0x20]\n" + "ldr q11, [x5, #0x30]\n" "blt 97f\n" "96:" // Height 5: Multiply loop: Main loop head "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr d9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr x8, [x16, #0x18]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr d10, [x16, #0x20]\n" + "add x15, x15, #0x10\n" "fmla v27.4s, v8.4s, v3.s[0]\n" - "ldr x21, [x16, #0x28]\n" + "add x14, x14, #0x10\n" "fmla v28.4s, v8.4s, v4.s[0]\n" - "mov v9.d[1], x8\n" - "ldr d11, [x16, #0x30]\n" - "add x11, x11, #0x10\n" + "add x13, x13, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "mov v10.d[1], x21\n" + "add x5, x5, #0x40\n" + "ldr d8, [x5, #0x0]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "ldr x8, [x16, #0x38]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "prfm pldl1keep, [x11, #0x80]\n" + "ldr x27, [x5, #0x8]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "ldr x10, [x11, #0x8]\n" + "ldr x26, [x5, #0x18]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" - "mov v11.d[1], x8\n" + "ldr d9, [x5, #0x10]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "add x9, x9, #0x10\n" + "ldr x12, [x5, #0x28]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "ldr x11, [x17, #0x8]\n" "fmla v26.4s, v10.4s, v2.s[2]\n" - "ldr x28, [x9, #0x8]\n" + "ldr x10, [x16, #0x8]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" - "add x27, x27, #0x10\n" + "ldr x9, [x15, #0x8]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "ldr d10, [x5, #0x20]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" - "ldr d0, [x11, #0x0]\n" + "ldr d0, [x17, #0x0]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" - "ldr d1, [x9, #0x0]\n" + "ldr d1, [x16, #0x0]\n" "fmla v26.4s, v11.4s, v2.s[3]\n" - "ldr d2, [x27, #0x0]\n" + "ldr d2, [x15, #0x0]\n" "fmla v27.4s, v11.4s, v3.s[3]\n" - "mov v0.d[1], x10\n" + "ldr d3, [x14, #0x0]\n" + "ldr x28, [x14, #0x8]\n" "fmla v28.4s, v11.4s, v4.s[3]\n" - "mov v1.d[1], x28\n" - "ldr x26, [x27, #0x8]\n" - "add x25, x25, #0x10\n" - "prfm pldl1keep, [x25, #0x80]\n" - "add x24, x24, #0x10\n" - "prfm pldl1keep, [x24, #0x80]\n" - "sub x12, x12, #0x4\n" - "mov v2.d[1], x26\n" - "ldr d3, [x25, #0x0]\n" - "ldr x8, [x25, #0x8]\n" - "cmp x12, #0x8\n" - "ldr d4, [x24, #0x0]\n" - "add x16, x16, #0x40\n" - "ldr x21, [x24, #0x8]\n" - "mov v3.d[1], x8\n" - "ldr d8, [x16, #0x0]\n" - "ldr x26, [x16, #0x8]\n" - "mov v4.d[1], x21\n" - "mov v8.d[1], x26\n" + "ldr d4, [x13, #0x0]\n" + "sub x8, x8, #0x4\n" + "ldr d11, [x5, #0x30]\n" + "cmp x8, #0x8\n" + "mov v8.d[1], x27\n" + "ldr x27, [x13, #0x8]\n" + "mov v9.d[1], x26\n" + "ldr x26, [x5, #0x38]\n" + "prfm pldl1keep, [x17, #0x80]\n" + "mov v10.d[1], x12\n" + "prfm pldl1keep, [x16, #0x80]\n" + "mov v0.d[1], x11\n" + "prfm pldl1keep, [x15, #0x80]\n" + "mov v1.d[1], x10\n" + "prfm pldl1keep, [x14, #0x80]\n" + "mov v2.d[1], x9\n" + "mov v3.d[1], x28\n" + "prfm pldl1keep, [x13, #0x80]\n" + "mov v4.d[1], x27\n" + "mov v11.d[1], x26\n" "bge 96b\n" "97:" // Height 5: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr q9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr q10, [x16, #0x20]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr q11, [x16, #0x30]\n" + "add x15, x15, #0x10\n" "fmla v27.4s, v8.4s, v3.s[0]\n" - "sub x12, x12, #0x4\n" + "add x14, x14, #0x10\n" "fmla v28.4s, v8.4s, v4.s[0]\n" - "add x11, x11, #0x10\n" + "add x13, x13, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "prfm pldl1keep, [x11, #0x80]\n" + "sub x8, x8, #0x4\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "add x9, x9, #0x10\n" + "prfm pldl1keep, [x17, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "prfm pldl1keep, [x16, #0x80]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "add x27, x27, #0x10\n" + "prfm pldl1keep, [x15, #0x80]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x14, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "add x25, x25, #0x10\n" + "prfm pldl1keep, [x13, #0x80]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "add x5, x5, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" - "add x24, x24, #0x10\n" "fmla v27.4s, v10.4s, v3.s[2]\n" - "prfm pldl1keep, [x24, #0x80]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" - "add x16, x16, #0x40\n" "fmla v24.4s, v11.4s, v0.s[3]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" "fmla v26.4s, v11.4s, v2.s[3]\n" "fmla v27.4s, v11.4s, v3.s[3]\n" "fmla v28.4s, v11.4s, v4.s[3]\n" "98:" // Height 5: Multiply loop: Main loop skip - "cbz x12, 100f\n" + "cbz x8, 100f\n" "99:" // Height 5: Multiply loop: Odd block loop - "ldr s0, [x11], #0x4\n" - "sub x12, x12, #0x1\n" - "ldr s1, [x9], #0x4\n" - "ldr s2, [x27], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr q12, [x16, #0x0]\n" - "add x16, x16, #0x10\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" - "fmla v28.4s, v12.4s, v4.s[0]\n" - "cbnz x12, 99b\n" + "ldr s21, [x17], #0x4\n" + "sub x8, x8, #0x1\n" + "ldr s20, [x16], #0x4\n" + "ldr s19, [x15], #0x4\n" + "ldr s18, [x14], #0x4\n" + "ldr s17, [x13], #0x4\n" + "ldr q16, [x5, #0x0]\n" + "fmla v24.4s, v16.4s, v21.s[0]\n" + "fmla v25.4s, v16.4s, v20.s[0]\n" + "add x5, x5, #0x10\n" + "fmla v26.4s, v16.4s, v19.s[0]\n" + "fmla v27.4s, v16.4s, v18.s[0]\n" + "fmla v28.4s, v16.4s, v17.s[0]\n" + "cbnz x8, 99b\n" "100:" // Height 5: Multiply loop: No odd multiplies - "ldr w8, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x13, x13, #0x1\n" - "cmp x13, x8\n" + "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x7, x7, #0x1\n" + "cmp x7, x26\n" "bne 93b\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x14, #0x0]\n" - "add x27, x14, x8, LSL #2\n" - "prfm pstl1keep, [x27, #0x0]\n" - "add x26, x27, x8, LSL #2\n" - "prfm pstl1keep, [x26, #0x0]\n" - "add x25, x26, x8, LSL #2\n" - "prfm pstl1keep, [x25, #0x0]\n" - "add x24, x25, x8, LSL #2\n" - "prfm pstl1keep, [x24, #0x0]\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "add x12, x13, x26, LSL #2\n" + "add x11, x12, x26, LSL #2\n" + "add x10, x11, x26, LSL #2\n" + "prfm pstl1keep, [x6, #0x0]\n" + "prfm pstl1keep, [x13, #0x0]\n" + "prfm pstl1keep, [x12, #0x0]\n" + "prfm pstl1keep, [x11, #0x0]\n" + "prfm pstl1keep, [x10, #0x0]\n" "tbz %x[flags], #1, 101f\n" - "add x20, %x[args_ptr], %[offset_min]\n" - "add x8, %x[args_ptr], %[offset_max]\n" - "ld1r { v17.4s }, [x20]\n" - "ld1r { v16.4s }, [x8]\n" + "add x26, %x[args_ptr], %[offset_max]\n" + "ld1r { v16.4s }, [x26]\n" "fmin v24.4s, v24.4s, v16.4s\n" "fmin v25.4s, v25.4s, v16.4s\n" "fmin v26.4s, v26.4s, v16.4s\n" "fmin v27.4s, v27.4s, v16.4s\n" "fmin v28.4s, v28.4s, v16.4s\n" - "fmax v24.4s, v24.4s, v17.4s\n" - "fmax v25.4s, v25.4s, v17.4s\n" - "fmax v26.4s, v26.4s, v17.4s\n" - "fmax v27.4s, v27.4s, v17.4s\n" - "fmax v28.4s, v28.4s, v17.4s\n" + "add x26, %x[args_ptr], %[offset_min]\n" + "ld1r { v16.4s }, [x26]\n" + "fmax v24.4s, v24.4s, v16.4s\n" + "fmax v25.4s, v25.4s, v16.4s\n" + "fmax v26.4s, v26.4s, v16.4s\n" + "fmax v27.4s, v27.4s, v16.4s\n" + "fmax v28.4s, v28.4s, v16.4s\n" "101:" // Height 5: No activation - "cmp x17, #0x4\n" + "cmp x4, #0x4\n" "bge 104f\n" - "tbz x17, #1, 102f\n" - "str d24, [x14], #0x8\n" - "str d25, [x27], #0x8\n" - "str d26, [x26], #0x8\n" - "str d27, [x25], #0x8\n" - "str d28, [x24], #0x8\n" - "tbz x17, #0, 103f\n" - "st1 { v24.s }[2], [x14]\n" - "st1 { v25.s }[2], [x27]\n" - "st1 { v26.s }[2], [x26]\n" - "st1 { v27.s }[2], [x25]\n" - "st1 { v28.s }[2], [x24]\n" + "tbz x4, #1, 102f\n" + "str d24, [x6], #0x8\n" + "str d25, [x13], #0x8\n" + "str d26, [x12], #0x8\n" + "str d27, [x11], #0x8\n" + "str d28, [x10], #0x8\n" + "tbz x4, #0, 103f\n" + "st1 { v24.s }[2], [x6]\n" + "st1 { v25.s }[2], [x13]\n" + "st1 { v26.s }[2], [x12]\n" + "st1 { v27.s }[2], [x11]\n" + "st1 { v28.s }[2], [x10]\n" "b 103f\n" "102:" // Height 5: Partial direct writeback: partial_1_0 - "str s24, [x14, #0x0]\n" - "str s25, [x27, #0x0]\n" - "str s26, [x26, #0x0]\n" - "str s27, [x25, #0x0]\n" - "str s28, [x24, #0x0]\n" + "str s24, [x6, #0x0]\n" + "str s25, [x13, #0x0]\n" + "str s26, [x12, #0x0]\n" + "str s27, [x11, #0x0]\n" + "str s28, [x10, #0x0]\n" "103:" // Height 5: Partial direct writeback: Done "b 105f\n" "104:" // Height 5: Full writeback - "str q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" - "str q25, [x27, #0x0]\n" - "str q26, [x26, #0x0]\n" - "str q27, [x25, #0x0]\n" - "str q28, [x24, #0x0]\n" + "str q24, [x6, #0x0]\n" + "add x6, x6, #0x10\n" + "str q25, [x13, #0x0]\n" + "str q26, [x12, #0x0]\n" + "str q27, [x11, #0x0]\n" + "str q28, [x10, #0x0]\n" "105:" // Height 5: Writeback done - "subs x17, x17, #0x4\n" + "subs x4, x4, #0x4\n" "bgt 86b\n" "b 170f\n" "106:" // Height 6 - "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" - "mov x15, %x[bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[output_ptr]\n" + "mov x3, %x[bias]\n" + "ldr x4, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x6, %x[output_ptr]\n" "107:" // Height 6: Column loop - "cbz x15, 108f\n" - "ldr q24, [x15, #0x0]\n" - "add x15, x15, #0x10\n" + "cbz x3, 108f\n" + "ldr q24, [x3, #0x0]\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x3, x3, #0x10\n" "mov v27.16b, v24.16b\n" "mov v28.16b, v24.16b\n" "mov v29.16b, v24.16b\n" "b 113f\n" "108:" // Height 6: no bias "tbz %x[flags], #0, 112f\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x17, #0x4\n" - "add x27, x14, x8, LSL #2\n" - "add x26, x27, x8, LSL #2\n" - "add x25, x26, x8, LSL #2\n" - "add x24, x25, x8, LSL #2\n" - "add x23, x24, x8, LSL #2\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "add x12, x13, x26, LSL #2\n" + "add x11, x12, x26, LSL #2\n" + "add x10, x11, x26, LSL #2\n" + "cmp x4, #0x4\n" + "add x9, x10, x26, LSL #2\n" "bge 111f\n" - "tbz x17, #1, 109f\n" - "ldr d24, [x14], #0x8\n" - "ldr d25, [x27], #0x8\n" - "mov x8, #0x8\n" - "ldr d26, [x26], #0x8\n" - "ldr d27, [x25], #0x8\n" - "ldr d28, [x24], #0x8\n" - "ldr d29, [x23], #0x8\n" - "tbz x17, #0, 110f\n" - "ld1 { v24.s }[2], [x14]\n" - "ld1 { v25.s }[2], [x27]\n" - "ld1 { v26.s }[2], [x26]\n" - "ld1 { v27.s }[2], [x25]\n" - "ld1 { v28.s }[2], [x24]\n" - "ld1 { v29.s }[2], [x23]\n" + "tbz x4, #1, 109f\n" + "ldr d24, [x6], #0x8\n" + "mov x26, #0x8\n" + "ldr d25, [x13], #0x8\n" + "ldr d26, [x12], #0x8\n" + "ldr d27, [x11], #0x8\n" + "ldr d28, [x10], #0x8\n" + "ldr d29, [x9], #0x8\n" + "tbz x4, #0, 110f\n" + "ld1 { v24.s }[2], [x6]\n" + "ld1 { v25.s }[2], [x13]\n" + "ld1 { v26.s }[2], [x12]\n" + "ld1 { v27.s }[2], [x11]\n" + "ld1 { v28.s }[2], [x10]\n" + "ld1 { v29.s }[2], [x9]\n" "b 110f\n" "109:" // Height 6: Partial accumulate: partial_1_0 - "ldr s24, [x14, #0x0]\n" - "mov x8, #0x0\n" - "ldr s25, [x27, #0x0]\n" - "ldr s26, [x26, #0x0]\n" - "ldr s27, [x25, #0x0]\n" - "ldr s28, [x24, #0x0]\n" - "ldr s29, [x23, #0x0]\n" + "ldr s24, [x6, #0x0]\n" + "mov x26, #0x0\n" + "ldr s25, [x13, #0x0]\n" + "ldr s26, [x12, #0x0]\n" + "ldr s27, [x11, #0x0]\n" + "ldr s28, [x10, #0x0]\n" + "ldr s29, [x9, #0x0]\n" "110:" // Height 6: Partial accumulate: Done - "sub x14, x14, x8\n" + "sub x6, x6, x26\n" "b 113f\n" "111:" // Height 6: full accumulate - "ldr q24, [x14, #0x0]\n" - "ldr q25, [x27, #0x0]\n" - "ldr q26, [x26, #0x0]\n" - "ldr q27, [x25, #0x0]\n" - "ldr q28, [x24, #0x0]\n" - "ldr q29, [x23, #0x0]\n" + "ldr q24, [x6, #0x0]\n" + "ldr q25, [x13, #0x0]\n" + "ldr q26, [x12, #0x0]\n" + "ldr q27, [x11, #0x0]\n" + "ldr q28, [x10, #0x0]\n" + "ldr q29, [x9, #0x0]\n" "b 113f\n" "112:" // Height 6: no accumulate "movi v24.16b, #0x0\n" @@ -1205,154 +1204,154 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "movi v28.16b, #0x0\n" "movi v29.16b, #0x0\n" "113:" // Height 6: setup done - "mov x13, #0x0\n" + "mov x7, #0x0\n" "114:" // Height 6: String loop - "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w12, [x20, x13, LSL #0x2]\n" + "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr w8, [x26, x7, LSL #0x2]\n" + "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 115f\n" - "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n" - "add x20, x20, x8, LSL #3\n" - "ldr x11, [x20, #0x0]\n" - "ldr x9, [x20, #0x8]\n" - "ldr x27, [x20, #0x10]\n" - "ldr x25, [x20, #0x18]\n" - "ldr x24, [x20, #0x20]\n" - "ldr x23, [x20, #0x28]\n" - "cbnz x13, 116f\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x11, x11, x8, LSL #2\n" - "add x9, x9, x8, LSL #2\n" - "add x27, x27, x8, LSL #2\n" - "add x25, x25, x8, LSL #2\n" - "add x24, x24, x8, LSL #2\n" - "add x23, x23, x8, LSL #2\n" + "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n" + "add x26, x26, x27, LSL #3\n" + "ldr x17, [x26, #0x0]\n" + "ldr x16, [x26, #0x8]\n" + "ldr x15, [x26, #0x10]\n" + "ldr x14, [x26, #0x18]\n" + "ldr x13, [x26, #0x20]\n" + "ldr x12, [x26, #0x28]\n" + "cbnz x7, 116f\n" + "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x17, x17, x26, LSL #2\n" + "add x16, x16, x26, LSL #2\n" + "add x15, x15, x26, LSL #2\n" + "add x14, x14, x26, LSL #2\n" + "add x13, x13, x26, LSL #2\n" + "add x12, x12, x26, LSL #2\n" "b 116f\n" "115:" // Height 6: setup direct input - "mov x11, %x[input_ptr]\n" - "add x9, x11, x8, LSL #2\n" - "add x27, x9, x8, LSL #2\n" - "add x25, x27, x8, LSL #2\n" - "add x24, x25, x8, LSL #2\n" - "add x23, x24, x8, LSL #2\n" + "mov x17, %x[input_ptr]\n" + "add x16, x17, x27, LSL #2\n" + "add x15, x16, x27, LSL #2\n" + "add x14, x15, x27, LSL #2\n" + "add x13, x14, x27, LSL #2\n" + "add x12, x13, x27, LSL #2\n" "116:" // Height 6: input setup done - "cmp x12, #0x4\n" + "cmp x8, #0x4\n" "blt 119f\n" - "ldr q0, [x11, #0x0]\n" - "ldr q1, [x9, #0x0]\n" - "cmp x12, #0x8\n" - "ldr q2, [x27, #0x0]\n" - "ldr q3, [x25, #0x0]\n" - "ldr q4, [x24, #0x0]\n" - "ldr q5, [x23, #0x0]\n" - "ldr q8, [x16, #0x0]\n" + "ldr q0, [x17, #0x0]\n" + "cmp x8, #0x8\n" + "ldr q1, [x16, #0x0]\n" + "ldr q2, [x15, #0x0]\n" + "ldr q3, [x14, #0x0]\n" + "ldr q4, [x13, #0x0]\n" + "ldr q5, [x12, #0x0]\n" + "ldr q8, [x5, #0x0]\n" + "ldr q9, [x5, #0x10]\n" + "ldr q10, [x5, #0x20]\n" + "ldr q11, [x5, #0x30]\n" "blt 118f\n" "117:" // Height 6: Multiply loop: Main loop head "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr d9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr x8, [x16, #0x18]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr d10, [x16, #0x20]\n" + "add x15, x15, #0x10\n" "fmla v27.4s, v8.4s, v3.s[0]\n" - "ldr x21, [x16, #0x28]\n" + "add x14, x14, #0x10\n" "fmla v28.4s, v8.4s, v4.s[0]\n" - "mov v9.d[1], x8\n" + "add x13, x13, #0x10\n" "fmla v29.4s, v8.4s, v5.s[0]\n" - "ldr d11, [x16, #0x30]\n" + "add x12, x12, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "mov v10.d[1], x21\n" + "add x5, x5, #0x40\n" + "ldr d8, [x5, #0x0]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "ldr x8, [x16, #0x38]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "add x11, x11, #0x10\n" + "ldr x9, [x5, #0x8]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "prfm pldl1keep, [x11, #0x80]\n" + "ldr x28, [x5, #0x18]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" - "mov v11.d[1], x8\n" + "ldr x27, [x5, #0x28]\n" "fmla v29.4s, v9.4s, v5.s[1]\n" - "ldr x10, [x11, #0x8]\n" + "ldr d9, [x5, #0x10]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "add x9, x9, #0x10\n" + "ldr x26, [x17, #0x8]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "ldr x11, [x16, #0x8]\n" "fmla v26.4s, v10.4s, v2.s[2]\n" - "ldr x28, [x9, #0x8]\n" + "ldr x10, [x15, #0x8]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" - "add x27, x27, #0x10\n" + "sub x8, x8, #0x4\n" "fmla v28.4s, v10.4s, v4.s[2]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "cmp x8, #0x8\n" "fmla v29.4s, v10.4s, v5.s[2]\n" - "ldr x26, [x27, #0x8]\n" + "ldr d10, [x5, #0x20]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" - "ldr d0, [x11, #0x0]\n" + "ldr d0, [x17, #0x0]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" - "ldr d1, [x9, #0x0]\n" + "ldr d1, [x16, #0x0]\n" "fmla v26.4s, v11.4s, v2.s[3]\n" - "ldr d2, [x27, #0x0]\n" + "ldr d2, [x15, #0x0]\n" "fmla v27.4s, v11.4s, v3.s[3]\n" - "mov v0.d[1], x10\n" + "ldr d3, [x14, #0x0]\n" "fmla v28.4s, v11.4s, v4.s[3]\n" - "mov v1.d[1], x28\n" + "ldr d4, [x13, #0x0]\n" "fmla v29.4s, v11.4s, v5.s[3]\n" - "mov v2.d[1], x26\n" - "add x25, x25, #0x10\n" - "add x24, x24, #0x10\n" - "prfm pldl1keep, [x25, #0x80]\n" - "add x23, x23, #0x10\n" - "prfm pldl1keep, [x24, #0x80]\n" - "sub x12, x12, #0x4\n" - "prfm pldl1keep, [x23, #0x80]\n" - "cmp x12, #0x8\n" - "ldr d3, [x25, #0x0]\n" - "add x16, x16, #0x40\n" - "ldr x8, [x25, #0x8]\n" - "ldr d4, [x24, #0x0]\n" - "ldr x21, [x24, #0x8]\n" - "mov v3.d[1], x8\n" - "ldr d5, [x23, #0x0]\n" - "ldr x8, [x23, #0x8]\n" - "mov v4.d[1], x21\n" - "ldr d8, [x16, #0x0]\n" - "ldr x26, [x16, #0x8]\n" - "mov v5.d[1], x8\n" - "mov v8.d[1], x26\n" + "ldr d5, [x12, #0x0]\n" + "ldr d11, [x5, #0x30]\n" + "mov v8.d[1], x9\n" + "ldr x9, [x14, #0x8]\n" + "mov v9.d[1], x28\n" + "ldr x28, [x13, #0x8]\n" + "mov v10.d[1], x27\n" + "ldr x27, [x12, #0x8]\n" + "mov v0.d[1], x26\n" + "ldr x26, [x5, #0x38]\n" + "mov v1.d[1], x11\n" + "prfm pldl1keep, [x17, #0x80]\n" + "mov v2.d[1], x10\n" + "prfm pldl1keep, [x16, #0x80]\n" + "mov v3.d[1], x9\n" + "prfm pldl1keep, [x15, #0x80]\n" + "mov v4.d[1], x28\n" + "prfm pldl1keep, [x14, #0x80]\n" + "mov v5.d[1], x27\n" + "prfm pldl1keep, [x13, #0x80]\n" + "mov v11.d[1], x26\n" + "prfm pldl1keep, [x12, #0x80]\n" "bge 117b\n" "118:" // Height 6: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr q9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr q10, [x16, #0x20]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr q11, [x16, #0x30]\n" + "add x15, x15, #0x10\n" "fmla v27.4s, v8.4s, v3.s[0]\n" - "sub x12, x12, #0x4\n" + "add x14, x14, #0x10\n" "fmla v28.4s, v8.4s, v4.s[0]\n" - "add x11, x11, #0x10\n" + "add x13, x13, #0x10\n" "fmla v29.4s, v8.4s, v5.s[0]\n" - "prfm pldl1keep, [x11, #0x80]\n" + "add x12, x12, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "add x9, x9, #0x10\n" + "sub x8, x8, #0x4\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "prfm pldl1keep, [x17, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "add x27, x27, #0x10\n" + "prfm pldl1keep, [x16, #0x80]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x15, #0x80]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" - "add x25, x25, #0x10\n" + "prfm pldl1keep, [x14, #0x80]\n" "fmla v29.4s, v9.4s, v5.s[1]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x13, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "add x24, x24, #0x10\n" + "prfm pldl1keep, [x12, #0x80]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "add x5, x5, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" - "add x23, x23, #0x10\n" "fmla v27.4s, v10.4s, v3.s[2]\n" - "prfm pldl1keep, [x23, #0x80]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" - "add x16, x16, #0x40\n" "fmla v29.4s, v10.4s, v5.s[2]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" @@ -1361,108 +1360,108 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "fmla v28.4s, v11.4s, v4.s[3]\n" "fmla v29.4s, v11.4s, v5.s[3]\n" "119:" // Height 6: Multiply loop: Main loop skip - "cbz x12, 121f\n" + "cbz x8, 121f\n" "120:" // Height 6: Multiply loop: Odd block loop - "ldr s0, [x11], #0x4\n" - "sub x12, x12, #0x1\n" - "ldr s1, [x9], #0x4\n" - "ldr s2, [x27], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr s5, [x23], #0x4\n" - "ldr q12, [x16, #0x0]\n" - "add x16, x16, #0x10\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" - "fmla v28.4s, v12.4s, v4.s[0]\n" - "fmla v29.4s, v12.4s, v5.s[0]\n" - "cbnz x12, 120b\n" + "ldr s22, [x17], #0x4\n" + "sub x8, x8, #0x1\n" + "ldr s21, [x16], #0x4\n" + "ldr s20, [x15], #0x4\n" + "ldr s19, [x14], #0x4\n" + "ldr s18, [x13], #0x4\n" + "ldr s17, [x12], #0x4\n" + "ldr q16, [x5, #0x0]\n" + "fmla v24.4s, v16.4s, v22.s[0]\n" + "fmla v25.4s, v16.4s, v21.s[0]\n" + "add x5, x5, #0x10\n" + "fmla v26.4s, v16.4s, v20.s[0]\n" + "fmla v27.4s, v16.4s, v19.s[0]\n" + "fmla v28.4s, v16.4s, v18.s[0]\n" + "fmla v29.4s, v16.4s, v17.s[0]\n" + "cbnz x8, 120b\n" "121:" // Height 6: Multiply loop: No odd multiplies - "ldr w8, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x13, x13, #0x1\n" - "cmp x13, x8\n" + "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x7, x7, #0x1\n" + "cmp x7, x26\n" "bne 114b\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x14, #0x0]\n" - "add x27, x14, x8, LSL #2\n" - "prfm pstl1keep, [x27, #0x0]\n" - "add x26, x27, x8, LSL #2\n" - "prfm pstl1keep, [x26, #0x0]\n" - "add x25, x26, x8, LSL #2\n" - "prfm pstl1keep, [x25, #0x0]\n" - "add x24, x25, x8, LSL #2\n" - "prfm pstl1keep, [x24, #0x0]\n" - "add x23, x24, x8, LSL #2\n" - "prfm pstl1keep, [x23, #0x0]\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "add x12, x13, x26, LSL #2\n" + "add x11, x12, x26, LSL #2\n" + "add x10, x11, x26, LSL #2\n" + "add x9, x10, x26, LSL #2\n" + "prfm pstl1keep, [x6, #0x0]\n" + "prfm pstl1keep, [x13, #0x0]\n" + "prfm pstl1keep, [x12, #0x0]\n" + "prfm pstl1keep, [x11, #0x0]\n" + "prfm pstl1keep, [x10, #0x0]\n" + "prfm pstl1keep, [x9, #0x0]\n" "tbz %x[flags], #1, 122f\n" - "add x20, %x[args_ptr], %[offset_min]\n" - "add x8, %x[args_ptr], %[offset_max]\n" - "ld1r { v17.4s }, [x20]\n" - "ld1r { v16.4s }, [x8]\n" + "add x26, %x[args_ptr], %[offset_max]\n" + "ld1r { v16.4s }, [x26]\n" "fmin v24.4s, v24.4s, v16.4s\n" "fmin v25.4s, v25.4s, v16.4s\n" "fmin v26.4s, v26.4s, v16.4s\n" "fmin v27.4s, v27.4s, v16.4s\n" "fmin v28.4s, v28.4s, v16.4s\n" "fmin v29.4s, v29.4s, v16.4s\n" - "fmax v24.4s, v24.4s, v17.4s\n" - "fmax v25.4s, v25.4s, v17.4s\n" - "fmax v26.4s, v26.4s, v17.4s\n" - "fmax v27.4s, v27.4s, v17.4s\n" - "fmax v28.4s, v28.4s, v17.4s\n" - "fmax v29.4s, v29.4s, v17.4s\n" + "add x26, %x[args_ptr], %[offset_min]\n" + "ld1r { v16.4s }, [x26]\n" + "fmax v24.4s, v24.4s, v16.4s\n" + "fmax v25.4s, v25.4s, v16.4s\n" + "fmax v26.4s, v26.4s, v16.4s\n" + "fmax v27.4s, v27.4s, v16.4s\n" + "fmax v28.4s, v28.4s, v16.4s\n" + "fmax v29.4s, v29.4s, v16.4s\n" "122:" // Height 6: No activation - "cmp x17, #0x4\n" + "cmp x4, #0x4\n" "bge 125f\n" - "tbz x17, #1, 123f\n" - "str d24, [x14], #0x8\n" - "str d25, [x27], #0x8\n" - "str d26, [x26], #0x8\n" - "str d27, [x25], #0x8\n" - "str d28, [x24], #0x8\n" - "str d29, [x23], #0x8\n" - "tbz x17, #0, 124f\n" - "st1 { v24.s }[2], [x14]\n" - "st1 { v25.s }[2], [x27]\n" - "st1 { v26.s }[2], [x26]\n" - "st1 { v27.s }[2], [x25]\n" - "st1 { v28.s }[2], [x24]\n" - "st1 { v29.s }[2], [x23]\n" + "tbz x4, #1, 123f\n" + "str d24, [x6], #0x8\n" + "str d25, [x13], #0x8\n" + "str d26, [x12], #0x8\n" + "str d27, [x11], #0x8\n" + "str d28, [x10], #0x8\n" + "str d29, [x9], #0x8\n" + "tbz x4, #0, 124f\n" + "st1 { v24.s }[2], [x6]\n" + "st1 { v25.s }[2], [x13]\n" + "st1 { v26.s }[2], [x12]\n" + "st1 { v27.s }[2], [x11]\n" + "st1 { v28.s }[2], [x10]\n" + "st1 { v29.s }[2], [x9]\n" "b 124f\n" "123:" // Height 6: Partial direct writeback: partial_1_0 - "str s24, [x14, #0x0]\n" - "str s25, [x27, #0x0]\n" - "str s26, [x26, #0x0]\n" - "str s27, [x25, #0x0]\n" - "str s28, [x24, #0x0]\n" - "str s29, [x23, #0x0]\n" + "str s24, [x6, #0x0]\n" + "str s25, [x13, #0x0]\n" + "str s26, [x12, #0x0]\n" + "str s27, [x11, #0x0]\n" + "str s28, [x10, #0x0]\n" + "str s29, [x9, #0x0]\n" "124:" // Height 6: Partial direct writeback: Done "b 126f\n" "125:" // Height 6: Full writeback - "str q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" - "str q25, [x27, #0x0]\n" - "str q26, [x26, #0x0]\n" - "str q27, [x25, #0x0]\n" - "str q28, [x24, #0x0]\n" - "str q29, [x23, #0x0]\n" + "str q24, [x6, #0x0]\n" + "add x6, x6, #0x10\n" + "str q25, [x13, #0x0]\n" + "str q26, [x12, #0x0]\n" + "str q27, [x11, #0x0]\n" + "str q28, [x10, #0x0]\n" + "str q29, [x9, #0x0]\n" "126:" // Height 6: Writeback done - "subs x17, x17, #0x4\n" + "subs x4, x4, #0x4\n" "bgt 107b\n" "b 170f\n" "127:" // Height 7 - "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" - "mov x15, %x[bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[output_ptr]\n" + "mov x3, %x[bias]\n" + "ldr x4, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x6, %x[output_ptr]\n" "128:" // Height 7: Column loop - "cbz x15, 129f\n" - "ldr q24, [x15, #0x0]\n" - "add x15, x15, #0x10\n" + "cbz x3, 129f\n" + "ldr q24, [x3, #0x0]\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x3, x3, #0x10\n" "mov v27.16b, v24.16b\n" "mov v28.16b, v24.16b\n" "mov v29.16b, v24.16b\n" @@ -1470,53 +1469,53 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "b 134f\n" "129:" // Height 7: no bias "tbz %x[flags], #0, 133f\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x17, #0x4\n" - "add x27, x14, x8, LSL #2\n" - "add x26, x27, x8, LSL #2\n" - "add x25, x26, x8, LSL #2\n" - "add x24, x25, x8, LSL #2\n" - "add x23, x24, x8, LSL #2\n" - "add x22, x23, x8, LSL #2\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "add x12, x13, x26, LSL #2\n" + "add x11, x12, x26, LSL #2\n" + "add x10, x11, x26, LSL #2\n" + "add x9, x10, x26, LSL #2\n" + "cmp x4, #0x4\n" + "add x28, x9, x26, LSL #2\n" "bge 132f\n" - "tbz x17, #1, 130f\n" - "ldr d24, [x14], #0x8\n" - "ldr d25, [x27], #0x8\n" - "mov x8, #0x8\n" - "ldr d26, [x26], #0x8\n" - "ldr d27, [x25], #0x8\n" - "ldr d28, [x24], #0x8\n" - "ldr d29, [x23], #0x8\n" - "ldr d30, [x22], #0x8\n" - "tbz x17, #0, 131f\n" - "ld1 { v24.s }[2], [x14]\n" - "ld1 { v25.s }[2], [x27]\n" - "ld1 { v26.s }[2], [x26]\n" - "ld1 { v27.s }[2], [x25]\n" - "ld1 { v28.s }[2], [x24]\n" - "ld1 { v29.s }[2], [x23]\n" - "ld1 { v30.s }[2], [x22]\n" + "tbz x4, #1, 130f\n" + "ldr d24, [x6], #0x8\n" + "mov x26, #0x8\n" + "ldr d25, [x13], #0x8\n" + "ldr d26, [x12], #0x8\n" + "ldr d27, [x11], #0x8\n" + "ldr d28, [x10], #0x8\n" + "ldr d29, [x9], #0x8\n" + "ldr d30, [x28], #0x8\n" + "tbz x4, #0, 131f\n" + "ld1 { v24.s }[2], [x6]\n" + "ld1 { v25.s }[2], [x13]\n" + "ld1 { v26.s }[2], [x12]\n" + "ld1 { v27.s }[2], [x11]\n" + "ld1 { v28.s }[2], [x10]\n" + "ld1 { v29.s }[2], [x9]\n" + "ld1 { v30.s }[2], [x28]\n" "b 131f\n" "130:" // Height 7: Partial accumulate: partial_1_0 - "ldr s24, [x14, #0x0]\n" - "mov x8, #0x0\n" - "ldr s25, [x27, #0x0]\n" - "ldr s26, [x26, #0x0]\n" - "ldr s27, [x25, #0x0]\n" - "ldr s28, [x24, #0x0]\n" - "ldr s29, [x23, #0x0]\n" - "ldr s30, [x22, #0x0]\n" + "ldr s24, [x6, #0x0]\n" + "mov x26, #0x0\n" + "ldr s25, [x13, #0x0]\n" + "ldr s26, [x12, #0x0]\n" + "ldr s27, [x11, #0x0]\n" + "ldr s28, [x10, #0x0]\n" + "ldr s29, [x9, #0x0]\n" + "ldr s30, [x28, #0x0]\n" "131:" // Height 7: Partial accumulate: Done - "sub x14, x14, x8\n" + "sub x6, x6, x26\n" "b 134f\n" "132:" // Height 7: full accumulate - "ldr q24, [x14, #0x0]\n" - "ldr q25, [x27, #0x0]\n" - "ldr q26, [x26, #0x0]\n" - "ldr q27, [x25, #0x0]\n" - "ldr q28, [x24, #0x0]\n" - "ldr q29, [x23, #0x0]\n" - "ldr q30, [x22, #0x0]\n" + "ldr q24, [x6, #0x0]\n" + "ldr q25, [x13, #0x0]\n" + "ldr q26, [x12, #0x0]\n" + "ldr q27, [x11, #0x0]\n" + "ldr q28, [x10, #0x0]\n" + "ldr q29, [x9, #0x0]\n" + "ldr q30, [x28, #0x0]\n" "b 134f\n" "133:" // Height 7: no accumulate "movi v24.16b, #0x0\n" @@ -1527,171 +1526,171 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "movi v29.16b, #0x0\n" "movi v30.16b, #0x0\n" "134:" // Height 7: setup done - "mov x13, #0x0\n" + "mov x7, #0x0\n" "135:" // Height 7: String loop - "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w12, [x20, x13, LSL #0x2]\n" + "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr w8, [x26, x7, LSL #0x2]\n" + "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 136f\n" - "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n" - "add x20, x20, x8, LSL #3\n" - "ldr x11, [x20, #0x0]\n" - "ldr x9, [x20, #0x8]\n" - "ldr x27, [x20, #0x10]\n" - "ldr x25, [x20, #0x18]\n" - "ldr x24, [x20, #0x20]\n" - "ldr x23, [x20, #0x28]\n" - "ldr x22, [x20, #0x30]\n" - "cbnz x13, 137f\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x11, x11, x8, LSL #2\n" - "add x9, x9, x8, LSL #2\n" - "add x27, x27, x8, LSL #2\n" - "add x25, x25, x8, LSL #2\n" - "add x24, x24, x8, LSL #2\n" - "add x23, x23, x8, LSL #2\n" - "add x22, x22, x8, LSL #2\n" + "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n" + "add x26, x26, x27, LSL #3\n" + "ldr x17, [x26, #0x0]\n" + "ldr x16, [x26, #0x8]\n" + "ldr x15, [x26, #0x10]\n" + "ldr x14, [x26, #0x18]\n" + "ldr x13, [x26, #0x20]\n" + "ldr x12, [x26, #0x28]\n" + "ldr x11, [x26, #0x30]\n" + "cbnz x7, 137f\n" + "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x17, x17, x26, LSL #2\n" + "add x16, x16, x26, LSL #2\n" + "add x15, x15, x26, LSL #2\n" + "add x14, x14, x26, LSL #2\n" + "add x13, x13, x26, LSL #2\n" + "add x12, x12, x26, LSL #2\n" + "add x11, x11, x26, LSL #2\n" "b 137f\n" "136:" // Height 7: setup direct input - "mov x11, %x[input_ptr]\n" - "add x9, x11, x8, LSL #2\n" - "add x27, x9, x8, LSL #2\n" - "add x25, x27, x8, LSL #2\n" - "add x24, x25, x8, LSL #2\n" - "add x23, x24, x8, LSL #2\n" - "add x22, x23, x8, LSL #2\n" + "mov x17, %x[input_ptr]\n" + "add x16, x17, x27, LSL #2\n" + "add x15, x16, x27, LSL #2\n" + "add x14, x15, x27, LSL #2\n" + "add x13, x14, x27, LSL #2\n" + "add x12, x13, x27, LSL #2\n" + "add x11, x12, x27, LSL #2\n" "137:" // Height 7: input setup done - "cmp x12, #0x4\n" + "cmp x8, #0x4\n" "blt 140f\n" - "ldr q0, [x11, #0x0]\n" - "ldr q1, [x9, #0x0]\n" - "cmp x12, #0x8\n" - "ldr q2, [x27, #0x0]\n" - "ldr q3, [x25, #0x0]\n" - "ldr q4, [x24, #0x0]\n" - "ldr q5, [x23, #0x0]\n" - "ldr q6, [x22, #0x0]\n" - "ldr q8, [x16, #0x0]\n" + "ldr q0, [x17, #0x0]\n" + "cmp x8, #0x8\n" + "ldr q1, [x16, #0x0]\n" + "ldr q2, [x15, #0x0]\n" + "ldr q3, [x14, #0x0]\n" + "ldr q4, [x13, #0x0]\n" + "ldr q5, [x12, #0x0]\n" + "ldr q6, [x11, #0x0]\n" + "ldr q8, [x5, #0x0]\n" + "ldr q9, [x5, #0x10]\n" + "ldr q10, [x5, #0x20]\n" + "ldr q11, [x5, #0x30]\n" "blt 139f\n" "138:" // Height 7: Multiply loop: Main loop head "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr d9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr x8, [x16, #0x18]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr d10, [x16, #0x20]\n" + "add x15, x15, #0x10\n" "fmla v27.4s, v8.4s, v3.s[0]\n" - "ldr x21, [x16, #0x28]\n" + "add x14, x14, #0x10\n" "fmla v28.4s, v8.4s, v4.s[0]\n" - "mov v9.d[1], x8\n" + "add x13, x13, #0x10\n" "fmla v29.4s, v8.4s, v5.s[0]\n" - "ldr d11, [x16, #0x30]\n" + "add x12, x12, #0x10\n" "fmla v30.4s, v8.4s, v6.s[0]\n" - "mov v10.d[1], x21\n" + "add x11, x11, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "ldr x8, [x16, #0x38]\n" + "add x5, x5, #0x40\n" + "ldr d8, [x5, #0x0]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "add x11, x11, #0x10\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "prfm pldl1keep, [x11, #0x80]\n" + "ldr x26, [x5, #0x8]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "mov v11.d[1], x8\n" + "ldr x10, [x5, #0x18]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" - "ldr x10, [x11, #0x8]\n" + "ldr x9, [x5, #0x28]\n" "fmla v29.4s, v9.4s, v5.s[1]\n" - "add x9, x9, #0x10\n" + "ldr x28, [x17, #0x8]\n" "fmla v30.4s, v9.4s, v6.s[1]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "ldr d9, [x5, #0x10]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "ldr x28, [x9, #0x8]\n" + "ldr x27, [x16, #0x8]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "add x27, x27, #0x10\n" + "sub x8, x8, #0x4\n" "fmla v26.4s, v10.4s, v2.s[2]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "cmp x8, #0x8\n" "fmla v27.4s, v10.4s, v3.s[2]\n" - "ldr x26, [x27, #0x8]\n" + "mov v8.d[1], x26\n" "fmla v28.4s, v10.4s, v4.s[2]\n" - "add x25, x25, #0x10\n" + "ldr x26, [x15, #0x8]\n" "fmla v29.4s, v10.4s, v5.s[2]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x17, #0x80]\n" "fmla v30.4s, v10.4s, v6.s[2]\n" - "ldr x8, [x25, #0x8]\n" + "ldr d10, [x5, #0x20]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" - "ldr d0, [x11, #0x0]\n" + "ldr d0, [x17, #0x0]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" - "ldr d1, [x9, #0x0]\n" + "ldr d1, [x16, #0x0]\n" "fmla v26.4s, v11.4s, v2.s[3]\n" - "ldr d2, [x27, #0x0]\n" + "ldr d2, [x15, #0x0]\n" "fmla v27.4s, v11.4s, v3.s[3]\n" - "mov v0.d[1], x10\n" + "ldr d3, [x14, #0x0]\n" "fmla v28.4s, v11.4s, v4.s[3]\n" - "mov v1.d[1], x28\n" + "ldr d4, [x13, #0x0]\n" "fmla v29.4s, v11.4s, v5.s[3]\n" - "mov v2.d[1], x26\n" + "ldr d5, [x12, #0x0]\n" "fmla v30.4s, v11.4s, v6.s[3]\n" - "ldr d3, [x25, #0x0]\n" - "add x24, x24, #0x10\n" - "add x23, x23, #0x10\n" - "prfm pldl1keep, [x24, #0x80]\n" - "add x22, x22, #0x10\n" - "mov v3.d[1], x8\n" - "prfm pldl1keep, [x23, #0x80]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "sub x12, x12, #0x4\n" - "ldr d4, [x24, #0x0]\n" - "cmp x12, #0x8\n" - "ldr x21, [x24, #0x8]\n" - "add x16, x16, #0x40\n" - "ldr d8, [x16, #0x0]\n" - "ldr x26, [x16, #0x8]\n" - "mov v4.d[1], x21\n" - "ldr d5, [x23, #0x0]\n" - "ldr x8, [x23, #0x8]\n" - "mov v8.d[1], x26\n" - "ldr d6, [x22, #0x0]\n" - "ldr x21, [x22, #0x8]\n" - "mov v5.d[1], x8\n" - "mov v6.d[1], x21\n" + "ldr d6, [x11, #0x0]\n" + "ldr d11, [x5, #0x30]\n" + "mov v9.d[1], x10\n" + "ldr x10, [x14, #0x8]\n" + "mov v10.d[1], x9\n" + "ldr x9, [x13, #0x8]\n" + "mov v0.d[1], x28\n" + "ldr x28, [x12, #0x8]\n" + "mov v1.d[1], x27\n" + "ldr x27, [x11, #0x8]\n" + "mov v2.d[1], x26\n" + "ldr x26, [x5, #0x38]\n" + "mov v3.d[1], x10\n" + "prfm pldl1keep, [x16, #0x80]\n" + "mov v4.d[1], x9\n" + "prfm pldl1keep, [x15, #0x80]\n" + "mov v5.d[1], x28\n" + "prfm pldl1keep, [x14, #0x80]\n" + "mov v6.d[1], x27\n" + "prfm pldl1keep, [x13, #0x80]\n" + "mov v11.d[1], x26\n" + "prfm pldl1keep, [x12, #0x80]\n" + "prfm pldl1keep, [x11, #0x80]\n" "bge 138b\n" "139:" // Height 7: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr q9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr q10, [x16, #0x20]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr q11, [x16, #0x30]\n" + "add x15, x15, #0x10\n" "fmla v27.4s, v8.4s, v3.s[0]\n" - "sub x12, x12, #0x4\n" + "add x14, x14, #0x10\n" "fmla v28.4s, v8.4s, v4.s[0]\n" - "add x11, x11, #0x10\n" + "add x13, x13, #0x10\n" "fmla v29.4s, v8.4s, v5.s[0]\n" - "prfm pldl1keep, [x11, #0x80]\n" + "add x12, x12, #0x10\n" "fmla v30.4s, v8.4s, v6.s[0]\n" - "add x9, x9, #0x10\n" + "add x11, x11, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "sub x8, x8, #0x4\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "add x27, x27, #0x10\n" + "prfm pldl1keep, [x17, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x16, #0x80]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "add x25, x25, #0x10\n" + "prfm pldl1keep, [x15, #0x80]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x14, #0x80]\n" "fmla v29.4s, v9.4s, v5.s[1]\n" - "add x24, x24, #0x10\n" + "prfm pldl1keep, [x13, #0x80]\n" "fmla v30.4s, v9.4s, v6.s[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x12, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "add x23, x23, #0x10\n" + "prfm pldl1keep, [x11, #0x80]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "prfm pldl1keep, [x23, #0x80]\n" + "add x5, x5, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" - "add x22, x22, #0x10\n" "fmla v27.4s, v10.4s, v3.s[2]\n" - "prfm pldl1keep, [x22, #0x80]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" - "add x16, x16, #0x40\n" "fmla v29.4s, v10.4s, v5.s[2]\n" "fmla v30.4s, v10.4s, v6.s[2]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" @@ -1702,50 +1701,48 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "fmla v29.4s, v11.4s, v5.s[3]\n" "fmla v30.4s, v11.4s, v6.s[3]\n" "140:" // Height 7: Multiply loop: Main loop skip - "cbz x12, 142f\n" + "cbz x8, 142f\n" "141:" // Height 7: Multiply loop: Odd block loop - "ldr s0, [x11], #0x4\n" - "sub x12, x12, #0x1\n" - "ldr s1, [x9], #0x4\n" - "ldr s2, [x27], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr s5, [x23], #0x4\n" - "ldr s6, [x22], #0x4\n" - "ldr q12, [x16, #0x0]\n" - "add x16, x16, #0x10\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" - "fmla v28.4s, v12.4s, v4.s[0]\n" - "fmla v29.4s, v12.4s, v5.s[0]\n" - "fmla v30.4s, v12.4s, v6.s[0]\n" - "cbnz x12, 141b\n" + "ldr s23, [x17], #0x4\n" + "sub x8, x8, #0x1\n" + "ldr s22, [x16], #0x4\n" + "ldr s21, [x15], #0x4\n" + "ldr s20, [x14], #0x4\n" + "ldr s19, [x13], #0x4\n" + "ldr s18, [x12], #0x4\n" + "ldr s17, [x11], #0x4\n" + "ldr q16, [x5, #0x0]\n" + "fmla v24.4s, v16.4s, v23.s[0]\n" + "fmla v25.4s, v16.4s, v22.s[0]\n" + "add x5, x5, #0x10\n" + "fmla v26.4s, v16.4s, v21.s[0]\n" + "fmla v27.4s, v16.4s, v20.s[0]\n" + "fmla v28.4s, v16.4s, v19.s[0]\n" + "fmla v29.4s, v16.4s, v18.s[0]\n" + "fmla v30.4s, v16.4s, v17.s[0]\n" + "cbnz x8, 141b\n" "142:" // Height 7: Multiply loop: No odd multiplies - "ldr w8, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x13, x13, #0x1\n" - "cmp x13, x8\n" + "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x7, x7, #0x1\n" + "cmp x7, x26\n" "bne 135b\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x14, #0x0]\n" - "add x27, x14, x8, LSL #2\n" - "prfm pstl1keep, [x27, #0x0]\n" - "add x26, x27, x8, LSL #2\n" - "prfm pstl1keep, [x26, #0x0]\n" - "add x25, x26, x8, LSL #2\n" - "prfm pstl1keep, [x25, #0x0]\n" - "add x24, x25, x8, LSL #2\n" - "prfm pstl1keep, [x24, #0x0]\n" - "add x23, x24, x8, LSL #2\n" - "prfm pstl1keep, [x23, #0x0]\n" - "add x22, x23, x8, LSL #2\n" - "prfm pstl1keep, [x22, #0x0]\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "add x12, x13, x26, LSL #2\n" + "add x11, x12, x26, LSL #2\n" + "add x10, x11, x26, LSL #2\n" + "add x9, x10, x26, LSL #2\n" + "add x28, x9, x26, LSL #2\n" + "prfm pstl1keep, [x6, #0x0]\n" + "prfm pstl1keep, [x13, #0x0]\n" + "prfm pstl1keep, [x12, #0x0]\n" + "prfm pstl1keep, [x11, #0x0]\n" + "prfm pstl1keep, [x10, #0x0]\n" + "prfm pstl1keep, [x9, #0x0]\n" + "prfm pstl1keep, [x28, #0x0]\n" "tbz %x[flags], #1, 143f\n" - "add x20, %x[args_ptr], %[offset_min]\n" - "add x8, %x[args_ptr], %[offset_max]\n" - "ld1r { v17.4s }, [x20]\n" - "ld1r { v16.4s }, [x8]\n" + "add x26, %x[args_ptr], %[offset_max]\n" + "ld1r { v16.4s }, [x26]\n" "fmin v24.4s, v24.4s, v16.4s\n" "fmin v25.4s, v25.4s, v16.4s\n" "fmin v26.4s, v26.4s, v16.4s\n" @@ -1753,70 +1750,72 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "fmin v28.4s, v28.4s, v16.4s\n" "fmin v29.4s, v29.4s, v16.4s\n" "fmin v30.4s, v30.4s, v16.4s\n" - "fmax v24.4s, v24.4s, v17.4s\n" - "fmax v25.4s, v25.4s, v17.4s\n" - "fmax v26.4s, v26.4s, v17.4s\n" - "fmax v27.4s, v27.4s, v17.4s\n" - "fmax v28.4s, v28.4s, v17.4s\n" - "fmax v29.4s, v29.4s, v17.4s\n" - "fmax v30.4s, v30.4s, v17.4s\n" + "add x26, %x[args_ptr], %[offset_min]\n" + "ld1r { v16.4s }, [x26]\n" + "fmax v24.4s, v24.4s, v16.4s\n" + "fmax v25.4s, v25.4s, v16.4s\n" + "fmax v26.4s, v26.4s, v16.4s\n" + "fmax v27.4s, v27.4s, v16.4s\n" + "fmax v28.4s, v28.4s, v16.4s\n" + "fmax v29.4s, v29.4s, v16.4s\n" + "fmax v30.4s, v30.4s, v16.4s\n" "143:" // Height 7: No activation - "cmp x17, #0x4\n" + "cmp x4, #0x4\n" "bge 146f\n" - "tbz x17, #1, 144f\n" - "str d24, [x14], #0x8\n" - "str d25, [x27], #0x8\n" - "str d26, [x26], #0x8\n" - "str d27, [x25], #0x8\n" - "str d28, [x24], #0x8\n" - "str d29, [x23], #0x8\n" - "str d30, [x22], #0x8\n" - "tbz x17, #0, 145f\n" - "st1 { v24.s }[2], [x14]\n" - "st1 { v25.s }[2], [x27]\n" - "st1 { v26.s }[2], [x26]\n" - "st1 { v27.s }[2], [x25]\n" - "st1 { v28.s }[2], [x24]\n" - "st1 { v29.s }[2], [x23]\n" - "st1 { v30.s }[2], [x22]\n" + "tbz x4, #1, 144f\n" + "str d24, [x6], #0x8\n" + "str d25, [x13], #0x8\n" + "str d26, [x12], #0x8\n" + "str d27, [x11], #0x8\n" + "str d28, [x10], #0x8\n" + "str d29, [x9], #0x8\n" + "str d30, [x28], #0x8\n" + "tbz x4, #0, 145f\n" + "st1 { v24.s }[2], [x6]\n" + "st1 { v25.s }[2], [x13]\n" + "st1 { v26.s }[2], [x12]\n" + "st1 { v27.s }[2], [x11]\n" + "st1 { v28.s }[2], [x10]\n" + "st1 { v29.s }[2], [x9]\n" + "st1 { v30.s }[2], [x28]\n" "b 145f\n" "144:" // Height 7: Partial direct writeback: partial_1_0 - "str s24, [x14, #0x0]\n" - "str s25, [x27, #0x0]\n" - "str s26, [x26, #0x0]\n" - "str s27, [x25, #0x0]\n" - "str s28, [x24, #0x0]\n" - "str s29, [x23, #0x0]\n" - "str s30, [x22, #0x0]\n" + "str s24, [x6, #0x0]\n" + "str s25, [x13, #0x0]\n" + "str s26, [x12, #0x0]\n" + "str s27, [x11, #0x0]\n" + "str s28, [x10, #0x0]\n" + "str s29, [x9, #0x0]\n" + "str s30, [x28, #0x0]\n" "145:" // Height 7: Partial direct writeback: Done "b 147f\n" "146:" // Height 7: Full writeback - "str q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" - "str q25, [x27, #0x0]\n" - "str q26, [x26, #0x0]\n" - "str q27, [x25, #0x0]\n" - "str q28, [x24, #0x0]\n" - "str q29, [x23, #0x0]\n" - "str q30, [x22, #0x0]\n" + "str q24, [x6, #0x0]\n" + "add x6, x6, #0x10\n" + "str q25, [x13, #0x0]\n" + "str q26, [x12, #0x0]\n" + "str q27, [x11, #0x0]\n" + "str q28, [x10, #0x0]\n" + "str q29, [x9, #0x0]\n" + "str q30, [x28, #0x0]\n" "147:" // Height 7: Writeback done - "subs x17, x17, #0x4\n" + "subs x4, x4, #0x4\n" "bgt 128b\n" "b 170f\n" "148:" // Height 8 - "ldr x17, [%x[args_ptr], %[offsetof_N]]\n" - "mov x15, %x[bias]\n" - "ldr x16, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x14, %x[output_ptr]\n" - "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "mov x8, #0x20\n" - "madd %x[output_ptr], x20, x8, %x[output_ptr]\n" + "ldr x27, [%x[args_ptr], %[offsetof_output_offset]]\n" + "mov x26, #0x20\n" + "mov x3, %x[bias]\n" + "ldr x4, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x5, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "mov x6, %x[output_ptr]\n" + "madd %x[output_ptr], x27, x26, %x[output_ptr]\n" "149:" // Height 8: Column loop - "cbz x15, 150f\n" - "ldr q24, [x15, #0x0]\n" - "add x15, x15, #0x10\n" + "cbz x3, 150f\n" + "ldr q24, [x3, #0x0]\n" "mov v25.16b, v24.16b\n" "mov v26.16b, v24.16b\n" + "add x3, x3, #0x10\n" "mov v27.16b, v24.16b\n" "mov v28.16b, v24.16b\n" "mov v29.16b, v24.16b\n" @@ -1825,58 +1824,58 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "b 155f\n" "150:" // Height 8: no bias "tbz %x[flags], #0, 154f\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "cmp x17, #0x4\n" - "add x27, x14, x8, LSL #2\n" - "add x26, x27, x8, LSL #2\n" - "add x25, x26, x8, LSL #2\n" - "add x24, x25, x8, LSL #2\n" - "add x23, x24, x8, LSL #2\n" - "add x22, x23, x8, LSL #2\n" - "add x21, x22, x8, LSL #2\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "add x12, x13, x26, LSL #2\n" + "add x11, x12, x26, LSL #2\n" + "add x10, x11, x26, LSL #2\n" + "add x9, x10, x26, LSL #2\n" + "add x28, x9, x26, LSL #2\n" + "cmp x4, #0x4\n" + "add x27, x28, x26, LSL #2\n" "bge 153f\n" - "tbz x17, #1, 151f\n" - "ldr d24, [x14], #0x8\n" - "ldr d25, [x27], #0x8\n" - "mov x8, #0x8\n" - "ldr d26, [x26], #0x8\n" - "ldr d27, [x25], #0x8\n" - "ldr d28, [x24], #0x8\n" - "ldr d29, [x23], #0x8\n" - "ldr d30, [x22], #0x8\n" - "ldr d31, [x21], #0x8\n" - "tbz x17, #0, 152f\n" - "ld1 { v24.s }[2], [x14]\n" - "ld1 { v25.s }[2], [x27]\n" - "ld1 { v26.s }[2], [x26]\n" - "ld1 { v27.s }[2], [x25]\n" - "ld1 { v28.s }[2], [x24]\n" - "ld1 { v29.s }[2], [x23]\n" - "ld1 { v30.s }[2], [x22]\n" - "ld1 { v31.s }[2], [x21]\n" + "tbz x4, #1, 151f\n" + "ldr d24, [x6], #0x8\n" + "mov x26, #0x8\n" + "ldr d25, [x13], #0x8\n" + "ldr d26, [x12], #0x8\n" + "ldr d27, [x11], #0x8\n" + "ldr d28, [x10], #0x8\n" + "ldr d29, [x9], #0x8\n" + "ldr d30, [x28], #0x8\n" + "ldr d31, [x27], #0x8\n" + "tbz x4, #0, 152f\n" + "ld1 { v24.s }[2], [x6]\n" + "ld1 { v25.s }[2], [x13]\n" + "ld1 { v26.s }[2], [x12]\n" + "ld1 { v27.s }[2], [x11]\n" + "ld1 { v28.s }[2], [x10]\n" + "ld1 { v29.s }[2], [x9]\n" + "ld1 { v30.s }[2], [x28]\n" + "ld1 { v31.s }[2], [x27]\n" "b 152f\n" "151:" // Height 8: Partial accumulate: partial_1_0 - "ldr s24, [x14, #0x0]\n" - "mov x8, #0x0\n" - "ldr s25, [x27, #0x0]\n" - "ldr s26, [x26, #0x0]\n" - "ldr s27, [x25, #0x0]\n" - "ldr s28, [x24, #0x0]\n" - "ldr s29, [x23, #0x0]\n" - "ldr s30, [x22, #0x0]\n" - "ldr s31, [x21, #0x0]\n" + "ldr s24, [x6, #0x0]\n" + "mov x26, #0x0\n" + "ldr s25, [x13, #0x0]\n" + "ldr s26, [x12, #0x0]\n" + "ldr s27, [x11, #0x0]\n" + "ldr s28, [x10, #0x0]\n" + "ldr s29, [x9, #0x0]\n" + "ldr s30, [x28, #0x0]\n" + "ldr s31, [x27, #0x0]\n" "152:" // Height 8: Partial accumulate: Done - "sub x14, x14, x8\n" + "sub x6, x6, x26\n" "b 155f\n" "153:" // Height 8: full accumulate - "ldr q24, [x14, #0x0]\n" - "ldr q25, [x27, #0x0]\n" - "ldr q26, [x26, #0x0]\n" - "ldr q27, [x25, #0x0]\n" - "ldr q28, [x24, #0x0]\n" - "ldr q29, [x23, #0x0]\n" - "ldr q30, [x22, #0x0]\n" - "ldr q31, [x21, #0x0]\n" + "ldr q24, [x6, #0x0]\n" + "ldr q25, [x13, #0x0]\n" + "ldr q26, [x12, #0x0]\n" + "ldr q27, [x11, #0x0]\n" + "ldr q28, [x10, #0x0]\n" + "ldr q29, [x9, #0x0]\n" + "ldr q30, [x28, #0x0]\n" + "ldr q31, [x27, #0x0]\n" "b 155f\n" "154:" // Height 8: no accumulate "movi v24.16b, #0x0\n" @@ -1888,188 +1887,188 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "movi v30.16b, #0x0\n" "movi v31.16b, #0x0\n" "155:" // Height 8: setup done - "mov x13, #0x0\n" + "mov x7, #0x0\n" "156:" // Height 8: String loop - "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_offset]]\n" - "ldr w12, [x20, x13, LSL #0x2]\n" + "ldr x26, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr w8, [x26, x7, LSL #0x2]\n" + "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 157f\n" - "ldr x20, [%x[input_ptr], x13, LSL #0x3]\n" - "add x20, x20, x8, LSL #3\n" - "ldr x11, [x20, #0x0]\n" - "ldr x9, [x20, #0x8]\n" - "ldr x27, [x20, #0x10]\n" - "ldr x25, [x20, #0x18]\n" - "ldr x24, [x20, #0x20]\n" - "ldr x23, [x20, #0x28]\n" - "ldr x22, [x20, #0x30]\n" - "ldr x20, [x20, #0x38]\n" - "cbnz x13, 158f\n" - "ldr x8, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x11, x11, x8, LSL #2\n" - "add x9, x9, x8, LSL #2\n" - "add x27, x27, x8, LSL #2\n" - "add x25, x25, x8, LSL #2\n" - "add x24, x24, x8, LSL #2\n" - "add x23, x23, x8, LSL #2\n" - "add x22, x22, x8, LSL #2\n" - "add x20, x20, x8, LSL #2\n" + "ldr x26, [%x[input_ptr], x7, LSL #0x3]\n" + "add x26, x26, x27, LSL #3\n" + "ldr x17, [x26, #0x0]\n" + "ldr x16, [x26, #0x8]\n" + "ldr x15, [x26, #0x10]\n" + "ldr x14, [x26, #0x18]\n" + "ldr x13, [x26, #0x20]\n" + "ldr x12, [x26, #0x28]\n" + "ldr x11, [x26, #0x30]\n" + "ldr x27, [x26, #0x38]\n" + "cbnz x7, 158f\n" + "ldr x26, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x17, x17, x26, LSL #2\n" + "add x16, x16, x26, LSL #2\n" + "add x15, x15, x26, LSL #2\n" + "add x14, x14, x26, LSL #2\n" + "add x13, x13, x26, LSL #2\n" + "add x12, x12, x26, LSL #2\n" + "add x11, x11, x26, LSL #2\n" + "add x27, x27, x26, LSL #2\n" "b 158f\n" "157:" // Height 8: setup direct input - "mov x11, %x[input_ptr]\n" - "add x9, x11, x8, LSL #2\n" - "add x27, x9, x8, LSL #2\n" - "add x25, x27, x8, LSL #2\n" - "add x24, x25, x8, LSL #2\n" - "add x23, x24, x8, LSL #2\n" - "add x22, x23, x8, LSL #2\n" - "add x20, x22, x8, LSL #2\n" + "mov x17, %x[input_ptr]\n" + "add x16, x17, x27, LSL #2\n" + "add x15, x16, x27, LSL #2\n" + "add x14, x15, x27, LSL #2\n" + "add x13, x14, x27, LSL #2\n" + "add x12, x13, x27, LSL #2\n" + "add x11, x12, x27, LSL #2\n" + "add x27, x11, x27, LSL #2\n" "158:" // Height 8: input setup done - "cmp x12, #0x4\n" + "cmp x8, #0x4\n" "blt 161f\n" - "ldr q0, [x11, #0x0]\n" - "ldr q1, [x9, #0x0]\n" - "cmp x12, #0x8\n" - "ldr q2, [x27, #0x0]\n" - "ldr q3, [x25, #0x0]\n" - "ldr q4, [x24, #0x0]\n" - "ldr q5, [x23, #0x0]\n" - "ldr q6, [x22, #0x0]\n" - "ldr q7, [x20, #0x0]\n" - "ldr q8, [x16, #0x0]\n" + "ldr q0, [x17, #0x0]\n" + "cmp x8, #0x8\n" + "ldr q1, [x16, #0x0]\n" + "ldr q2, [x15, #0x0]\n" + "ldr q3, [x14, #0x0]\n" + "ldr q4, [x13, #0x0]\n" + "ldr q5, [x12, #0x0]\n" + "ldr q6, [x11, #0x0]\n" + "ldr q7, [x27, #0x0]\n" + "ldr q8, [x5, #0x0]\n" + "ldr q9, [x5, #0x10]\n" + "ldr q10, [x5, #0x20]\n" + "ldr q11, [x5, #0x30]\n" "blt 160f\n" "159:" // Height 8: Multiply loop: Main loop head "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr d9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr x8, [x16, #0x18]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr d10, [x16, #0x20]\n" + "add x15, x15, #0x10\n" "fmla v27.4s, v8.4s, v3.s[0]\n" - "ldr x21, [x16, #0x28]\n" + "add x14, x14, #0x10\n" "fmla v28.4s, v8.4s, v4.s[0]\n" - "mov v9.d[1], x8\n" + "add x13, x13, #0x10\n" "fmla v29.4s, v8.4s, v5.s[0]\n" - "ldr d11, [x16, #0x30]\n" + "add x12, x12, #0x10\n" "fmla v30.4s, v8.4s, v6.s[0]\n" - "mov v10.d[1], x21\n" + "add x11, x11, #0x10\n" "fmla v31.4s, v8.4s, v7.s[0]\n" - "ldr x8, [x16, #0x38]\n" + "add x27, x27, #0x10\n" "fmla v24.4s, v9.4s, v0.s[1]\n" - "add x11, x11, #0x10\n" + "add x5, x5, #0x40\n" + "ldr d8, [x5, #0x0]\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "prfm pldl1keep, [x11, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "mov v11.d[1], x8\n" + "ldr x26, [x5, #0x8]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "ldr x10, [x11, #0x8]\n" + "sub x8, x8, #0x4\n" "fmla v28.4s, v9.4s, v4.s[1]\n" - "add x9, x9, #0x10\n" + "cmp x8, #0x8\n" "fmla v29.4s, v9.4s, v5.s[1]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "mov v8.d[1], x26\n" "fmla v30.4s, v9.4s, v6.s[1]\n" - "ldr x28, [x9, #0x8]\n" + "ldr x26, [x5, #0x18]\n" "fmla v31.4s, v9.4s, v7.s[1]\n" - "add x27, x27, #0x10\n" + "ldr d9, [x5, #0x10]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x17, #0x80]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "ldr x26, [x27, #0x8]\n" + "prfm pldl1keep, [x16, #0x80]\n" "fmla v26.4s, v10.4s, v2.s[2]\n" - "add x25, x25, #0x10\n" + "prfm pldl1keep, [x15, #0x80]\n" "fmla v27.4s, v10.4s, v3.s[2]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "mov v9.d[1], x26\n" "fmla v28.4s, v10.4s, v4.s[2]\n" - "ldr x8, [x25, #0x8]\n" + "ldr x26, [x5, #0x28]\n" "fmla v29.4s, v10.4s, v5.s[2]\n" - "add x24, x24, #0x10\n" + "prfm pldl1keep, [x14, #0x80]\n" "fmla v30.4s, v10.4s, v6.s[2]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x13, #0x80]\n" "fmla v31.4s, v10.4s, v7.s[2]\n" - "ldr x21, [x24, #0x8]\n" + "ldr d10, [x5, #0x20]\n" "fmla v24.4s, v11.4s, v0.s[3]\n" - "ldr d0, [x11, #0x0]\n" + "ldr d0, [x17, #0x0]\n" "fmla v25.4s, v11.4s, v1.s[3]\n" - "ldr d1, [x9, #0x0]\n" + "ldr d1, [x16, #0x0]\n" "fmla v26.4s, v11.4s, v2.s[3]\n" - "ldr d2, [x27, #0x0]\n" + "ldr d2, [x15, #0x0]\n" "fmla v27.4s, v11.4s, v3.s[3]\n" - "mov v0.d[1], x10\n" + "ldr d3, [x14, #0x0]\n" "fmla v28.4s, v11.4s, v4.s[3]\n" - "mov v1.d[1], x28\n" + "ldr d4, [x13, #0x0]\n" "fmla v29.4s, v11.4s, v5.s[3]\n" - "mov v2.d[1], x26\n" + "ldr d5, [x12, #0x0]\n" "fmla v30.4s, v11.4s, v6.s[3]\n" - "ldr d3, [x25, #0x0]\n" + "ldr d6, [x11, #0x0]\n" "fmla v31.4s, v11.4s, v7.s[3]\n" - "ldr d4, [x24, #0x0]\n" - "add x23, x23, #0x10\n" - "add x22, x22, #0x10\n" - "mov v3.d[1], x8\n" - "prfm pldl1keep, [x23, #0x80]\n" - "mov v4.d[1], x21\n" - "prfm pldl1keep, [x22, #0x80]\n" - "ldr d5, [x23, #0x0]\n" - "add x20, x20, #0x10\n" - "prfm pldl1keep, [x20, #0x80]\n" - "sub x12, x12, #0x4\n" - "ldr x8, [x23, #0x8]\n" - "cmp x12, #0x8\n" - "ldr d6, [x22, #0x0]\n" - "add x16, x16, #0x40\n" - "ldr d8, [x16, #0x0]\n" - "mov v5.d[1], x8\n" + "ldr d7, [x27, #0x0]\n" + "ldr d11, [x5, #0x30]\n" + "mov v10.d[1], x26\n" + "ldr x26, [x17, #0x8]\n" + "mov v0.d[1], x26\n" "ldr x26, [x16, #0x8]\n" - "ldr x21, [x22, #0x8]\n" - "ldr d7, [x20, #0x0]\n" - "mov v8.d[1], x26\n" - "ldr x8, [x20, #0x8]\n" - "mov v6.d[1], x21\n" - "mov v7.d[1], x8\n" + "mov v1.d[1], x26\n" + "ldr x26, [x15, #0x8]\n" + "mov v2.d[1], x26\n" + "ldr x26, [x14, #0x8]\n" + "mov v3.d[1], x26\n" + "ldr x26, [x13, #0x8]\n" + "mov v4.d[1], x26\n" + "ldr x26, [x12, #0x8]\n" + "mov v5.d[1], x26\n" + "ldr x26, [x11, #0x8]\n" + "mov v6.d[1], x26\n" + "ldr x26, [x27, #0x8]\n" + "mov v7.d[1], x26\n" + "ldr x26, [x5, #0x38]\n" + "mov v11.d[1], x26\n" + "prfm pldl1keep, [x12, #0x80]\n" + "prfm pldl1keep, [x11, #0x80]\n" + "prfm pldl1keep, [x27, #0x80]\n" "bge 159b\n" "160:" // Height 8: Multiply loop: Single iteration only "fmla v24.4s, v8.4s, v0.s[0]\n" - "ldr q9, [x16, #0x10]\n" + "add x17, x17, #0x10\n" "fmla v25.4s, v8.4s, v1.s[0]\n" - "ldr q10, [x16, #0x20]\n" + "add x16, x16, #0x10\n" "fmla v26.4s, v8.4s, v2.s[0]\n" - "ldr q11, [x16, #0x30]\n" + "add x15, x15, #0x10\n" "fmla v27.4s, v8.4s, v3.s[0]\n" - "sub x12, x12, #0x4\n" + "add x14, x14, #0x10\n" "fmla v28.4s, v8.4s, v4.s[0]\n" - "add x11, x11, #0x10\n" + "add x13, x13, #0x10\n" "fmla v29.4s, v8.4s, v5.s[0]\n" - "prfm pldl1keep, [x11, #0x80]\n" + "add x12, x12, #0x10\n" "fmla v30.4s, v8.4s, v6.s[0]\n" - "add x9, x9, #0x10\n" + "add x11, x11, #0x10\n" "fmla v31.4s, v8.4s, v7.s[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" - "fmla v24.4s, v9.4s, v0.s[1]\n" "add x27, x27, #0x10\n" + "fmla v24.4s, v9.4s, v0.s[1]\n" + "sub x8, x8, #0x4\n" "fmla v25.4s, v9.4s, v1.s[1]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x17, #0x80]\n" "fmla v26.4s, v9.4s, v2.s[1]\n" - "add x25, x25, #0x10\n" + "prfm pldl1keep, [x16, #0x80]\n" "fmla v27.4s, v9.4s, v3.s[1]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x15, #0x80]\n" "fmla v28.4s, v9.4s, v4.s[1]\n" - "add x24, x24, #0x10\n" + "prfm pldl1keep, [x14, #0x80]\n" "fmla v29.4s, v9.4s, v5.s[1]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x13, #0x80]\n" "fmla v30.4s, v9.4s, v6.s[1]\n" - "add x23, x23, #0x10\n" + "prfm pldl1keep, [x12, #0x80]\n" "fmla v31.4s, v9.4s, v7.s[1]\n" - "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x11, #0x80]\n" "fmla v24.4s, v10.4s, v0.s[2]\n" - "add x22, x22, #0x10\n" + "prfm pldl1keep, [x27, #0x80]\n" "fmla v25.4s, v10.4s, v1.s[2]\n" - "prfm pldl1keep, [x22, #0x80]\n" + "add x5, x5, #0x40\n" "fmla v26.4s, v10.4s, v2.s[2]\n" - "add x20, x20, #0x10\n" "fmla v27.4s, v10.4s, v3.s[2]\n" - "prfm pldl1keep, [x20, #0x80]\n" "fmla v28.4s, v10.4s, v4.s[2]\n" - "add x16, x16, #0x40\n" "fmla v29.4s, v10.4s, v5.s[2]\n" "fmla v30.4s, v10.4s, v6.s[2]\n" "fmla v31.4s, v10.4s, v7.s[2]\n" @@ -2082,54 +2081,52 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "fmla v30.4s, v11.4s, v6.s[3]\n" "fmla v31.4s, v11.4s, v7.s[3]\n" "161:" // Height 8: Multiply loop: Main loop skip - "cbz x12, 163f\n" + "cbz x8, 163f\n" "162:" // Height 8: Multiply loop: Odd block loop - "ldr s0, [x11], #0x4\n" - "sub x12, x12, #0x1\n" - "ldr s1, [x9], #0x4\n" - "ldr s2, [x27], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr s5, [x23], #0x4\n" - "ldr s6, [x22], #0x4\n" - "ldr s7, [x20], #0x4\n" - "ldr q12, [x16, #0x0]\n" - "add x16, x16, #0x10\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" - "fmla v28.4s, v12.4s, v4.s[0]\n" - "fmla v29.4s, v12.4s, v5.s[0]\n" - "fmla v30.4s, v12.4s, v6.s[0]\n" - "fmla v31.4s, v12.4s, v7.s[0]\n" - "cbnz x12, 162b\n" + "ldr s0, [x17], #0x4\n" + "sub x8, x8, #0x1\n" + "ldr s23, [x16], #0x4\n" + "ldr s22, [x15], #0x4\n" + "ldr s21, [x14], #0x4\n" + "ldr s20, [x13], #0x4\n" + "ldr s19, [x12], #0x4\n" + "ldr s18, [x11], #0x4\n" + "ldr s17, [x27], #0x4\n" + "ldr q16, [x5, #0x0]\n" + "fmla v24.4s, v16.4s, v0.s[0]\n" + "fmla v25.4s, v16.4s, v23.s[0]\n" + "add x5, x5, #0x10\n" + "fmla v26.4s, v16.4s, v22.s[0]\n" + "fmla v27.4s, v16.4s, v21.s[0]\n" + "fmla v28.4s, v16.4s, v20.s[0]\n" + "fmla v29.4s, v16.4s, v19.s[0]\n" + "fmla v30.4s, v16.4s, v18.s[0]\n" + "fmla v31.4s, v16.4s, v17.s[0]\n" + "cbnz x8, 162b\n" "163:" // Height 8: Multiply loop: No odd multiplies - "ldr w8, [%x[args_ptr], %[offsetof_num_strings]]\n" - "add x13, x13, #0x1\n" - "cmp x13, x8\n" + "ldr w26, [%x[args_ptr], %[offsetof_num_strings]]\n" + "add x7, x7, #0x1\n" + "cmp x7, x26\n" "bne 156b\n" - "ldr x8, [%x[args_ptr], %[offsetof_output_offset]]\n" - "prfm pstl1keep, [x14, #0x0]\n" - "add x27, x14, x8, LSL #2\n" + "ldr x26, [%x[args_ptr], %[offsetof_output_offset]]\n" + "add x13, x6, x26, LSL #2\n" + "add x12, x13, x26, LSL #2\n" + "add x11, x12, x26, LSL #2\n" + "add x10, x11, x26, LSL #2\n" + "add x9, x10, x26, LSL #2\n" + "add x28, x9, x26, LSL #2\n" + "add x27, x28, x26, LSL #2\n" + "prfm pstl1keep, [x6, #0x0]\n" + "prfm pstl1keep, [x13, #0x0]\n" + "prfm pstl1keep, [x12, #0x0]\n" + "prfm pstl1keep, [x11, #0x0]\n" + "prfm pstl1keep, [x10, #0x0]\n" + "prfm pstl1keep, [x9, #0x0]\n" + "prfm pstl1keep, [x28, #0x0]\n" "prfm pstl1keep, [x27, #0x0]\n" - "add x26, x27, x8, LSL #2\n" - "prfm pstl1keep, [x26, #0x0]\n" - "add x25, x26, x8, LSL #2\n" - "prfm pstl1keep, [x25, #0x0]\n" - "add x24, x25, x8, LSL #2\n" - "prfm pstl1keep, [x24, #0x0]\n" - "add x23, x24, x8, LSL #2\n" - "prfm pstl1keep, [x23, #0x0]\n" - "add x22, x23, x8, LSL #2\n" - "prfm pstl1keep, [x22, #0x0]\n" - "add x21, x22, x8, LSL #2\n" - "prfm pstl1keep, [x21, #0x0]\n" "tbz %x[flags], #1, 164f\n" - "add x20, %x[args_ptr], %[offset_min]\n" - "add x8, %x[args_ptr], %[offset_max]\n" - "ld1r { v17.4s }, [x20]\n" - "ld1r { v16.4s }, [x8]\n" + "add x26, %x[args_ptr], %[offset_max]\n" + "ld1r { v16.4s }, [x26]\n" "fmin v24.4s, v24.4s, v16.4s\n" "fmin v25.4s, v25.4s, v16.4s\n" "fmin v26.4s, v26.4s, v16.4s\n" @@ -2138,76 +2135,77 @@ void a64_hybrid_fp32_mla_8x4_a55 ( "fmin v29.4s, v29.4s, v16.4s\n" "fmin v30.4s, v30.4s, v16.4s\n" "fmin v31.4s, v31.4s, v16.4s\n" - "fmax v24.4s, v24.4s, v17.4s\n" - "fmax v25.4s, v25.4s, v17.4s\n" - "fmax v26.4s, v26.4s, v17.4s\n" - "fmax v27.4s, v27.4s, v17.4s\n" - "fmax v28.4s, v28.4s, v17.4s\n" - "fmax v29.4s, v29.4s, v17.4s\n" - "fmax v30.4s, v30.4s, v17.4s\n" - "fmax v31.4s, v31.4s, v17.4s\n" + "add x26, %x[args_ptr], %[offset_min]\n" + "ld1r { v16.4s }, [x26]\n" + "fmax v24.4s, v24.4s, v16.4s\n" + "fmax v25.4s, v25.4s, v16.4s\n" + "fmax v26.4s, v26.4s, v16.4s\n" + "fmax v27.4s, v27.4s, v16.4s\n" + "fmax v28.4s, v28.4s, v16.4s\n" + "fmax v29.4s, v29.4s, v16.4s\n" + "fmax v30.4s, v30.4s, v16.4s\n" + "fmax v31.4s, v31.4s, v16.4s\n" "164:" // Height 8: No activation - "cmp x17, #0x4\n" + "cmp x4, #0x4\n" "bge 167f\n" - "tbz x17, #1, 165f\n" - "str d24, [x14], #0x8\n" - "str d25, [x27], #0x8\n" - "str d26, [x26], #0x8\n" - "str d27, [x25], #0x8\n" - "str d28, [x24], #0x8\n" - "str d29, [x23], #0x8\n" - "str d30, [x22], #0x8\n" - "str d31, [x21], #0x8\n" - "tbz x17, #0, 166f\n" - "st1 { v24.s }[2], [x14]\n" - "st1 { v25.s }[2], [x27]\n" - "st1 { v26.s }[2], [x26]\n" - "st1 { v27.s }[2], [x25]\n" - "st1 { v28.s }[2], [x24]\n" - "st1 { v29.s }[2], [x23]\n" - "st1 { v30.s }[2], [x22]\n" - "st1 { v31.s }[2], [x21]\n" + "tbz x4, #1, 165f\n" + "str d24, [x6], #0x8\n" + "str d25, [x13], #0x8\n" + "str d26, [x12], #0x8\n" + "str d27, [x11], #0x8\n" + "str d28, [x10], #0x8\n" + "str d29, [x9], #0x8\n" + "str d30, [x28], #0x8\n" + "str d31, [x27], #0x8\n" + "tbz x4, #0, 166f\n" + "st1 { v24.s }[2], [x6]\n" + "st1 { v25.s }[2], [x13]\n" + "st1 { v26.s }[2], [x12]\n" + "st1 { v27.s }[2], [x11]\n" + "st1 { v28.s }[2], [x10]\n" + "st1 { v29.s }[2], [x9]\n" + "st1 { v30.s }[2], [x28]\n" + "st1 { v31.s }[2], [x27]\n" "b 166f\n" "165:" // Height 8: Partial direct writeback: partial_1_0 - "str s24, [x14, #0x0]\n" - "str s25, [x27, #0x0]\n" - "str s26, [x26, #0x0]\n" - "str s27, [x25, #0x0]\n" - "str s28, [x24, #0x0]\n" - "str s29, [x23, #0x0]\n" - "str s30, [x22, #0x0]\n" - "str s31, [x21, #0x0]\n" + "str s24, [x6, #0x0]\n" + "str s25, [x13, #0x0]\n" + "str s26, [x12, #0x0]\n" + "str s27, [x11, #0x0]\n" + "str s28, [x10, #0x0]\n" + "str s29, [x9, #0x0]\n" + "str s30, [x28, #0x0]\n" + "str s31, [x27, #0x0]\n" "166:" // Height 8: Partial direct writeback: Done "b 168f\n" "167:" // Height 8: Full writeback - "str q24, [x14, #0x0]\n" - "add x14, x14, #0x10\n" - "str q25, [x27, #0x0]\n" - "str q26, [x26, #0x0]\n" - "str q27, [x25, #0x0]\n" - "str q28, [x24, #0x0]\n" - "str q29, [x23, #0x0]\n" - "str q30, [x22, #0x0]\n" - "str q31, [x21, #0x0]\n" + "str q24, [x6, #0x0]\n" + "add x6, x6, #0x10\n" + "str q25, [x13, #0x0]\n" + "str q26, [x12, #0x0]\n" + "str q27, [x11, #0x0]\n" + "str q28, [x10, #0x0]\n" + "str q29, [x9, #0x0]\n" + "str q30, [x28, #0x0]\n" + "str q31, [x27, #0x0]\n" "168:" // Height 8: Writeback done - "subs x17, x17, #0x4\n" + "subs x4, x4, #0x4\n" "bgt 149b\n" "subs %x[M], %x[M], #0x8\n" "beq 170f\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x27, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 169f\n" - "add x20, x20, #0x8\n" - "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "add x27, x27, #0x8\n" + "str x27, [%x[args_ptr], %[offsetof_input_offset]]\n" "b 1b\n" "169:" // Update direct input - "mov x8, #0x20\n" - "madd %x[input_ptr], x8, x20, %x[input_ptr]\n" + "mov x26, #0x20\n" + "madd %x[input_ptr], x26, x27, %x[input_ptr]\n" "b 1b\n" "170:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, 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)) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v16", "v17", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x8", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x3", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x26", "x27", "x28" ); } diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp index bd22336c8d..004e5d7f23 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_8x4/generic.cpp @@ -92,7 +92,6 @@ void a64_hybrid_fp32_mla_8x4 ( break; } __asm__ __volatile__( - "1:" // Row loop "cmp %x[M], #0x8\n" "bge 148f\n" @@ -140,11 +139,11 @@ void a64_hybrid_fp32_mla_8x4 ( "9:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 10f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" "cbnz x10, 11f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -189,10 +188,10 @@ void a64_hybrid_fp32_mla_8x4 ( "14:" // Height 1: Multiply loop: Main loop skip "cbz x9, 16f\n" "15:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr q12, [x12, #0x0]\n" + "ldr s17, [x28], #0x4\n" + "ldr q16, [x12, #0x0]\n" "sub x9, x9, #0x1\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" + "fmla v24.4s, v16.4s, v17.s[0]\n" "add x12, x12, #0x10\n" "cbnz x9, 15b\n" "16:" // Height 1: Multiply loop: No odd multiplies @@ -271,12 +270,12 @@ void a64_hybrid_fp32_mla_8x4 ( "30:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 31f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" "cbnz x10, 32f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -284,7 +283,7 @@ void a64_hybrid_fp32_mla_8x4 ( "b 32f\n" "31:" // Height 2: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, x20, LSL #2\n" + "add x27, x28, x21, LSL #2\n" "32:" // Height 2: input setup done "cmp x9, #0x4\n" "blt 35f\n" @@ -337,12 +336,12 @@ void a64_hybrid_fp32_mla_8x4 ( "35:" // Height 2: Multiply loop: Main loop skip "cbz x9, 37f\n" "36:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s18, [x28], #0x4\n" + "ldr s17, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" + "ldr q16, [x12, #0x0]\n" + "fmla v24.4s, v16.4s, v18.s[0]\n" + "fmla v25.4s, v16.4s, v17.s[0]\n" "add x12, x12, #0x10\n" "cbnz x9, 36b\n" "37:" // Height 2: Multiply loop: No odd multiplies @@ -437,13 +436,13 @@ void a64_hybrid_fp32_mla_8x4 ( "51:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 52f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" "cbnz x10, 53f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -452,8 +451,8 @@ void a64_hybrid_fp32_mla_8x4 ( "b 53f\n" "52:" // Height 3: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, x20, LSL #2\n" - "add x26, x27, x20, LSL #2\n" + "add x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" "53:" // Height 3: input setup done "cmp x9, #0x4\n" "blt 56f\n" @@ -520,14 +519,14 @@ void a64_hybrid_fp32_mla_8x4 ( "56:" // Height 3: Multiply loop: Main loop skip "cbz x9, 58f\n" "57:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s19, [x28], #0x4\n" + "ldr s18, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" + "ldr s17, [x26], #0x4\n" + "ldr q16, [x12, #0x0]\n" + "fmla v24.4s, v16.4s, v19.s[0]\n" + "fmla v25.4s, v16.4s, v18.s[0]\n" + "fmla v26.4s, v16.4s, v17.s[0]\n" "add x12, x12, #0x10\n" "cbnz x9, 57b\n" "58:" // Height 3: Multiply loop: No odd multiplies @@ -637,14 +636,14 @@ void a64_hybrid_fp32_mla_8x4 ( "72:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 73f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x25, [x20, #0x18]\n" "cbnz x10, 74f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -654,9 +653,9 @@ void a64_hybrid_fp32_mla_8x4 ( "b 74f\n" "73:" // Height 4: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, x20, LSL #2\n" - "add x26, x27, x20, LSL #2\n" - "add x25, x26, x20, LSL #2\n" + "add x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" + "add x25, x26, x21, LSL #2\n" "74:" // Height 4: input setup done "cmp x9, #0x4\n" "blt 77f\n" @@ -737,17 +736,17 @@ void a64_hybrid_fp32_mla_8x4 ( "77:" // Height 4: Multiply loop: Main loop skip "cbz x9, 79f\n" "78:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s20, [x28], #0x4\n" + "ldr s19, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" + "ldr s18, [x26], #0x4\n" + "ldr s17, [x25], #0x4\n" + "ldr q16, [x12, #0x0]\n" + "fmla v24.4s, v16.4s, v20.s[0]\n" + "fmla v25.4s, v16.4s, v19.s[0]\n" "add x12, x12, #0x10\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" + "fmla v26.4s, v16.4s, v18.s[0]\n" + "fmla v27.4s, v16.4s, v17.s[0]\n" "cbnz x9, 78b\n" "79:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -871,15 +870,15 @@ void a64_hybrid_fp32_mla_8x4 ( "93:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 94f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" - "ldr x24, [x21, #0x20]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x25, [x20, #0x18]\n" + "ldr x24, [x20, #0x20]\n" "cbnz x10, 95f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -890,10 +889,10 @@ void a64_hybrid_fp32_mla_8x4 ( "b 95f\n" "94:" // Height 5: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, x20, LSL #2\n" - "add x26, x27, x20, LSL #2\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" "95:" // Height 5: input setup done "cmp x9, #0x4\n" "blt 98f\n" @@ -988,19 +987,19 @@ void a64_hybrid_fp32_mla_8x4 ( "98:" // Height 5: Multiply loop: Main loop skip "cbz x9, 100f\n" "99:" // Height 5: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s21, [x28], #0x4\n" + "ldr s20, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" + "ldr s19, [x26], #0x4\n" + "ldr s18, [x25], #0x4\n" + "ldr s17, [x24], #0x4\n" + "ldr q16, [x12, #0x0]\n" + "fmla v24.4s, v16.4s, v21.s[0]\n" + "fmla v25.4s, v16.4s, v20.s[0]\n" + "fmla v26.4s, v16.4s, v19.s[0]\n" + "fmla v27.4s, v16.4s, v18.s[0]\n" "add x12, x12, #0x10\n" - "fmla v28.4s, v12.4s, v4.s[0]\n" + "fmla v28.4s, v16.4s, v17.s[0]\n" "cbnz x9, 99b\n" "100:" // Height 5: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1139,16 +1138,16 @@ void a64_hybrid_fp32_mla_8x4 ( "114:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 115f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" - "ldr x24, [x21, #0x20]\n" - "ldr x23, [x21, #0x28]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x25, [x20, #0x18]\n" + "ldr x24, [x20, #0x20]\n" + "ldr x23, [x20, #0x28]\n" "cbnz x10, 116f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -1160,11 +1159,11 @@ void a64_hybrid_fp32_mla_8x4 ( "b 116f\n" "115:" // Height 6: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, x20, LSL #2\n" - "add x26, x27, x20, LSL #2\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" + "add x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "116:" // Height 6: input setup done "cmp x9, #0x4\n" "blt 119f\n" @@ -1273,21 +1272,21 @@ void a64_hybrid_fp32_mla_8x4 ( "119:" // Height 6: Multiply loop: Main loop skip "cbz x9, 121f\n" "120:" // Height 6: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s22, [x28], #0x4\n" + "ldr s21, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr s5, [x23], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" + "ldr s20, [x26], #0x4\n" + "ldr s19, [x25], #0x4\n" + "ldr s18, [x24], #0x4\n" + "ldr s17, [x23], #0x4\n" + "ldr q16, [x12, #0x0]\n" + "fmla v24.4s, v16.4s, v22.s[0]\n" + "fmla v25.4s, v16.4s, v21.s[0]\n" "add x12, x12, #0x10\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" - "fmla v28.4s, v12.4s, v4.s[0]\n" - "fmla v29.4s, v12.4s, v5.s[0]\n" + "fmla v26.4s, v16.4s, v20.s[0]\n" + "fmla v27.4s, v16.4s, v19.s[0]\n" + "fmla v28.4s, v16.4s, v18.s[0]\n" + "fmla v29.4s, v16.4s, v17.s[0]\n" "cbnz x9, 120b\n" "121:" // Height 6: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1441,17 +1440,17 @@ void a64_hybrid_fp32_mla_8x4 ( "135:" // Height 7: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 136f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" - "ldr x24, [x21, #0x20]\n" - "ldr x23, [x21, #0x28]\n" - "ldr x22, [x21, #0x30]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x25, [x20, #0x18]\n" + "ldr x24, [x20, #0x20]\n" + "ldr x23, [x20, #0x28]\n" + "ldr x22, [x20, #0x30]\n" "cbnz x10, 137f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -1464,12 +1463,12 @@ void a64_hybrid_fp32_mla_8x4 ( "b 137f\n" "136:" // Height 7: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, x20, LSL #2\n" - "add x26, x27, x20, LSL #2\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "137:" // Height 7: input setup done "cmp x9, #0x4\n" "blt 140f\n" @@ -1592,23 +1591,23 @@ void a64_hybrid_fp32_mla_8x4 ( "140:" // Height 7: Multiply loop: Main loop skip "cbz x9, 142f\n" "141:" // Height 7: Multiply loop: Odd block loop - "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s23, [x28], #0x4\n" + "ldr s22, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr s5, [x23], #0x4\n" - "ldr s6, [x22], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" + "ldr s21, [x26], #0x4\n" + "ldr s20, [x25], #0x4\n" + "ldr s19, [x24], #0x4\n" + "ldr s18, [x23], #0x4\n" + "ldr s17, [x22], #0x4\n" + "ldr q16, [x12, #0x0]\n" + "fmla v24.4s, v16.4s, v23.s[0]\n" + "fmla v25.4s, v16.4s, v22.s[0]\n" + "fmla v26.4s, v16.4s, v21.s[0]\n" + "fmla v27.4s, v16.4s, v20.s[0]\n" "add x12, x12, #0x10\n" - "fmla v28.4s, v12.4s, v4.s[0]\n" - "fmla v29.4s, v12.4s, v5.s[0]\n" - "fmla v30.4s, v12.4s, v6.s[0]\n" + "fmla v28.4s, v16.4s, v19.s[0]\n" + "fmla v29.4s, v16.4s, v18.s[0]\n" + "fmla v30.4s, v16.4s, v17.s[0]\n" "cbnz x9, 141b\n" "142:" // Height 7: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1780,18 +1779,18 @@ void a64_hybrid_fp32_mla_8x4 ( "156:" // Height 8: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w9, [x20, x10, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 157f\n" - "ldr x21, [%x[input_ptr], x10, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x28, [x21, #0x0]\n" - "ldr x27, [x21, #0x8]\n" - "ldr x26, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" - "ldr x24, [x21, #0x20]\n" - "ldr x23, [x21, #0x28]\n" - "ldr x22, [x21, #0x30]\n" - "ldr x21, [x21, #0x38]\n" + "ldr x20, [%x[input_ptr], x10, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x28, [x20, #0x0]\n" + "ldr x27, [x20, #0x8]\n" + "ldr x26, [x20, #0x10]\n" + "ldr x25, [x20, #0x18]\n" + "ldr x24, [x20, #0x20]\n" + "ldr x23, [x20, #0x28]\n" + "ldr x22, [x20, #0x30]\n" + "ldr x21, [x20, #0x38]\n" "cbnz x10, 158f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x28, x28, x20, LSL #2\n" @@ -1805,13 +1804,13 @@ void a64_hybrid_fp32_mla_8x4 ( "b 158f\n" "157:" // Height 8: setup direct input "mov x28, %x[input_ptr]\n" - "add x27, x28, x20, LSL #2\n" - "add x26, x27, x20, LSL #2\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" + "add x27, x28, x21, LSL #2\n" + "add x26, x27, x21, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "158:" // Height 8: input setup done "cmp x9, #0x4\n" "blt 161f\n" @@ -1949,24 +1948,24 @@ void a64_hybrid_fp32_mla_8x4 ( "cbz x9, 163f\n" "162:" // Height 8: Multiply loop: Odd block loop "ldr s0, [x28], #0x4\n" - "ldr s1, [x27], #0x4\n" + "ldr s23, [x27], #0x4\n" "sub x9, x9, #0x1\n" - "ldr s2, [x26], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x24], #0x4\n" - "ldr s5, [x23], #0x4\n" - "ldr s6, [x22], #0x4\n" - "ldr s7, [x21], #0x4\n" - "ldr q12, [x12, #0x0]\n" - "fmla v24.4s, v12.4s, v0.s[0]\n" - "fmla v25.4s, v12.4s, v1.s[0]\n" + "ldr s22, [x26], #0x4\n" + "ldr s21, [x25], #0x4\n" + "ldr s20, [x24], #0x4\n" + "ldr s19, [x23], #0x4\n" + "ldr s18, [x22], #0x4\n" + "ldr s17, [x21], #0x4\n" + "ldr q16, [x12, #0x0]\n" + "fmla v24.4s, v16.4s, v0.s[0]\n" + "fmla v25.4s, v16.4s, v23.s[0]\n" "add x12, x12, #0x10\n" - "fmla v26.4s, v12.4s, v2.s[0]\n" - "fmla v27.4s, v12.4s, v3.s[0]\n" - "fmla v28.4s, v12.4s, v4.s[0]\n" - "fmla v29.4s, v12.4s, v5.s[0]\n" - "fmla v30.4s, v12.4s, v6.s[0]\n" - "fmla v31.4s, v12.4s, v7.s[0]\n" + "fmla v26.4s, v16.4s, v22.s[0]\n" + "fmla v27.4s, v16.4s, v21.s[0]\n" + "fmla v28.4s, v16.4s, v20.s[0]\n" + "fmla v29.4s, v16.4s, v19.s[0]\n" + "fmla v30.4s, v16.4s, v18.s[0]\n" + "fmla v31.4s, v16.4s, v17.s[0]\n" "cbnz x9, 162b\n" "163:" // Height 8: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -2068,10 +2067,9 @@ void a64_hybrid_fp32_mla_8x4 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "170:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, 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)) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v16", "v17", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } -- cgit v1.2.1