diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp | 1479 |
1 files changed, 741 insertions, 738 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp index bb84a50282..e411da6874 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021, 2023 Arm Limited. + * Copyright (c) 2019-2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,18 +47,19 @@ void a64_hybrid_fp32_mla_6x16 ( size_t output_offset = {}; size_t input_initial_col = {}; size_t input_offset = {}; + void *output_ptr = {}; + const float *bias = {}; } ka; unsigned long flags=0; - void *output_ptr; void *input_ptr; if (output_arg.is_indirect) { - output_ptr=(void *)(output_arg.indirect.ptr); + ka.output_ptr=(void *)(output_arg.indirect.ptr); ka.output_offset=output_arg.indirect.offset; flags |= 0x4; } else { - output_ptr=(void *)(output_arg.direct.base); + ka.output_ptr=(void *)(output_arg.direct.base); ka.output_offset=output_arg.direct.stride; } @@ -79,6 +80,7 @@ void a64_hybrid_fp32_mla_6x16 ( ka.string_lengths = string_lengths; ka.N = N; ka.B_ptr = B_ptr; + ka.bias = bias; switch(act.type) { default: case Activation::Type::None: @@ -101,10 +103,10 @@ void a64_hybrid_fp32_mla_6x16 ( "cmp %x[M], #0x2\n" "bgt 67f\n" "beq 34f\n" - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "2:" // Height 1: Column loop "cbz x12, 3f\n" "ldr q8, [x12, #0x0]\n" @@ -187,8 +189,8 @@ void a64_hybrid_fp32_mla_6x16 ( "mov x28, #0x0\n" "15:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 16f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -212,6 +214,10 @@ void a64_hybrid_fp32_mla_6x16 ( "ldr q17, [x10, #0x20]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "ldr q16, [x10, #0x30]\n" + "sub x27, x27, #0x4\n" + "add x26, x26, #0x10\n" + "cmp x27, #0x8\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla v10.4s, v17.4s, v0.s[0]\n" "ldr q17, [x10, #0x40]\n" "fmla v11.4s, v16.4s, v0.s[0]\n" @@ -236,22 +242,21 @@ void a64_hybrid_fp32_mla_6x16 ( "ldr q17, [x10, #0xe0]\n" "fmla v9.4s, v16.4s, v0.s[3]\n" "ldr q16, [x10, #0xf0]\n" - "sub x27, x27, #0x4\n" - "add x26, x26, #0x10\n" + "add x10, x10, #0x100\n" "fmla v10.4s, v17.4s, v0.s[3]\n" + "ldr q6, [x10, #0x0]\n" "fmla v11.4s, v16.4s, v0.s[3]\n" "ldr q0, [x26, #0x0]\n" - "cmp x27, #0x8\n" - "add x10, x10, #0x100\n" - "ldr q6, [x10, #0x0]\n" "ldr q7, [x10, #0x10]\n" - "prfm pldl1keep, [x26, #0x80]\n" "bge 18b\n" "19:" // Height 1: Multiply loop: Single iteration only "fmla v8.4s, v6.4s, v0.s[0]\n" "ldr q17, [x10, #0x20]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "ldr q16, [x10, #0x30]\n" + "add x26, x26, #0x10\n" + "sub x27, x27, #0x4\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla v10.4s, v17.4s, v0.s[0]\n" "ldr q17, [x10, #0x40]\n" "fmla v11.4s, v16.4s, v0.s[0]\n" @@ -276,26 +281,23 @@ void a64_hybrid_fp32_mla_6x16 ( "ldr q17, [x10, #0xe0]\n" "fmla v9.4s, v16.4s, v0.s[3]\n" "ldr q16, [x10, #0xf0]\n" - "add x26, x26, #0x10\n" - "sub x27, x27, #0x4\n" + "add x10, x10, #0x100\n" "fmla v10.4s, v17.4s, v0.s[3]\n" "fmla v11.4s, v16.4s, v0.s[3]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x10, x10, #0x100\n" "20:" // Height 1: Multiply loop: Main loop skip "cbz x27, 22f\n" "21:" // Height 1: Multiply loop: Odd block loop "ldr s18, [x26], #0x4\n" - "ldr q16, [x10, #0x0]\n" - "fmla v8.4s, v16.4s, v18.s[0]\n" + "ldr q17, [x10, #0x0]\n" "sub x27, x27, #0x1\n" - "ldr q17, [x10, #0x10]\n" - "ldr q16, [x10, #0x20]\n" - "fmla v9.4s, v17.4s, v18.s[0]\n" - "fmla v10.4s, v16.4s, v18.s[0]\n" + "ldr q16, [x10, #0x10]\n" + "fmla v8.4s, v17.4s, v18.s[0]\n" + "ldr q17, [x10, #0x20]\n" + "fmla v9.4s, v16.4s, v18.s[0]\n" "ldr q16, [x10, #0x30]\n" - "fmla v11.4s, v16.4s, v18.s[0]\n" "add x10, x10, #0x40\n" + "fmla v10.4s, v17.4s, v18.s[0]\n" + "fmla v11.4s, v16.4s, v18.s[0]\n" "cbnz x27, 21b\n" "22:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -304,9 +306,9 @@ void a64_hybrid_fp32_mla_6x16 ( "bne 15b\n" "prfm pstl1keep, [x9, #0x0]\n" "tbz %x[flags], #1, 23f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v17.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v8.4s, v8.4s, v17.4s\n" "fmin v9.4s, v9.4s, v17.4s\n" @@ -376,95 +378,95 @@ void a64_hybrid_fp32_mla_6x16 ( "bgt 2b\n" "b 200f\n" "34:" // Height 2 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "35:" // Height 2: Column loop "cbz x12, 36f\n" "ldr q8, [x12, #0x0]\n" "ldr q9, [x12, #0x10]\n" - "mov v12.16b, v8.16b\n" - "mov v13.16b, v9.16b\n" "ldr q10, [x12, #0x20]\n" "ldr q11, [x12, #0x30]\n" + "add x12, x12, #0x40\n" + "mov v12.16b, v8.16b\n" + "mov v13.16b, v9.16b\n" "mov v14.16b, v10.16b\n" "mov v15.16b, v11.16b\n" - "add x12, x12, #0x40\n" "b 47f\n" "36:" // Height 2: no bias "tbz %x[flags], #0, 46f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "cmp x11, #0x10\n" - "add x25, x9, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" "bge 45f\n" "tbz x11, #3, 40f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" "ld1 { v9.4s }, [x9], #0x10\n" - "ld1 { v13.4s }, [x25], #0x10\n" + "ld1 { v13.4s }, [x26], #0x10\n" "tbz x11, #2, 38f\n" "ld1 { v10.4s }, [x9], #0x10\n" - "ld1 { v14.4s }, [x25], #0x10\n" + "ld1 { v14.4s }, [x26], #0x10\n" "tbz x11, #1, 37f\n" "ldr d11, [x9], #0x8\n" - "ldr d15, [x25], #0x8\n" + "ldr d15, [x26], #0x8\n" "mov x20, #0x38\n" "tbz x11, #0, 44f\n" "ld1 { v11.s }[2], [x9]\n" - "ld1 { v15.s }[2], [x25]\n" + "ld1 { v15.s }[2], [x26]\n" "b 44f\n" "37:" // Height 2: Partial accumulate: partial_1_12 "mov x20, #0x30\n" "tbz x11, #0, 44f\n" "ldr s11, [x9, #0x0]\n" - "ldr s15, [x25, #0x0]\n" + "ldr s15, [x26, #0x0]\n" "b 44f\n" "38:" // Height 2: Partial accumulate: partial_2_8 "tbz x11, #1, 39f\n" "ldr d10, [x9], #0x8\n" - "ldr d14, [x25], #0x8\n" + "ldr d14, [x26], #0x8\n" "mov x20, #0x28\n" "tbz x11, #0, 44f\n" "ld1 { v10.s }[2], [x9]\n" - "ld1 { v14.s }[2], [x25]\n" + "ld1 { v14.s }[2], [x26]\n" "b 44f\n" "39:" // Height 2: Partial accumulate: partial_1_8 "mov x20, #0x20\n" "tbz x11, #0, 44f\n" "ldr s10, [x9, #0x0]\n" - "ldr s14, [x25, #0x0]\n" + "ldr s14, [x26, #0x0]\n" "b 44f\n" "40:" // Height 2: Partial accumulate: partial_4_0 "tbz x11, #2, 42f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" "tbz x11, #1, 41f\n" "ldr d9, [x9], #0x8\n" - "ldr d13, [x25], #0x8\n" + "ldr d13, [x26], #0x8\n" "mov x20, #0x18\n" "tbz x11, #0, 44f\n" "ld1 { v9.s }[2], [x9]\n" - "ld1 { v13.s }[2], [x25]\n" + "ld1 { v13.s }[2], [x26]\n" "b 44f\n" "41:" // Height 2: Partial accumulate: partial_1_4 "mov x20, #0x10\n" "tbz x11, #0, 44f\n" "ldr s9, [x9, #0x0]\n" - "ldr s13, [x25, #0x0]\n" + "ldr s13, [x26, #0x0]\n" "b 44f\n" "42:" // Height 2: Partial accumulate: partial_2_0 "tbz x11, #1, 43f\n" "ldr d8, [x9], #0x8\n" - "ldr d12, [x25], #0x8\n" + "ldr d12, [x26], #0x8\n" "mov x20, #0x8\n" "tbz x11, #0, 44f\n" "ld1 { v8.s }[2], [x9]\n" - "ld1 { v12.s }[2], [x25]\n" + "ld1 { v12.s }[2], [x26]\n" "b 44f\n" "43:" // Height 2: Partial accumulate: partial_1_0 "ldr s8, [x9, #0x0]\n" - "ldr s12, [x25, #0x0]\n" + "ldr s12, [x26, #0x0]\n" "mov x20, #0x0\n" "44:" // Height 2: Partial accumulate: Done "sub x9, x9, x20\n" @@ -474,10 +476,10 @@ void a64_hybrid_fp32_mla_6x16 ( "ldr q9, [x9, #0x10]\n" "ldr q10, [x9, #0x20]\n" "ldr q11, [x9, #0x30]\n" - "ldr q12, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "ldr q14, [x25, #0x20]\n" - "ldr q15, [x25, #0x30]\n" + "ldr q12, [x26, #0x0]\n" + "ldr q13, [x26, #0x10]\n" + "ldr q14, [x26, #0x20]\n" + "ldr q15, [x26, #0x30]\n" "b 47f\n" "46:" // Height 2: no accumulate "movi v8.16b, #0x0\n" @@ -492,8 +494,8 @@ void a64_hybrid_fp32_mla_6x16 ( "mov x28, #0x0\n" "48:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 49f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -525,22 +527,22 @@ void a64_hybrid_fp32_mla_6x16 ( "fmla v13.4s, v7.4s, v1.s[0]\n" "ldr q16, [x10, #0x30]\n" "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" + "cmp x27, #0x8\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla v10.4s, v17.4s, v0.s[0]\n" "fmla v14.4s, v17.4s, v1.s[0]\n" "ldr q17, [x10, #0x40]\n" - "add x25, x25, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla v11.4s, v16.4s, v0.s[0]\n" "fmla v15.4s, v16.4s, v1.s[0]\n" "ldr q16, [x10, #0x50]\n" - "cmp x27, #0x8\n" "fmla v8.4s, v17.4s, v0.s[1]\n" "fmla v12.4s, v17.4s, v1.s[1]\n" "ldr q17, [x10, #0x60]\n" - "prfm pldl1keep, [x26, #0x80]\n" "fmla v9.4s, v16.4s, v0.s[1]\n" "fmla v13.4s, v16.4s, v1.s[1]\n" "ldr q16, [x10, #0x70]\n" - "prfm pldl1keep, [x25, #0x80]\n" "fmla v10.4s, v17.4s, v0.s[1]\n" "fmla v14.4s, v17.4s, v1.s[1]\n" "ldr q17, [x10, #0x80]\n" @@ -584,18 +586,18 @@ void a64_hybrid_fp32_mla_6x16 ( "fmla v13.4s, v7.4s, v1.s[0]\n" "ldr q16, [x10, #0x30]\n" "add x25, x25, #0x10\n" + "sub x27, x27, #0x4\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla v10.4s, v17.4s, v0.s[0]\n" "fmla v14.4s, v17.4s, v1.s[0]\n" "ldr q17, [x10, #0x40]\n" - "sub x27, x27, #0x4\n" "fmla v11.4s, v16.4s, v0.s[0]\n" "fmla v15.4s, v16.4s, v1.s[0]\n" "ldr q16, [x10, #0x50]\n" - "prfm pldl1keep, [x26, #0x80]\n" "fmla v8.4s, v17.4s, v0.s[1]\n" "fmla v12.4s, v17.4s, v1.s[1]\n" "ldr q17, [x10, #0x60]\n" - "prfm pldl1keep, [x25, #0x80]\n" "fmla v9.4s, v16.4s, v0.s[1]\n" "fmla v13.4s, v16.4s, v1.s[1]\n" "ldr q16, [x10, #0x70]\n" @@ -642,9 +644,9 @@ void a64_hybrid_fp32_mla_6x16 ( "fmla v9.4s, v16.4s, v19.s[0]\n" "fmla v13.4s, v16.4s, v18.s[0]\n" "ldr q16, [x10, #0x30]\n" + "add x10, x10, #0x40\n" "fmla v10.4s, v17.4s, v19.s[0]\n" "fmla v14.4s, v17.4s, v18.s[0]\n" - "add x10, x10, #0x40\n" "fmla v11.4s, v16.4s, v19.s[0]\n" "fmla v15.4s, v16.4s, v18.s[0]\n" "cbnz x27, 54b\n" @@ -654,13 +656,13 @@ void a64_hybrid_fp32_mla_6x16 ( "cmp x28, x20\n" "bne 48b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" "prfm pstl1keep, [x9, #0x0]\n" - "prfm pstl1keep, [x25, #0x0]\n" + "add x26, x9, x20, LSL #2\n" + "prfm pstl1keep, [x26, #0x0]\n" "tbz %x[flags], #1, 56f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v17.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v8.4s, v8.4s, v17.4s\n" "fmin v9.4s, v9.4s, v17.4s\n" @@ -684,63 +686,63 @@ void a64_hybrid_fp32_mla_6x16 ( "tbz x11, #3, 60f\n" "st1 { v8.4s }, [x9], #0x10\n" "st1 { v9.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v13.4s }, [x25], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v13.4s }, [x26], #0x10\n" "tbz x11, #2, 58f\n" "st1 { v10.4s }, [x9], #0x10\n" - "st1 { v14.4s }, [x25], #0x10\n" + "st1 { v14.4s }, [x26], #0x10\n" "tbz x11, #1, 57f\n" "str d11, [x9], #0x8\n" - "str d15, [x25], #0x8\n" + "str d15, [x26], #0x8\n" "tbz x11, #0, 64f\n" "st1 { v11.s }[2], [x9]\n" - "st1 { v15.s }[2], [x25]\n" + "st1 { v15.s }[2], [x26]\n" "b 64f\n" "57:" // Height 2: Partial direct writeback: partial_1_12 "tbz x11, #0, 64f\n" "str s11, [x9, #0x0]\n" - "str s15, [x25, #0x0]\n" + "str s15, [x26, #0x0]\n" "b 64f\n" "58:" // Height 2: Partial direct writeback: partial_2_8 "tbz x11, #1, 59f\n" "str d10, [x9], #0x8\n" - "str d14, [x25], #0x8\n" + "str d14, [x26], #0x8\n" "tbz x11, #0, 64f\n" "st1 { v10.s }[2], [x9]\n" - "st1 { v14.s }[2], [x25]\n" + "st1 { v14.s }[2], [x26]\n" "b 64f\n" "59:" // Height 2: Partial direct writeback: partial_1_8 "tbz x11, #0, 64f\n" "str s10, [x9, #0x0]\n" - "str s14, [x25, #0x0]\n" + "str s14, [x26, #0x0]\n" "b 64f\n" "60:" // Height 2: Partial direct writeback: partial_4_0 "tbz x11, #2, 62f\n" "st1 { v8.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" "tbz x11, #1, 61f\n" "str d9, [x9], #0x8\n" - "str d13, [x25], #0x8\n" + "str d13, [x26], #0x8\n" "tbz x11, #0, 64f\n" "st1 { v9.s }[2], [x9]\n" - "st1 { v13.s }[2], [x25]\n" + "st1 { v13.s }[2], [x26]\n" "b 64f\n" "61:" // Height 2: Partial direct writeback: partial_1_4 "tbz x11, #0, 64f\n" "str s9, [x9, #0x0]\n" - "str s13, [x25, #0x0]\n" + "str s13, [x26, #0x0]\n" "b 64f\n" "62:" // Height 2: Partial direct writeback: partial_2_0 "tbz x11, #1, 63f\n" "str d8, [x9], #0x8\n" - "str d12, [x25], #0x8\n" + "str d12, [x26], #0x8\n" "tbz x11, #0, 64f\n" "st1 { v8.s }[2], [x9]\n" - "st1 { v12.s }[2], [x25]\n" + "st1 { v12.s }[2], [x26]\n" "b 64f\n" "63:" // Height 2: Partial direct writeback: partial_1_0 "str s8, [x9, #0x0]\n" - "str s12, [x25, #0x0]\n" + "str s12, [x26, #0x0]\n" "64:" // Height 2: Partial direct writeback: Done "b 66f\n" "65:" // Height 2: Full writeback @@ -749,126 +751,126 @@ void a64_hybrid_fp32_mla_6x16 ( "str q10, [x9, #0x20]\n" "str q11, [x9, #0x30]\n" "add x9, x9, #0x40\n" - "str q12, [x25, #0x0]\n" - "str q13, [x25, #0x10]\n" - "str q14, [x25, #0x20]\n" - "str q15, [x25, #0x30]\n" + "str q12, [x26, #0x0]\n" + "str q13, [x26, #0x10]\n" + "str q14, [x26, #0x20]\n" + "str q15, [x26, #0x30]\n" "66:" // Height 2: Writeback done "subs x11, x11, #0x10\n" "bgt 35b\n" "b 200f\n" "67:" // Height 3 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "68:" // Height 3: Column loop "cbz x12, 69f\n" "ldr q8, [x12, #0x0]\n" "ldr q9, [x12, #0x10]\n" - "mov v12.16b, v8.16b\n" - "mov v13.16b, v9.16b\n" "ldr q10, [x12, #0x20]\n" "ldr q11, [x12, #0x30]\n" - "mov v14.16b, v10.16b\n" - "mov v15.16b, v11.16b\n" + "add x12, x12, #0x40\n" + "mov v12.16b, v8.16b\n" + "mov v13.16b, v9.16b\n" "mov v16.16b, v8.16b\n" "mov v17.16b, v9.16b\n" - "add x12, x12, #0x40\n" + "mov v14.16b, v10.16b\n" + "mov v15.16b, v11.16b\n" "mov v18.16b, v10.16b\n" "mov v19.16b, v11.16b\n" "b 80f\n" "69:" // Height 3: no bias "tbz %x[flags], #0, 79f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" "cmp x11, #0x10\n" - "add x24, x25, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" "bge 78f\n" "tbz x11, #3, 73f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" "ld1 { v9.4s }, [x9], #0x10\n" - "ld1 { v13.4s }, [x25], #0x10\n" - "ld1 { v17.4s }, [x24], #0x10\n" + "ld1 { v13.4s }, [x26], #0x10\n" + "ld1 { v17.4s }, [x25], #0x10\n" "tbz x11, #2, 71f\n" "ld1 { v10.4s }, [x9], #0x10\n" - "ld1 { v14.4s }, [x25], #0x10\n" - "ld1 { v18.4s }, [x24], #0x10\n" + "ld1 { v14.4s }, [x26], #0x10\n" + "ld1 { v18.4s }, [x25], #0x10\n" "tbz x11, #1, 70f\n" "ldr d11, [x9], #0x8\n" - "ldr d15, [x25], #0x8\n" + "ldr d15, [x26], #0x8\n" "mov x20, #0x38\n" - "ldr d19, [x24], #0x8\n" + "ldr d19, [x25], #0x8\n" "tbz x11, #0, 77f\n" "ld1 { v11.s }[2], [x9]\n" - "ld1 { v15.s }[2], [x25]\n" - "ld1 { v19.s }[2], [x24]\n" + "ld1 { v15.s }[2], [x26]\n" + "ld1 { v19.s }[2], [x25]\n" "b 77f\n" "70:" // Height 3: Partial accumulate: partial_1_12 "mov x20, #0x30\n" "tbz x11, #0, 77f\n" "ldr s11, [x9, #0x0]\n" - "ldr s15, [x25, #0x0]\n" - "ldr s19, [x24, #0x0]\n" + "ldr s15, [x26, #0x0]\n" + "ldr s19, [x25, #0x0]\n" "b 77f\n" "71:" // Height 3: Partial accumulate: partial_2_8 "tbz x11, #1, 72f\n" "ldr d10, [x9], #0x8\n" - "ldr d14, [x25], #0x8\n" + "ldr d14, [x26], #0x8\n" "mov x20, #0x28\n" - "ldr d18, [x24], #0x8\n" + "ldr d18, [x25], #0x8\n" "tbz x11, #0, 77f\n" "ld1 { v10.s }[2], [x9]\n" - "ld1 { v14.s }[2], [x25]\n" - "ld1 { v18.s }[2], [x24]\n" + "ld1 { v14.s }[2], [x26]\n" + "ld1 { v18.s }[2], [x25]\n" "b 77f\n" "72:" // Height 3: Partial accumulate: partial_1_8 "mov x20, #0x20\n" "tbz x11, #0, 77f\n" "ldr s10, [x9, #0x0]\n" - "ldr s14, [x25, #0x0]\n" - "ldr s18, [x24, #0x0]\n" + "ldr s14, [x26, #0x0]\n" + "ldr s18, [x25, #0x0]\n" "b 77f\n" "73:" // Height 3: Partial accumulate: partial_4_0 "tbz x11, #2, 75f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" "tbz x11, #1, 74f\n" "ldr d9, [x9], #0x8\n" - "ldr d13, [x25], #0x8\n" + "ldr d13, [x26], #0x8\n" "mov x20, #0x18\n" - "ldr d17, [x24], #0x8\n" + "ldr d17, [x25], #0x8\n" "tbz x11, #0, 77f\n" "ld1 { v9.s }[2], [x9]\n" - "ld1 { v13.s }[2], [x25]\n" - "ld1 { v17.s }[2], [x24]\n" + "ld1 { v13.s }[2], [x26]\n" + "ld1 { v17.s }[2], [x25]\n" "b 77f\n" "74:" // Height 3: Partial accumulate: partial_1_4 "mov x20, #0x10\n" "tbz x11, #0, 77f\n" "ldr s9, [x9, #0x0]\n" - "ldr s13, [x25, #0x0]\n" - "ldr s17, [x24, #0x0]\n" + "ldr s13, [x26, #0x0]\n" + "ldr s17, [x25, #0x0]\n" "b 77f\n" "75:" // Height 3: Partial accumulate: partial_2_0 "tbz x11, #1, 76f\n" "ldr d8, [x9], #0x8\n" - "ldr d12, [x25], #0x8\n" + "ldr d12, [x26], #0x8\n" "mov x20, #0x8\n" - "ldr d16, [x24], #0x8\n" + "ldr d16, [x25], #0x8\n" "tbz x11, #0, 77f\n" "ld1 { v8.s }[2], [x9]\n" - "ld1 { v12.s }[2], [x25]\n" - "ld1 { v16.s }[2], [x24]\n" + "ld1 { v12.s }[2], [x26]\n" + "ld1 { v16.s }[2], [x25]\n" "b 77f\n" "76:" // Height 3: Partial accumulate: partial_1_0 "ldr s8, [x9, #0x0]\n" - "ldr s12, [x25, #0x0]\n" + "ldr s12, [x26, #0x0]\n" "mov x20, #0x0\n" - "ldr s16, [x24, #0x0]\n" + "ldr s16, [x25, #0x0]\n" "77:" // Height 3: Partial accumulate: Done "sub x9, x9, x20\n" "b 80f\n" @@ -877,14 +879,14 @@ void a64_hybrid_fp32_mla_6x16 ( "ldr q9, [x9, #0x10]\n" "ldr q10, [x9, #0x20]\n" "ldr q11, [x9, #0x30]\n" - "ldr q12, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "ldr q14, [x25, #0x20]\n" - "ldr q15, [x25, #0x30]\n" - "ldr q16, [x24, #0x0]\n" - "ldr q17, [x24, #0x10]\n" - "ldr q18, [x24, #0x20]\n" - "ldr q19, [x24, #0x30]\n" + "ldr q12, [x26, #0x0]\n" + "ldr q13, [x26, #0x10]\n" + "ldr q14, [x26, #0x20]\n" + "ldr q15, [x26, #0x30]\n" + "ldr q16, [x25, #0x0]\n" + "ldr q17, [x25, #0x10]\n" + "ldr q18, [x25, #0x20]\n" + "ldr q19, [x25, #0x30]\n" "b 80f\n" "79:" // Height 3: no accumulate "movi v8.16b, #0x0\n" @@ -903,8 +905,8 @@ void a64_hybrid_fp32_mla_6x16 ( "mov x28, #0x0\n" "81:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 82f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -944,18 +946,18 @@ void a64_hybrid_fp32_mla_6x16 ( "fmla v17.4s, v7.4s, v2.s[0]\n" "ldr q20, [x10, #0x30]\n" "add x24, x24, #0x10\n" - "fmla v10.4s, v21.4s, v0.s[0]\n" - "fmla v14.4s, v21.4s, v1.s[0]\n" "cmp x27, #0x8\n" "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "fmla v10.4s, v21.4s, v0.s[0]\n" + "fmla v14.4s, v21.4s, v1.s[0]\n" "fmla v18.4s, v21.4s, v2.s[0]\n" "ldr q21, [x10, #0x40]\n" + "prfm pldl1keep, [x24, #0x80]\n" "fmla v11.4s, v20.4s, v0.s[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" "fmla v15.4s, v20.4s, v1.s[0]\n" "fmla v19.4s, v20.4s, v2.s[0]\n" "ldr q20, [x10, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" "fmla v8.4s, v21.4s, v0.s[1]\n" "fmla v12.4s, v21.4s, v1.s[1]\n" "fmla v16.4s, v21.4s, v2.s[1]\n" @@ -1022,14 +1024,14 @@ void a64_hybrid_fp32_mla_6x16 ( "fmla v17.4s, v7.4s, v2.s[0]\n" "ldr q20, [x10, #0x30]\n" "sub x27, x27, #0x4\n" - "fmla v10.4s, v21.4s, v0.s[0]\n" - "fmla v14.4s, v21.4s, v1.s[0]\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" + "fmla v10.4s, v21.4s, v0.s[0]\n" + "fmla v14.4s, v21.4s, v1.s[0]\n" "fmla v18.4s, v21.4s, v2.s[0]\n" "ldr q21, [x10, #0x40]\n" - "fmla v11.4s, v20.4s, v0.s[0]\n" "prfm pldl1keep, [x24, #0x80]\n" + "fmla v11.4s, v20.4s, v0.s[0]\n" "fmla v15.4s, v20.4s, v1.s[0]\n" "fmla v19.4s, v20.4s, v2.s[0]\n" "ldr q20, [x10, #0x50]\n" @@ -1088,9 +1090,9 @@ void a64_hybrid_fp32_mla_6x16 ( "sub x27, x27, #0x1\n" "ldr s22, [x24], #0x4\n" "ldr q21, [x10, #0x0]\n" + "ldr q20, [x10, #0x10]\n" "fmla v8.4s, v21.4s, v24.s[0]\n" "fmla v12.4s, v21.4s, v23.s[0]\n" - "ldr q20, [x10, #0x10]\n" "fmla v16.4s, v21.4s, v22.s[0]\n" "ldr q21, [x10, #0x20]\n" "fmla v9.4s, v20.4s, v24.s[0]\n" @@ -1111,15 +1113,15 @@ void a64_hybrid_fp32_mla_6x16 ( "cmp x28, x20\n" "bne 81b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x9, #0x0]\n" + "add x26, x9, x20, LSL #2\n" + "prfm pstl1keep, [x26, #0x0]\n" + "add x25, x26, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" - "prfm pstl1keep, [x24, #0x0]\n" "tbz %x[flags], #1, 89f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v21.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v21.4s }, [x21]\n" "ld1r { v20.4s }, [x20]\n" "fmin v8.4s, v8.4s, v21.4s\n" "fmin v9.4s, v9.4s, v21.4s\n" @@ -1151,79 +1153,79 @@ void a64_hybrid_fp32_mla_6x16 ( "tbz x11, #3, 93f\n" "st1 { v8.4s }, [x9], #0x10\n" "st1 { v9.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v13.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v17.4s }, [x24], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v13.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v17.4s }, [x25], #0x10\n" "tbz x11, #2, 91f\n" "st1 { v10.4s }, [x9], #0x10\n" - "st1 { v14.4s }, [x25], #0x10\n" - "st1 { v18.4s }, [x24], #0x10\n" + "st1 { v14.4s }, [x26], #0x10\n" + "st1 { v18.4s }, [x25], #0x10\n" "tbz x11, #1, 90f\n" "str d11, [x9], #0x8\n" - "str d15, [x25], #0x8\n" - "str d19, [x24], #0x8\n" + "str d15, [x26], #0x8\n" + "str d19, [x25], #0x8\n" "tbz x11, #0, 97f\n" "st1 { v11.s }[2], [x9]\n" - "st1 { v15.s }[2], [x25]\n" - "st1 { v19.s }[2], [x24]\n" + "st1 { v15.s }[2], [x26]\n" + "st1 { v19.s }[2], [x25]\n" "b 97f\n" "90:" // Height 3: Partial direct writeback: partial_1_12 "tbz x11, #0, 97f\n" "str s11, [x9, #0x0]\n" - "str s15, [x25, #0x0]\n" - "str s19, [x24, #0x0]\n" + "str s15, [x26, #0x0]\n" + "str s19, [x25, #0x0]\n" "b 97f\n" "91:" // Height 3: Partial direct writeback: partial_2_8 "tbz x11, #1, 92f\n" "str d10, [x9], #0x8\n" - "str d14, [x25], #0x8\n" - "str d18, [x24], #0x8\n" + "str d14, [x26], #0x8\n" + "str d18, [x25], #0x8\n" "tbz x11, #0, 97f\n" "st1 { v10.s }[2], [x9]\n" - "st1 { v14.s }[2], [x25]\n" - "st1 { v18.s }[2], [x24]\n" + "st1 { v14.s }[2], [x26]\n" + "st1 { v18.s }[2], [x25]\n" "b 97f\n" "92:" // Height 3: Partial direct writeback: partial_1_8 "tbz x11, #0, 97f\n" "str s10, [x9, #0x0]\n" - "str s14, [x25, #0x0]\n" - "str s18, [x24, #0x0]\n" + "str s14, [x26, #0x0]\n" + "str s18, [x25, #0x0]\n" "b 97f\n" "93:" // Height 3: Partial direct writeback: partial_4_0 "tbz x11, #2, 95f\n" "st1 { v8.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" "tbz x11, #1, 94f\n" "str d9, [x9], #0x8\n" - "str d13, [x25], #0x8\n" - "str d17, [x24], #0x8\n" + "str d13, [x26], #0x8\n" + "str d17, [x25], #0x8\n" "tbz x11, #0, 97f\n" "st1 { v9.s }[2], [x9]\n" - "st1 { v13.s }[2], [x25]\n" - "st1 { v17.s }[2], [x24]\n" + "st1 { v13.s }[2], [x26]\n" + "st1 { v17.s }[2], [x25]\n" "b 97f\n" "94:" // Height 3: Partial direct writeback: partial_1_4 "tbz x11, #0, 97f\n" "str s9, [x9, #0x0]\n" - "str s13, [x25, #0x0]\n" - "str s17, [x24, #0x0]\n" + "str s13, [x26, #0x0]\n" + "str s17, [x25, #0x0]\n" "b 97f\n" "95:" // Height 3: Partial direct writeback: partial_2_0 "tbz x11, #1, 96f\n" "str d8, [x9], #0x8\n" - "str d12, [x25], #0x8\n" - "str d16, [x24], #0x8\n" + "str d12, [x26], #0x8\n" + "str d16, [x25], #0x8\n" "tbz x11, #0, 97f\n" "st1 { v8.s }[2], [x9]\n" - "st1 { v12.s }[2], [x25]\n" - "st1 { v16.s }[2], [x24]\n" + "st1 { v12.s }[2], [x26]\n" + "st1 { v16.s }[2], [x25]\n" "b 97f\n" "96:" // Height 3: Partial direct writeback: partial_1_0 "str s8, [x9, #0x0]\n" - "str s12, [x25, #0x0]\n" - "str s16, [x24, #0x0]\n" + "str s12, [x26, #0x0]\n" + "str s16, [x25, #0x0]\n" "97:" // Height 3: Partial direct writeback: Done "b 99f\n" "98:" // Height 3: Full writeback @@ -1232,39 +1234,39 @@ void a64_hybrid_fp32_mla_6x16 ( "str q10, [x9, #0x20]\n" "str q11, [x9, #0x30]\n" "add x9, x9, #0x40\n" - "str q12, [x25, #0x0]\n" - "str q13, [x25, #0x10]\n" - "str q14, [x25, #0x20]\n" - "str q15, [x25, #0x30]\n" - "str q16, [x24, #0x0]\n" - "str q17, [x24, #0x10]\n" - "str q18, [x24, #0x20]\n" - "str q19, [x24, #0x30]\n" + "str q12, [x26, #0x0]\n" + "str q13, [x26, #0x10]\n" + "str q14, [x26, #0x20]\n" + "str q15, [x26, #0x30]\n" + "str q16, [x25, #0x0]\n" + "str q17, [x25, #0x10]\n" + "str q18, [x25, #0x20]\n" + "str q19, [x25, #0x30]\n" "99:" // Height 3: Writeback done "subs x11, x11, #0x10\n" "bgt 68b\n" "b 200f\n" "100:" // Height 4 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "101:" // Height 4: Column loop "cbz x12, 102f\n" "ldr q8, [x12, #0x0]\n" "ldr q9, [x12, #0x10]\n" - "mov v12.16b, v8.16b\n" - "mov v13.16b, v9.16b\n" "ldr q10, [x12, #0x20]\n" "ldr q11, [x12, #0x30]\n" - "mov v14.16b, v10.16b\n" - "mov v15.16b, v11.16b\n" + "add x12, x12, #0x40\n" + "mov v12.16b, v8.16b\n" + "mov v13.16b, v9.16b\n" "mov v16.16b, v8.16b\n" "mov v17.16b, v9.16b\n" - "add x12, x12, #0x40\n" + "mov v20.16b, v8.16b\n" + "mov v14.16b, v10.16b\n" + "mov v15.16b, v11.16b\n" "mov v18.16b, v10.16b\n" "mov v19.16b, v11.16b\n" - "mov v20.16b, v8.16b\n" "mov v21.16b, v9.16b\n" "mov v22.16b, v10.16b\n" "mov v23.16b, v11.16b\n" @@ -1272,111 +1274,111 @@ void a64_hybrid_fp32_mla_6x16 ( "102:" // Height 4: no bias "tbz %x[flags], #0, 112f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "cmp x11, #0x10\n" - "add x23, x24, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" "bge 111f\n" "tbz x11, #3, 106f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" "ld1 { v9.4s }, [x9], #0x10\n" - "ld1 { v13.4s }, [x25], #0x10\n" - "ld1 { v17.4s }, [x24], #0x10\n" - "ld1 { v21.4s }, [x23], #0x10\n" + "ld1 { v13.4s }, [x26], #0x10\n" + "ld1 { v17.4s }, [x25], #0x10\n" + "ld1 { v21.4s }, [x24], #0x10\n" "tbz x11, #2, 104f\n" "ld1 { v10.4s }, [x9], #0x10\n" - "ld1 { v14.4s }, [x25], #0x10\n" - "ld1 { v18.4s }, [x24], #0x10\n" - "ld1 { v22.4s }, [x23], #0x10\n" + "ld1 { v14.4s }, [x26], #0x10\n" + "ld1 { v18.4s }, [x25], #0x10\n" + "ld1 { v22.4s }, [x24], #0x10\n" "tbz x11, #1, 103f\n" "ldr d11, [x9], #0x8\n" - "ldr d15, [x25], #0x8\n" + "ldr d15, [x26], #0x8\n" "mov x20, #0x38\n" - "ldr d19, [x24], #0x8\n" - "ldr d23, [x23], #0x8\n" + "ldr d19, [x25], #0x8\n" + "ldr d23, [x24], #0x8\n" "tbz x11, #0, 110f\n" "ld1 { v11.s }[2], [x9]\n" - "ld1 { v15.s }[2], [x25]\n" - "ld1 { v19.s }[2], [x24]\n" - "ld1 { v23.s }[2], [x23]\n" + "ld1 { v15.s }[2], [x26]\n" + "ld1 { v19.s }[2], [x25]\n" + "ld1 { v23.s }[2], [x24]\n" "b 110f\n" "103:" // Height 4: Partial accumulate: partial_1_12 "mov x20, #0x30\n" "tbz x11, #0, 110f\n" "ldr s11, [x9, #0x0]\n" - "ldr s15, [x25, #0x0]\n" - "ldr s19, [x24, #0x0]\n" - "ldr s23, [x23, #0x0]\n" + "ldr s15, [x26, #0x0]\n" + "ldr s19, [x25, #0x0]\n" + "ldr s23, [x24, #0x0]\n" "b 110f\n" "104:" // Height 4: Partial accumulate: partial_2_8 "tbz x11, #1, 105f\n" "ldr d10, [x9], #0x8\n" - "ldr d14, [x25], #0x8\n" + "ldr d14, [x26], #0x8\n" "mov x20, #0x28\n" - "ldr d18, [x24], #0x8\n" - "ldr d22, [x23], #0x8\n" + "ldr d18, [x25], #0x8\n" + "ldr d22, [x24], #0x8\n" "tbz x11, #0, 110f\n" "ld1 { v10.s }[2], [x9]\n" - "ld1 { v14.s }[2], [x25]\n" - "ld1 { v18.s }[2], [x24]\n" - "ld1 { v22.s }[2], [x23]\n" + "ld1 { v14.s }[2], [x26]\n" + "ld1 { v18.s }[2], [x25]\n" + "ld1 { v22.s }[2], [x24]\n" "b 110f\n" "105:" // Height 4: Partial accumulate: partial_1_8 "mov x20, #0x20\n" "tbz x11, #0, 110f\n" "ldr s10, [x9, #0x0]\n" - "ldr s14, [x25, #0x0]\n" - "ldr s18, [x24, #0x0]\n" - "ldr s22, [x23, #0x0]\n" + "ldr s14, [x26, #0x0]\n" + "ldr s18, [x25, #0x0]\n" + "ldr s22, [x24, #0x0]\n" "b 110f\n" "106:" // Height 4: Partial accumulate: partial_4_0 "tbz x11, #2, 108f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" "tbz x11, #1, 107f\n" "ldr d9, [x9], #0x8\n" - "ldr d13, [x25], #0x8\n" + "ldr d13, [x26], #0x8\n" "mov x20, #0x18\n" - "ldr d17, [x24], #0x8\n" - "ldr d21, [x23], #0x8\n" + "ldr d17, [x25], #0x8\n" + "ldr d21, [x24], #0x8\n" "tbz x11, #0, 110f\n" "ld1 { v9.s }[2], [x9]\n" - "ld1 { v13.s }[2], [x25]\n" - "ld1 { v17.s }[2], [x24]\n" - "ld1 { v21.s }[2], [x23]\n" + "ld1 { v13.s }[2], [x26]\n" + "ld1 { v17.s }[2], [x25]\n" + "ld1 { v21.s }[2], [x24]\n" "b 110f\n" "107:" // Height 4: Partial accumulate: partial_1_4 "mov x20, #0x10\n" "tbz x11, #0, 110f\n" "ldr s9, [x9, #0x0]\n" - "ldr s13, [x25, #0x0]\n" - "ldr s17, [x24, #0x0]\n" - "ldr s21, [x23, #0x0]\n" + "ldr s13, [x26, #0x0]\n" + "ldr s17, [x25, #0x0]\n" + "ldr s21, [x24, #0x0]\n" "b 110f\n" "108:" // Height 4: Partial accumulate: partial_2_0 "tbz x11, #1, 109f\n" "ldr d8, [x9], #0x8\n" - "ldr d12, [x25], #0x8\n" + "ldr d12, [x26], #0x8\n" "mov x20, #0x8\n" - "ldr d16, [x24], #0x8\n" - "ldr d20, [x23], #0x8\n" + "ldr d16, [x25], #0x8\n" + "ldr d20, [x24], #0x8\n" "tbz x11, #0, 110f\n" "ld1 { v8.s }[2], [x9]\n" - "ld1 { v12.s }[2], [x25]\n" - "ld1 { v16.s }[2], [x24]\n" - "ld1 { v20.s }[2], [x23]\n" + "ld1 { v12.s }[2], [x26]\n" + "ld1 { v16.s }[2], [x25]\n" + "ld1 { v20.s }[2], [x24]\n" "b 110f\n" "109:" // Height 4: Partial accumulate: partial_1_0 "ldr s8, [x9, #0x0]\n" - "ldr s12, [x25, #0x0]\n" + "ldr s12, [x26, #0x0]\n" "mov x20, #0x0\n" - "ldr s16, [x24, #0x0]\n" - "ldr s20, [x23, #0x0]\n" + "ldr s16, [x25, #0x0]\n" + "ldr s20, [x24, #0x0]\n" "110:" // Height 4: Partial accumulate: Done "sub x9, x9, x20\n" "b 113f\n" @@ -1385,18 +1387,18 @@ void a64_hybrid_fp32_mla_6x16 ( "ldr q9, [x9, #0x10]\n" "ldr q10, [x9, #0x20]\n" "ldr q11, [x9, #0x30]\n" - "ldr q12, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "ldr q14, [x25, #0x20]\n" - "ldr q15, [x25, #0x30]\n" - "ldr q16, [x24, #0x0]\n" - "ldr q17, [x24, #0x10]\n" - "ldr q18, [x24, #0x20]\n" - "ldr q19, [x24, #0x30]\n" - "ldr q20, [x23, #0x0]\n" - "ldr q21, [x23, #0x10]\n" - "ldr q22, [x23, #0x20]\n" - "ldr q23, [x23, #0x30]\n" + "ldr q12, [x26, #0x0]\n" + "ldr q13, [x26, #0x10]\n" + "ldr q14, [x26, #0x20]\n" + "ldr q15, [x26, #0x30]\n" + "ldr q16, [x25, #0x0]\n" + "ldr q17, [x25, #0x10]\n" + "ldr q18, [x25, #0x20]\n" + "ldr q19, [x25, #0x30]\n" + "ldr q20, [x24, #0x0]\n" + "ldr q21, [x24, #0x10]\n" + "ldr q22, [x24, #0x20]\n" + "ldr q23, [x24, #0x30]\n" "b 113f\n" "112:" // Height 4: no accumulate "movi v8.16b, #0x0\n" @@ -1419,8 +1421,8 @@ void a64_hybrid_fp32_mla_6x16 ( "mov x28, #0x0\n" "114:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 115f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1560,14 +1562,14 @@ void a64_hybrid_fp32_mla_6x16 ( "fmla v9.4s, v7.4s, v0.s[0]\n" "fmla v13.4s, v7.4s, v1.s[0]\n" "add x23, x23, #0x10\n" - "sub x27, x27, #0x4\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla v17.4s, v7.4s, v2.s[0]\n" "fmla v21.4s, v7.4s, v3.s[0]\n" "ldr q24, [x10, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla v10.4s, v25.4s, v0.s[0]\n" "fmla v14.4s, v25.4s, v1.s[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "sub x27, x27, #0x4\n" "prfm pldl1keep, [x24, #0x80]\n" "fmla v18.4s, v25.4s, v2.s[0]\n" "fmla v22.4s, v25.4s, v3.s[0]\n" @@ -1673,17 +1675,17 @@ void a64_hybrid_fp32_mla_6x16 ( "cmp x28, x20\n" "bne 114b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x9, #0x0]\n" - "add x23, x24, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" + "prfm pstl1keep, [x26, #0x0]\n" + "add x25, x26, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" + "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x24, #0x0]\n" - "prfm pstl1keep, [x23, #0x0]\n" "tbz %x[flags], #1, 122f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v25.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v25.4s }, [x21]\n" "ld1r { v24.4s }, [x20]\n" "fmin v8.4s, v8.4s, v25.4s\n" "fmin v9.4s, v9.4s, v25.4s\n" @@ -1723,95 +1725,95 @@ void a64_hybrid_fp32_mla_6x16 ( "tbz x11, #3, 126f\n" "st1 { v8.4s }, [x9], #0x10\n" "st1 { v9.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v13.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v17.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" - "st1 { v21.4s }, [x23], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v13.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v17.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" + "st1 { v21.4s }, [x24], #0x10\n" "tbz x11, #2, 124f\n" "st1 { v10.4s }, [x9], #0x10\n" - "st1 { v14.4s }, [x25], #0x10\n" - "st1 { v18.4s }, [x24], #0x10\n" - "st1 { v22.4s }, [x23], #0x10\n" + "st1 { v14.4s }, [x26], #0x10\n" + "st1 { v18.4s }, [x25], #0x10\n" + "st1 { v22.4s }, [x24], #0x10\n" "tbz x11, #1, 123f\n" "str d11, [x9], #0x8\n" - "str d15, [x25], #0x8\n" - "str d19, [x24], #0x8\n" - "str d23, [x23], #0x8\n" + "str d15, [x26], #0x8\n" + "str d19, [x25], #0x8\n" + "str d23, [x24], #0x8\n" "tbz x11, #0, 130f\n" "st1 { v11.s }[2], [x9]\n" - "st1 { v15.s }[2], [x25]\n" - "st1 { v19.s }[2], [x24]\n" - "st1 { v23.s }[2], [x23]\n" + "st1 { v15.s }[2], [x26]\n" + "st1 { v19.s }[2], [x25]\n" + "st1 { v23.s }[2], [x24]\n" "b 130f\n" "123:" // Height 4: Partial direct writeback: partial_1_12 "tbz x11, #0, 130f\n" "str s11, [x9, #0x0]\n" - "str s15, [x25, #0x0]\n" - "str s19, [x24, #0x0]\n" - "str s23, [x23, #0x0]\n" + "str s15, [x26, #0x0]\n" + "str s19, [x25, #0x0]\n" + "str s23, [x24, #0x0]\n" "b 130f\n" "124:" // Height 4: Partial direct writeback: partial_2_8 "tbz x11, #1, 125f\n" "str d10, [x9], #0x8\n" - "str d14, [x25], #0x8\n" - "str d18, [x24], #0x8\n" - "str d22, [x23], #0x8\n" + "str d14, [x26], #0x8\n" + "str d18, [x25], #0x8\n" + "str d22, [x24], #0x8\n" "tbz x11, #0, 130f\n" "st1 { v10.s }[2], [x9]\n" - "st1 { v14.s }[2], [x25]\n" - "st1 { v18.s }[2], [x24]\n" - "st1 { v22.s }[2], [x23]\n" + "st1 { v14.s }[2], [x26]\n" + "st1 { v18.s }[2], [x25]\n" + "st1 { v22.s }[2], [x24]\n" "b 130f\n" "125:" // Height 4: Partial direct writeback: partial_1_8 "tbz x11, #0, 130f\n" "str s10, [x9, #0x0]\n" - "str s14, [x25, #0x0]\n" - "str s18, [x24, #0x0]\n" - "str s22, [x23, #0x0]\n" + "str s14, [x26, #0x0]\n" + "str s18, [x25, #0x0]\n" + "str s22, [x24, #0x0]\n" "b 130f\n" "126:" // Height 4: Partial direct writeback: partial_4_0 "tbz x11, #2, 128f\n" "st1 { v8.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" "tbz x11, #1, 127f\n" "str d9, [x9], #0x8\n" - "str d13, [x25], #0x8\n" - "str d17, [x24], #0x8\n" - "str d21, [x23], #0x8\n" + "str d13, [x26], #0x8\n" + "str d17, [x25], #0x8\n" + "str d21, [x24], #0x8\n" "tbz x11, #0, 130f\n" "st1 { v9.s }[2], [x9]\n" - "st1 { v13.s }[2], [x25]\n" - "st1 { v17.s }[2], [x24]\n" - "st1 { v21.s }[2], [x23]\n" + "st1 { v13.s }[2], [x26]\n" + "st1 { v17.s }[2], [x25]\n" + "st1 { v21.s }[2], [x24]\n" "b 130f\n" "127:" // Height 4: Partial direct writeback: partial_1_4 "tbz x11, #0, 130f\n" "str s9, [x9, #0x0]\n" - "str s13, [x25, #0x0]\n" - "str s17, [x24, #0x0]\n" - "str s21, [x23, #0x0]\n" + "str s13, [x26, #0x0]\n" + "str s17, [x25, #0x0]\n" + "str s21, [x24, #0x0]\n" "b 130f\n" "128:" // Height 4: Partial direct writeback: partial_2_0 "tbz x11, #1, 129f\n" "str d8, [x9], #0x8\n" - "str d12, [x25], #0x8\n" - "str d16, [x24], #0x8\n" - "str d20, [x23], #0x8\n" + "str d12, [x26], #0x8\n" + "str d16, [x25], #0x8\n" + "str d20, [x24], #0x8\n" "tbz x11, #0, 130f\n" "st1 { v8.s }[2], [x9]\n" - "st1 { v12.s }[2], [x25]\n" - "st1 { v16.s }[2], [x24]\n" - "st1 { v20.s }[2], [x23]\n" + "st1 { v12.s }[2], [x26]\n" + "st1 { v16.s }[2], [x25]\n" + "st1 { v20.s }[2], [x24]\n" "b 130f\n" "129:" // Height 4: Partial direct writeback: partial_1_0 "str s8, [x9, #0x0]\n" - "str s12, [x25, #0x0]\n" - "str s16, [x24, #0x0]\n" - "str s20, [x23, #0x0]\n" + "str s12, [x26, #0x0]\n" + "str s16, [x25, #0x0]\n" + "str s20, [x24, #0x0]\n" "130:" // Height 4: Partial direct writeback: Done "b 132f\n" "131:" // Height 4: Full writeback @@ -1820,43 +1822,43 @@ void a64_hybrid_fp32_mla_6x16 ( "str q10, [x9, #0x20]\n" "str q11, [x9, #0x30]\n" "add x9, x9, #0x40\n" - "str q12, [x25, #0x0]\n" - "str q13, [x25, #0x10]\n" - "str q14, [x25, #0x20]\n" - "str q15, [x25, #0x30]\n" - "str q16, [x24, #0x0]\n" - "str q17, [x24, #0x10]\n" - "str q18, [x24, #0x20]\n" - "str q19, [x24, #0x30]\n" - "str q20, [x23, #0x0]\n" - "str q21, [x23, #0x10]\n" - "str q22, [x23, #0x20]\n" - "str q23, [x23, #0x30]\n" + "str q12, [x26, #0x0]\n" + "str q13, [x26, #0x10]\n" + "str q14, [x26, #0x20]\n" + "str q15, [x26, #0x30]\n" + "str q16, [x25, #0x0]\n" + "str q17, [x25, #0x10]\n" + "str q18, [x25, #0x20]\n" + "str q19, [x25, #0x30]\n" + "str q20, [x24, #0x0]\n" + "str q21, [x24, #0x10]\n" + "str q22, [x24, #0x20]\n" + "str q23, [x24, #0x30]\n" "132:" // Height 4: Writeback done "subs x11, x11, #0x10\n" "bgt 101b\n" "b 200f\n" "133:" // Height 5 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "134:" // Height 5: Column loop "cbz x12, 135f\n" "ldr q8, [x12, #0x0]\n" "ldr q9, [x12, #0x10]\n" - "mov v12.16b, v8.16b\n" - "mov v13.16b, v9.16b\n" "ldr q10, [x12, #0x20]\n" "ldr q11, [x12, #0x30]\n" - "mov v14.16b, v10.16b\n" - "mov v15.16b, v11.16b\n" + "add x12, x12, #0x40\n" + "mov v12.16b, v8.16b\n" + "mov v13.16b, v9.16b\n" "mov v16.16b, v8.16b\n" "mov v17.16b, v9.16b\n" - "add x12, x12, #0x40\n" + "mov v20.16b, v8.16b\n" + "mov v14.16b, v10.16b\n" + "mov v15.16b, v11.16b\n" "mov v18.16b, v10.16b\n" "mov v19.16b, v11.16b\n" - "mov v20.16b, v8.16b\n" "mov v21.16b, v9.16b\n" "mov v22.16b, v10.16b\n" "mov v23.16b, v11.16b\n" @@ -1868,128 +1870,128 @@ void a64_hybrid_fp32_mla_6x16 ( "135:" // Height 5: no bias "tbz %x[flags], #0, 145f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" + "cmp x11, #0x10\n" + "add x26, x9, 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" - "cmp x11, #0x10\n" - "add x22, x23, x20, LSL #2\n" "bge 144f\n" "tbz x11, #3, 139f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" - "ld1 { v24.4s }, [x22], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" + "ld1 { v24.4s }, [x23], #0x10\n" "ld1 { v9.4s }, [x9], #0x10\n" - "ld1 { v13.4s }, [x25], #0x10\n" - "ld1 { v17.4s }, [x24], #0x10\n" - "ld1 { v21.4s }, [x23], #0x10\n" - "ld1 { v25.4s }, [x22], #0x10\n" + "ld1 { v13.4s }, [x26], #0x10\n" + "ld1 { v17.4s }, [x25], #0x10\n" + "ld1 { v21.4s }, [x24], #0x10\n" + "ld1 { v25.4s }, [x23], #0x10\n" "tbz x11, #2, 137f\n" "ld1 { v10.4s }, [x9], #0x10\n" - "ld1 { v14.4s }, [x25], #0x10\n" - "ld1 { v18.4s }, [x24], #0x10\n" - "ld1 { v22.4s }, [x23], #0x10\n" - "ld1 { v26.4s }, [x22], #0x10\n" + "ld1 { v14.4s }, [x26], #0x10\n" + "ld1 { v18.4s }, [x25], #0x10\n" + "ld1 { v22.4s }, [x24], #0x10\n" + "ld1 { v26.4s }, [x23], #0x10\n" "tbz x11, #1, 136f\n" "ldr d11, [x9], #0x8\n" - "ldr d15, [x25], #0x8\n" + "ldr d15, [x26], #0x8\n" "mov x20, #0x38\n" - "ldr d19, [x24], #0x8\n" - "ldr d23, [x23], #0x8\n" - "ldr d27, [x22], #0x8\n" + "ldr d19, [x25], #0x8\n" + "ldr d23, [x24], #0x8\n" + "ldr d27, [x23], #0x8\n" "tbz x11, #0, 143f\n" "ld1 { v11.s }[2], [x9]\n" - "ld1 { v15.s }[2], [x25]\n" - "ld1 { v19.s }[2], [x24]\n" - "ld1 { v23.s }[2], [x23]\n" - "ld1 { v27.s }[2], [x22]\n" + "ld1 { v15.s }[2], [x26]\n" + "ld1 { v19.s }[2], [x25]\n" + "ld1 { v23.s }[2], [x24]\n" + "ld1 { v27.s }[2], [x23]\n" "b 143f\n" "136:" // Height 5: Partial accumulate: partial_1_12 "mov x20, #0x30\n" "tbz x11, #0, 143f\n" "ldr s11, [x9, #0x0]\n" - "ldr s15, [x25, #0x0]\n" - "ldr s19, [x24, #0x0]\n" - "ldr s23, [x23, #0x0]\n" - "ldr s27, [x22, #0x0]\n" + "ldr s15, [x26, #0x0]\n" + "ldr s19, [x25, #0x0]\n" + "ldr s23, [x24, #0x0]\n" + "ldr s27, [x23, #0x0]\n" "b 143f\n" "137:" // Height 5: Partial accumulate: partial_2_8 "tbz x11, #1, 138f\n" "ldr d10, [x9], #0x8\n" - "ldr d14, [x25], #0x8\n" + "ldr d14, [x26], #0x8\n" "mov x20, #0x28\n" - "ldr d18, [x24], #0x8\n" - "ldr d22, [x23], #0x8\n" - "ldr d26, [x22], #0x8\n" + "ldr d18, [x25], #0x8\n" + "ldr d22, [x24], #0x8\n" + "ldr d26, [x23], #0x8\n" "tbz x11, #0, 143f\n" "ld1 { v10.s }[2], [x9]\n" - "ld1 { v14.s }[2], [x25]\n" - "ld1 { v18.s }[2], [x24]\n" - "ld1 { v22.s }[2], [x23]\n" - "ld1 { v26.s }[2], [x22]\n" + "ld1 { v14.s }[2], [x26]\n" + "ld1 { v18.s }[2], [x25]\n" + "ld1 { v22.s }[2], [x24]\n" + "ld1 { v26.s }[2], [x23]\n" "b 143f\n" "138:" // Height 5: Partial accumulate: partial_1_8 "mov x20, #0x20\n" "tbz x11, #0, 143f\n" "ldr s10, [x9, #0x0]\n" - "ldr s14, [x25, #0x0]\n" - "ldr s18, [x24, #0x0]\n" - "ldr s22, [x23, #0x0]\n" - "ldr s26, [x22, #0x0]\n" + "ldr s14, [x26, #0x0]\n" + "ldr s18, [x25, #0x0]\n" + "ldr s22, [x24, #0x0]\n" + "ldr s26, [x23, #0x0]\n" "b 143f\n" "139:" // Height 5: Partial accumulate: partial_4_0 "tbz x11, #2, 141f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" - "ld1 { v24.4s }, [x22], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" + "ld1 { v24.4s }, [x23], #0x10\n" "tbz x11, #1, 140f\n" "ldr d9, [x9], #0x8\n" - "ldr d13, [x25], #0x8\n" + "ldr d13, [x26], #0x8\n" "mov x20, #0x18\n" - "ldr d17, [x24], #0x8\n" - "ldr d21, [x23], #0x8\n" - "ldr d25, [x22], #0x8\n" + "ldr d17, [x25], #0x8\n" + "ldr d21, [x24], #0x8\n" + "ldr d25, [x23], #0x8\n" "tbz x11, #0, 143f\n" "ld1 { v9.s }[2], [x9]\n" - "ld1 { v13.s }[2], [x25]\n" - "ld1 { v17.s }[2], [x24]\n" - "ld1 { v21.s }[2], [x23]\n" - "ld1 { v25.s }[2], [x22]\n" + "ld1 { v13.s }[2], [x26]\n" + "ld1 { v17.s }[2], [x25]\n" + "ld1 { v21.s }[2], [x24]\n" + "ld1 { v25.s }[2], [x23]\n" "b 143f\n" "140:" // Height 5: Partial accumulate: partial_1_4 "mov x20, #0x10\n" "tbz x11, #0, 143f\n" "ldr s9, [x9, #0x0]\n" - "ldr s13, [x25, #0x0]\n" - "ldr s17, [x24, #0x0]\n" - "ldr s21, [x23, #0x0]\n" - "ldr s25, [x22, #0x0]\n" + "ldr s13, [x26, #0x0]\n" + "ldr s17, [x25, #0x0]\n" + "ldr s21, [x24, #0x0]\n" + "ldr s25, [x23, #0x0]\n" "b 143f\n" "141:" // Height 5: Partial accumulate: partial_2_0 "tbz x11, #1, 142f\n" "ldr d8, [x9], #0x8\n" - "ldr d12, [x25], #0x8\n" + "ldr d12, [x26], #0x8\n" "mov x20, #0x8\n" - "ldr d16, [x24], #0x8\n" - "ldr d20, [x23], #0x8\n" - "ldr d24, [x22], #0x8\n" + "ldr d16, [x25], #0x8\n" + "ldr d20, [x24], #0x8\n" + "ldr d24, [x23], #0x8\n" "tbz x11, #0, 143f\n" "ld1 { v8.s }[2], [x9]\n" - "ld1 { v12.s }[2], [x25]\n" - "ld1 { v16.s }[2], [x24]\n" - "ld1 { v20.s }[2], [x23]\n" - "ld1 { v24.s }[2], [x22]\n" + "ld1 { v12.s }[2], [x26]\n" + "ld1 { v16.s }[2], [x25]\n" + "ld1 { v20.s }[2], [x24]\n" + "ld1 { v24.s }[2], [x23]\n" "b 143f\n" "142:" // Height 5: Partial accumulate: partial_1_0 "ldr s8, [x9, #0x0]\n" - "ldr s12, [x25, #0x0]\n" + "ldr s12, [x26, #0x0]\n" "mov x20, #0x0\n" - "ldr s16, [x24, #0x0]\n" - "ldr s20, [x23, #0x0]\n" - "ldr s24, [x22, #0x0]\n" + "ldr s16, [x25, #0x0]\n" + "ldr s20, [x24, #0x0]\n" + "ldr s24, [x23, #0x0]\n" "143:" // Height 5: Partial accumulate: Done "sub x9, x9, x20\n" "b 146f\n" @@ -1998,22 +2000,22 @@ void a64_hybrid_fp32_mla_6x16 ( "ldr q9, [x9, #0x10]\n" "ldr q10, [x9, #0x20]\n" "ldr q11, [x9, #0x30]\n" - "ldr q12, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "ldr q14, [x25, #0x20]\n" - "ldr q15, [x25, #0x30]\n" - "ldr q16, [x24, #0x0]\n" - "ldr q17, [x24, #0x10]\n" - "ldr q18, [x24, #0x20]\n" - "ldr q19, [x24, #0x30]\n" - "ldr q20, [x23, #0x0]\n" - "ldr q21, [x23, #0x10]\n" - "ldr q22, [x23, #0x20]\n" - "ldr q23, [x23, #0x30]\n" - "ldr q24, [x22, #0x0]\n" - "ldr q25, [x22, #0x10]\n" - "ldr q26, [x22, #0x20]\n" - "ldr q27, [x22, #0x30]\n" + "ldr q12, [x26, #0x0]\n" + "ldr q13, [x26, #0x10]\n" + "ldr q14, [x26, #0x20]\n" + "ldr q15, [x26, #0x30]\n" + "ldr q16, [x25, #0x0]\n" + "ldr q17, [x25, #0x10]\n" + "ldr q18, [x25, #0x20]\n" + "ldr q19, [x25, #0x30]\n" + "ldr q20, [x24, #0x0]\n" + "ldr q21, [x24, #0x10]\n" + "ldr q22, [x24, #0x20]\n" + "ldr q23, [x24, #0x30]\n" + "ldr q24, [x23, #0x0]\n" + "ldr q25, [x23, #0x10]\n" + "ldr q26, [x23, #0x20]\n" + "ldr q27, [x23, #0x30]\n" "b 146f\n" "145:" // Height 5: no accumulate "movi v8.16b, #0x0\n" @@ -2040,8 +2042,8 @@ void a64_hybrid_fp32_mla_6x16 ( "mov x28, #0x0\n" "147:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 148f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -2207,12 +2209,12 @@ void a64_hybrid_fp32_mla_6x16 ( "add x22, x22, #0x10\n" "fmla v13.4s, v7.4s, v1.s[0]\n" "fmla v17.4s, v7.4s, v2.s[0]\n" - "sub x27, x27, #0x4\n" "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "fmla v21.4s, v7.4s, v3.s[0]\n" "fmla v25.4s, v7.4s, v4.s[0]\n" "ldr q28, [x10, #0x30]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "sub x27, x27, #0x4\n" "fmla v10.4s, v29.4s, v0.s[0]\n" "fmla v14.4s, v29.4s, v1.s[0]\n" "prfm pldl1keep, [x24, #0x80]\n" @@ -2309,9 +2311,9 @@ void a64_hybrid_fp32_mla_6x16 ( "ldr s31, [x23], #0x4\n" "ldr s30, [x22], #0x4\n" "ldr q29, [x10, #0x0]\n" + "ldr q28, [x10, #0x10]\n" "fmla v8.4s, v29.4s, v2.s[0]\n" "fmla v12.4s, v29.4s, v1.s[0]\n" - "ldr q28, [x10, #0x10]\n" "fmla v16.4s, v29.4s, v0.s[0]\n" "fmla v20.4s, v29.4s, v31.s[0]\n" "fmla v24.4s, v29.4s, v30.s[0]\n" @@ -2340,19 +2342,19 @@ void a64_hybrid_fp32_mla_6x16 ( "cmp x28, x20\n" "bne 147b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x9, #0x0]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" + "prfm pstl1keep, [x26, #0x0]\n" + "add x25, x26, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" + "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x24, #0x0]\n" + "add x23, x24, x20, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" - "prfm pstl1keep, [x22, #0x0]\n" "tbz %x[flags], #1, 155f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v29.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v29.4s }, [x21]\n" "ld1r { v28.4s }, [x20]\n" "fmin v8.4s, v8.4s, v29.4s\n" "fmin v9.4s, v9.4s, v29.4s\n" @@ -2400,111 +2402,111 @@ void a64_hybrid_fp32_mla_6x16 ( "tbz x11, #3, 159f\n" "st1 { v8.4s }, [x9], #0x10\n" "st1 { v9.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v13.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v17.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" - "st1 { v21.4s }, [x23], #0x10\n" - "st1 { v24.4s }, [x22], #0x10\n" - "st1 { v25.4s }, [x22], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v13.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v17.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" + "st1 { v21.4s }, [x24], #0x10\n" + "st1 { v24.4s }, [x23], #0x10\n" + "st1 { v25.4s }, [x23], #0x10\n" "tbz x11, #2, 157f\n" "st1 { v10.4s }, [x9], #0x10\n" - "st1 { v14.4s }, [x25], #0x10\n" - "st1 { v18.4s }, [x24], #0x10\n" - "st1 { v22.4s }, [x23], #0x10\n" - "st1 { v26.4s }, [x22], #0x10\n" + "st1 { v14.4s }, [x26], #0x10\n" + "st1 { v18.4s }, [x25], #0x10\n" + "st1 { v22.4s }, [x24], #0x10\n" + "st1 { v26.4s }, [x23], #0x10\n" "tbz x11, #1, 156f\n" "str d11, [x9], #0x8\n" - "str d15, [x25], #0x8\n" - "str d19, [x24], #0x8\n" - "str d23, [x23], #0x8\n" - "str d27, [x22], #0x8\n" + "str d15, [x26], #0x8\n" + "str d19, [x25], #0x8\n" + "str d23, [x24], #0x8\n" + "str d27, [x23], #0x8\n" "tbz x11, #0, 163f\n" "st1 { v11.s }[2], [x9]\n" - "st1 { v15.s }[2], [x25]\n" - "st1 { v19.s }[2], [x24]\n" - "st1 { v23.s }[2], [x23]\n" - "st1 { v27.s }[2], [x22]\n" + "st1 { v15.s }[2], [x26]\n" + "st1 { v19.s }[2], [x25]\n" + "st1 { v23.s }[2], [x24]\n" + "st1 { v27.s }[2], [x23]\n" "b 163f\n" "156:" // Height 5: Partial direct writeback: partial_1_12 "tbz x11, #0, 163f\n" "str s11, [x9, #0x0]\n" - "str s15, [x25, #0x0]\n" - "str s19, [x24, #0x0]\n" - "str s23, [x23, #0x0]\n" - "str s27, [x22, #0x0]\n" + "str s15, [x26, #0x0]\n" + "str s19, [x25, #0x0]\n" + "str s23, [x24, #0x0]\n" + "str s27, [x23, #0x0]\n" "b 163f\n" "157:" // Height 5: Partial direct writeback: partial_2_8 "tbz x11, #1, 158f\n" "str d10, [x9], #0x8\n" - "str d14, [x25], #0x8\n" - "str d18, [x24], #0x8\n" - "str d22, [x23], #0x8\n" - "str d26, [x22], #0x8\n" + "str d14, [x26], #0x8\n" + "str d18, [x25], #0x8\n" + "str d22, [x24], #0x8\n" + "str d26, [x23], #0x8\n" "tbz x11, #0, 163f\n" "st1 { v10.s }[2], [x9]\n" - "st1 { v14.s }[2], [x25]\n" - "st1 { v18.s }[2], [x24]\n" - "st1 { v22.s }[2], [x23]\n" - "st1 { v26.s }[2], [x22]\n" + "st1 { v14.s }[2], [x26]\n" + "st1 { v18.s }[2], [x25]\n" + "st1 { v22.s }[2], [x24]\n" + "st1 { v26.s }[2], [x23]\n" "b 163f\n" "158:" // Height 5: Partial direct writeback: partial_1_8 "tbz x11, #0, 163f\n" "str s10, [x9, #0x0]\n" - "str s14, [x25, #0x0]\n" - "str s18, [x24, #0x0]\n" - "str s22, [x23, #0x0]\n" - "str s26, [x22, #0x0]\n" + "str s14, [x26, #0x0]\n" + "str s18, [x25, #0x0]\n" + "str s22, [x24, #0x0]\n" + "str s26, [x23, #0x0]\n" "b 163f\n" "159:" // Height 5: Partial direct writeback: partial_4_0 "tbz x11, #2, 161f\n" "st1 { v8.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" - "st1 { v24.4s }, [x22], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" + "st1 { v24.4s }, [x23], #0x10\n" "tbz x11, #1, 160f\n" "str d9, [x9], #0x8\n" - "str d13, [x25], #0x8\n" - "str d17, [x24], #0x8\n" - "str d21, [x23], #0x8\n" - "str d25, [x22], #0x8\n" + "str d13, [x26], #0x8\n" + "str d17, [x25], #0x8\n" + "str d21, [x24], #0x8\n" + "str d25, [x23], #0x8\n" "tbz x11, #0, 163f\n" "st1 { v9.s }[2], [x9]\n" - "st1 { v13.s }[2], [x25]\n" - "st1 { v17.s }[2], [x24]\n" - "st1 { v21.s }[2], [x23]\n" - "st1 { v25.s }[2], [x22]\n" + "st1 { v13.s }[2], [x26]\n" + "st1 { v17.s }[2], [x25]\n" + "st1 { v21.s }[2], [x24]\n" + "st1 { v25.s }[2], [x23]\n" "b 163f\n" "160:" // Height 5: Partial direct writeback: partial_1_4 "tbz x11, #0, 163f\n" "str s9, [x9, #0x0]\n" - "str s13, [x25, #0x0]\n" - "str s17, [x24, #0x0]\n" - "str s21, [x23, #0x0]\n" - "str s25, [x22, #0x0]\n" + "str s13, [x26, #0x0]\n" + "str s17, [x25, #0x0]\n" + "str s21, [x24, #0x0]\n" + "str s25, [x23, #0x0]\n" "b 163f\n" "161:" // Height 5: Partial direct writeback: partial_2_0 "tbz x11, #1, 162f\n" "str d8, [x9], #0x8\n" - "str d12, [x25], #0x8\n" - "str d16, [x24], #0x8\n" - "str d20, [x23], #0x8\n" - "str d24, [x22], #0x8\n" + "str d12, [x26], #0x8\n" + "str d16, [x25], #0x8\n" + "str d20, [x24], #0x8\n" + "str d24, [x23], #0x8\n" "tbz x11, #0, 163f\n" "st1 { v8.s }[2], [x9]\n" - "st1 { v12.s }[2], [x25]\n" - "st1 { v16.s }[2], [x24]\n" - "st1 { v20.s }[2], [x23]\n" - "st1 { v24.s }[2], [x22]\n" + "st1 { v12.s }[2], [x26]\n" + "st1 { v16.s }[2], [x25]\n" + "st1 { v20.s }[2], [x24]\n" + "st1 { v24.s }[2], [x23]\n" "b 163f\n" "162:" // Height 5: Partial direct writeback: partial_1_0 "str s8, [x9, #0x0]\n" - "str s12, [x25, #0x0]\n" - "str s16, [x24, #0x0]\n" - "str s20, [x23, #0x0]\n" - "str s24, [x22, #0x0]\n" + "str s12, [x26, #0x0]\n" + "str s16, [x25, #0x0]\n" + "str s20, [x24, #0x0]\n" + "str s24, [x23, #0x0]\n" "163:" // Height 5: Partial direct writeback: Done "b 165f\n" "164:" // Height 5: Full writeback @@ -2513,50 +2515,51 @@ void a64_hybrid_fp32_mla_6x16 ( "str q10, [x9, #0x20]\n" "str q11, [x9, #0x30]\n" "add x9, x9, #0x40\n" - "str q12, [x25, #0x0]\n" - "str q13, [x25, #0x10]\n" - "str q14, [x25, #0x20]\n" - "str q15, [x25, #0x30]\n" - "str q16, [x24, #0x0]\n" - "str q17, [x24, #0x10]\n" - "str q18, [x24, #0x20]\n" - "str q19, [x24, #0x30]\n" - "str q20, [x23, #0x0]\n" - "str q21, [x23, #0x10]\n" - "str q22, [x23, #0x20]\n" - "str q23, [x23, #0x30]\n" - "str q24, [x22, #0x0]\n" - "str q25, [x22, #0x10]\n" - "str q26, [x22, #0x20]\n" - "str q27, [x22, #0x30]\n" + "str q12, [x26, #0x0]\n" + "str q13, [x26, #0x10]\n" + "str q14, [x26, #0x20]\n" + "str q15, [x26, #0x30]\n" + "str q16, [x25, #0x0]\n" + "str q17, [x25, #0x10]\n" + "str q18, [x25, #0x20]\n" + "str q19, [x25, #0x30]\n" + "str q20, [x24, #0x0]\n" + "str q21, [x24, #0x10]\n" + "str q22, [x24, #0x20]\n" + "str q23, [x24, #0x30]\n" + "str q24, [x23, #0x0]\n" + "str q25, [x23, #0x10]\n" + "str q26, [x23, #0x20]\n" + "str q27, [x23, #0x30]\n" "165:" // Height 5: Writeback done "subs x11, x11, #0x10\n" "bgt 134b\n" "b 200f\n" "166:" // Height 6 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "mov x20, #0x18\n" - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" - "madd %x[output_ptr], x21, x20, %x[output_ptr]\n" + "madd x20, x21, x20, x9\n" + "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n" "167:" // Height 6: Column loop "cbz x12, 168f\n" "ldr q8, [x12, #0x0]\n" "ldr q9, [x12, #0x10]\n" - "mov v12.16b, v8.16b\n" - "mov v13.16b, v9.16b\n" "ldr q10, [x12, #0x20]\n" "ldr q11, [x12, #0x30]\n" - "mov v14.16b, v10.16b\n" - "mov v15.16b, v11.16b\n" + "add x12, x12, #0x40\n" + "mov v12.16b, v8.16b\n" + "mov v13.16b, v9.16b\n" "mov v16.16b, v8.16b\n" "mov v17.16b, v9.16b\n" - "add x12, x12, #0x40\n" + "mov v20.16b, v8.16b\n" + "mov v14.16b, v10.16b\n" + "mov v15.16b, v11.16b\n" "mov v18.16b, v10.16b\n" "mov v19.16b, v11.16b\n" - "mov v20.16b, v8.16b\n" "mov v21.16b, v9.16b\n" "mov v22.16b, v10.16b\n" "mov v23.16b, v11.16b\n" @@ -2572,145 +2575,145 @@ void a64_hybrid_fp32_mla_6x16 ( "168:" // Height 6: no bias "tbz %x[flags], #0, 178f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" + "cmp x11, #0x10\n" + "add x26, x9, 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" - "cmp x11, #0x10\n" - "add x21, x22, x20, LSL #2\n" "bge 177f\n" "tbz x11, #3, 172f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" - "ld1 { v24.4s }, [x22], #0x10\n" - "ld1 { v28.4s }, [x21], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" + "ld1 { v24.4s }, [x23], #0x10\n" + "ld1 { v28.4s }, [x22], #0x10\n" "ld1 { v9.4s }, [x9], #0x10\n" - "ld1 { v13.4s }, [x25], #0x10\n" - "ld1 { v17.4s }, [x24], #0x10\n" - "ld1 { v21.4s }, [x23], #0x10\n" - "ld1 { v25.4s }, [x22], #0x10\n" - "ld1 { v29.4s }, [x21], #0x10\n" + "ld1 { v13.4s }, [x26], #0x10\n" + "ld1 { v17.4s }, [x25], #0x10\n" + "ld1 { v21.4s }, [x24], #0x10\n" + "ld1 { v25.4s }, [x23], #0x10\n" + "ld1 { v29.4s }, [x22], #0x10\n" "tbz x11, #2, 170f\n" "ld1 { v10.4s }, [x9], #0x10\n" - "ld1 { v14.4s }, [x25], #0x10\n" - "ld1 { v18.4s }, [x24], #0x10\n" - "ld1 { v22.4s }, [x23], #0x10\n" - "ld1 { v26.4s }, [x22], #0x10\n" - "ld1 { v30.4s }, [x21], #0x10\n" + "ld1 { v14.4s }, [x26], #0x10\n" + "ld1 { v18.4s }, [x25], #0x10\n" + "ld1 { v22.4s }, [x24], #0x10\n" + "ld1 { v26.4s }, [x23], #0x10\n" + "ld1 { v30.4s }, [x22], #0x10\n" "tbz x11, #1, 169f\n" "ldr d11, [x9], #0x8\n" - "ldr d15, [x25], #0x8\n" + "ldr d15, [x26], #0x8\n" "mov x20, #0x38\n" - "ldr d19, [x24], #0x8\n" - "ldr d23, [x23], #0x8\n" - "ldr d27, [x22], #0x8\n" - "ldr d31, [x21], #0x8\n" + "ldr d19, [x25], #0x8\n" + "ldr d23, [x24], #0x8\n" + "ldr d27, [x23], #0x8\n" + "ldr d31, [x22], #0x8\n" "tbz x11, #0, 176f\n" "ld1 { v11.s }[2], [x9]\n" - "ld1 { v15.s }[2], [x25]\n" - "ld1 { v19.s }[2], [x24]\n" - "ld1 { v23.s }[2], [x23]\n" - "ld1 { v27.s }[2], [x22]\n" - "ld1 { v31.s }[2], [x21]\n" + "ld1 { v15.s }[2], [x26]\n" + "ld1 { v19.s }[2], [x25]\n" + "ld1 { v23.s }[2], [x24]\n" + "ld1 { v27.s }[2], [x23]\n" + "ld1 { v31.s }[2], [x22]\n" "b 176f\n" "169:" // Height 6: Partial accumulate: partial_1_12 "mov x20, #0x30\n" "tbz x11, #0, 176f\n" "ldr s11, [x9, #0x0]\n" - "ldr s15, [x25, #0x0]\n" - "ldr s19, [x24, #0x0]\n" - "ldr s23, [x23, #0x0]\n" - "ldr s27, [x22, #0x0]\n" - "ldr s31, [x21, #0x0]\n" + "ldr s15, [x26, #0x0]\n" + "ldr s19, [x25, #0x0]\n" + "ldr s23, [x24, #0x0]\n" + "ldr s27, [x23, #0x0]\n" + "ldr s31, [x22, #0x0]\n" "b 176f\n" "170:" // Height 6: Partial accumulate: partial_2_8 "tbz x11, #1, 171f\n" "ldr d10, [x9], #0x8\n" - "ldr d14, [x25], #0x8\n" + "ldr d14, [x26], #0x8\n" "mov x20, #0x28\n" - "ldr d18, [x24], #0x8\n" - "ldr d22, [x23], #0x8\n" - "ldr d26, [x22], #0x8\n" - "ldr d30, [x21], #0x8\n" + "ldr d18, [x25], #0x8\n" + "ldr d22, [x24], #0x8\n" + "ldr d26, [x23], #0x8\n" + "ldr d30, [x22], #0x8\n" "tbz x11, #0, 176f\n" "ld1 { v10.s }[2], [x9]\n" - "ld1 { v14.s }[2], [x25]\n" - "ld1 { v18.s }[2], [x24]\n" - "ld1 { v22.s }[2], [x23]\n" - "ld1 { v26.s }[2], [x22]\n" - "ld1 { v30.s }[2], [x21]\n" + "ld1 { v14.s }[2], [x26]\n" + "ld1 { v18.s }[2], [x25]\n" + "ld1 { v22.s }[2], [x24]\n" + "ld1 { v26.s }[2], [x23]\n" + "ld1 { v30.s }[2], [x22]\n" "b 176f\n" "171:" // Height 6: Partial accumulate: partial_1_8 "mov x20, #0x20\n" "tbz x11, #0, 176f\n" "ldr s10, [x9, #0x0]\n" - "ldr s14, [x25, #0x0]\n" - "ldr s18, [x24, #0x0]\n" - "ldr s22, [x23, #0x0]\n" - "ldr s26, [x22, #0x0]\n" - "ldr s30, [x21, #0x0]\n" + "ldr s14, [x26, #0x0]\n" + "ldr s18, [x25, #0x0]\n" + "ldr s22, [x24, #0x0]\n" + "ldr s26, [x23, #0x0]\n" + "ldr s30, [x22, #0x0]\n" "b 176f\n" "172:" // Height 6: Partial accumulate: partial_4_0 "tbz x11, #2, 174f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" - "ld1 { v24.4s }, [x22], #0x10\n" - "ld1 { v28.4s }, [x21], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" + "ld1 { v24.4s }, [x23], #0x10\n" + "ld1 { v28.4s }, [x22], #0x10\n" "tbz x11, #1, 173f\n" "ldr d9, [x9], #0x8\n" - "ldr d13, [x25], #0x8\n" + "ldr d13, [x26], #0x8\n" "mov x20, #0x18\n" - "ldr d17, [x24], #0x8\n" - "ldr d21, [x23], #0x8\n" - "ldr d25, [x22], #0x8\n" - "ldr d29, [x21], #0x8\n" + "ldr d17, [x25], #0x8\n" + "ldr d21, [x24], #0x8\n" + "ldr d25, [x23], #0x8\n" + "ldr d29, [x22], #0x8\n" "tbz x11, #0, 176f\n" "ld1 { v9.s }[2], [x9]\n" - "ld1 { v13.s }[2], [x25]\n" - "ld1 { v17.s }[2], [x24]\n" - "ld1 { v21.s }[2], [x23]\n" - "ld1 { v25.s }[2], [x22]\n" - "ld1 { v29.s }[2], [x21]\n" + "ld1 { v13.s }[2], [x26]\n" + "ld1 { v17.s }[2], [x25]\n" + "ld1 { v21.s }[2], [x24]\n" + "ld1 { v25.s }[2], [x23]\n" + "ld1 { v29.s }[2], [x22]\n" "b 176f\n" "173:" // Height 6: Partial accumulate: partial_1_4 "mov x20, #0x10\n" "tbz x11, #0, 176f\n" "ldr s9, [x9, #0x0]\n" - "ldr s13, [x25, #0x0]\n" - "ldr s17, [x24, #0x0]\n" - "ldr s21, [x23, #0x0]\n" - "ldr s25, [x22, #0x0]\n" - "ldr s29, [x21, #0x0]\n" + "ldr s13, [x26, #0x0]\n" + "ldr s17, [x25, #0x0]\n" + "ldr s21, [x24, #0x0]\n" + "ldr s25, [x23, #0x0]\n" + "ldr s29, [x22, #0x0]\n" "b 176f\n" "174:" // Height 6: Partial accumulate: partial_2_0 "tbz x11, #1, 175f\n" "ldr d8, [x9], #0x8\n" - "ldr d12, [x25], #0x8\n" + "ldr d12, [x26], #0x8\n" "mov x20, #0x8\n" - "ldr d16, [x24], #0x8\n" - "ldr d20, [x23], #0x8\n" - "ldr d24, [x22], #0x8\n" - "ldr d28, [x21], #0x8\n" + "ldr d16, [x25], #0x8\n" + "ldr d20, [x24], #0x8\n" + "ldr d24, [x23], #0x8\n" + "ldr d28, [x22], #0x8\n" "tbz x11, #0, 176f\n" "ld1 { v8.s }[2], [x9]\n" - "ld1 { v12.s }[2], [x25]\n" - "ld1 { v16.s }[2], [x24]\n" - "ld1 { v20.s }[2], [x23]\n" - "ld1 { v24.s }[2], [x22]\n" - "ld1 { v28.s }[2], [x21]\n" + "ld1 { v12.s }[2], [x26]\n" + "ld1 { v16.s }[2], [x25]\n" + "ld1 { v20.s }[2], [x24]\n" + "ld1 { v24.s }[2], [x23]\n" + "ld1 { v28.s }[2], [x22]\n" "b 176f\n" "175:" // Height 6: Partial accumulate: partial_1_0 "ldr s8, [x9, #0x0]\n" - "ldr s12, [x25, #0x0]\n" + "ldr s12, [x26, #0x0]\n" "mov x20, #0x0\n" - "ldr s16, [x24, #0x0]\n" - "ldr s20, [x23, #0x0]\n" - "ldr s24, [x22, #0x0]\n" - "ldr s28, [x21, #0x0]\n" + "ldr s16, [x25, #0x0]\n" + "ldr s20, [x24, #0x0]\n" + "ldr s24, [x23, #0x0]\n" + "ldr s28, [x22, #0x0]\n" "176:" // Height 6: Partial accumulate: Done "sub x9, x9, x20\n" "b 179f\n" @@ -2719,26 +2722,26 @@ void a64_hybrid_fp32_mla_6x16 ( "ldr q9, [x9, #0x10]\n" "ldr q10, [x9, #0x20]\n" "ldr q11, [x9, #0x30]\n" - "ldr q12, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "ldr q14, [x25, #0x20]\n" - "ldr q15, [x25, #0x30]\n" - "ldr q16, [x24, #0x0]\n" - "ldr q17, [x24, #0x10]\n" - "ldr q18, [x24, #0x20]\n" - "ldr q19, [x24, #0x30]\n" - "ldr q20, [x23, #0x0]\n" - "ldr q21, [x23, #0x10]\n" - "ldr q22, [x23, #0x20]\n" - "ldr q23, [x23, #0x30]\n" - "ldr q24, [x22, #0x0]\n" - "ldr q25, [x22, #0x10]\n" - "ldr q26, [x22, #0x20]\n" - "ldr q27, [x22, #0x30]\n" - "ldr q28, [x21, #0x0]\n" - "ldr q29, [x21, #0x10]\n" - "ldr q30, [x21, #0x20]\n" - "ldr q31, [x21, #0x30]\n" + "ldr q12, [x26, #0x0]\n" + "ldr q13, [x26, #0x10]\n" + "ldr q14, [x26, #0x20]\n" + "ldr q15, [x26, #0x30]\n" + "ldr q16, [x25, #0x0]\n" + "ldr q17, [x25, #0x10]\n" + "ldr q18, [x25, #0x20]\n" + "ldr q19, [x25, #0x30]\n" + "ldr q20, [x24, #0x0]\n" + "ldr q21, [x24, #0x10]\n" + "ldr q22, [x24, #0x20]\n" + "ldr q23, [x24, #0x30]\n" + "ldr q24, [x23, #0x0]\n" + "ldr q25, [x23, #0x10]\n" + "ldr q26, [x23, #0x20]\n" + "ldr q27, [x23, #0x30]\n" + "ldr q28, [x22, #0x0]\n" + "ldr q29, [x22, #0x10]\n" + "ldr q30, [x22, #0x20]\n" + "ldr q31, [x22, #0x30]\n" "b 179f\n" "178:" // Height 6: no accumulate "movi v8.16b, #0x0\n" @@ -2769,8 +2772,8 @@ void a64_hybrid_fp32_mla_6x16 ( "mov x28, #0x0\n" "180:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 181f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -2960,18 +2963,18 @@ void a64_hybrid_fp32_mla_6x16 ( "fmla v9.4s, v7.4s, v0.s[0]\n" "fmla v13.4s, v7.4s, v1.s[0]\n" "add x21, x21, #0x10\n" - "sub x27, x27, #0x4\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla v17.4s, v7.4s, v2.s[0]\n" "fmla v21.4s, v7.4s, v3.s[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" "fmla v25.4s, v7.4s, v4.s[0]\n" "fmla v29.4s, v7.4s, v5.s[0]\n" "ldr q7, [x10, #0x30]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" "fmla v10.4s, v6.4s, v0.s[0]\n" "fmla v14.4s, v6.4s, v1.s[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" + "sub x27, x27, #0x4\n" "prfm pldl1keep, [x22, #0x80]\n" "fmla v18.4s, v6.4s, v2.s[0]\n" "fmla v22.4s, v6.4s, v3.s[0]\n" @@ -3115,21 +3118,21 @@ void a64_hybrid_fp32_mla_6x16 ( "cmp x28, x20\n" "bne 180b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x9, #0x0]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" + "prfm pstl1keep, [x26, #0x0]\n" + "add x25, x26, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" + "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x24, #0x0]\n" - "add x21, x22, x20, LSL #2\n" + "add x23, x24, x20, LSL #2\n" + "add x22, x23, x20, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "prfm pstl1keep, [x22, #0x0]\n" - "prfm pstl1keep, [x21, #0x0]\n" "tbz %x[flags], #1, 188f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v1.4s }, [x21]\n" "ld1r { v0.4s }, [x20]\n" "fmin v8.4s, v8.4s, v1.4s\n" "fmin v9.4s, v9.4s, v1.4s\n" @@ -3185,127 +3188,127 @@ void a64_hybrid_fp32_mla_6x16 ( "tbz x11, #3, 192f\n" "st1 { v8.4s }, [x9], #0x10\n" "st1 { v9.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v13.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v17.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" - "st1 { v21.4s }, [x23], #0x10\n" - "st1 { v24.4s }, [x22], #0x10\n" - "st1 { v25.4s }, [x22], #0x10\n" - "st1 { v28.4s }, [x21], #0x10\n" - "st1 { v29.4s }, [x21], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v13.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v17.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" + "st1 { v21.4s }, [x24], #0x10\n" + "st1 { v24.4s }, [x23], #0x10\n" + "st1 { v25.4s }, [x23], #0x10\n" + "st1 { v28.4s }, [x22], #0x10\n" + "st1 { v29.4s }, [x22], #0x10\n" "tbz x11, #2, 190f\n" "st1 { v10.4s }, [x9], #0x10\n" - "st1 { v14.4s }, [x25], #0x10\n" - "st1 { v18.4s }, [x24], #0x10\n" - "st1 { v22.4s }, [x23], #0x10\n" - "st1 { v26.4s }, [x22], #0x10\n" - "st1 { v30.4s }, [x21], #0x10\n" + "st1 { v14.4s }, [x26], #0x10\n" + "st1 { v18.4s }, [x25], #0x10\n" + "st1 { v22.4s }, [x24], #0x10\n" + "st1 { v26.4s }, [x23], #0x10\n" + "st1 { v30.4s }, [x22], #0x10\n" "tbz x11, #1, 189f\n" "str d11, [x9], #0x8\n" - "str d15, [x25], #0x8\n" - "str d19, [x24], #0x8\n" - "str d23, [x23], #0x8\n" - "str d27, [x22], #0x8\n" - "str d31, [x21], #0x8\n" + "str d15, [x26], #0x8\n" + "str d19, [x25], #0x8\n" + "str d23, [x24], #0x8\n" + "str d27, [x23], #0x8\n" + "str d31, [x22], #0x8\n" "tbz x11, #0, 196f\n" "st1 { v11.s }[2], [x9]\n" - "st1 { v15.s }[2], [x25]\n" - "st1 { v19.s }[2], [x24]\n" - "st1 { v23.s }[2], [x23]\n" - "st1 { v27.s }[2], [x22]\n" - "st1 { v31.s }[2], [x21]\n" + "st1 { v15.s }[2], [x26]\n" + "st1 { v19.s }[2], [x25]\n" + "st1 { v23.s }[2], [x24]\n" + "st1 { v27.s }[2], [x23]\n" + "st1 { v31.s }[2], [x22]\n" "b 196f\n" "189:" // Height 6: Partial direct writeback: partial_1_12 "tbz x11, #0, 196f\n" "str s11, [x9, #0x0]\n" - "str s15, [x25, #0x0]\n" - "str s19, [x24, #0x0]\n" - "str s23, [x23, #0x0]\n" - "str s27, [x22, #0x0]\n" - "str s31, [x21, #0x0]\n" + "str s15, [x26, #0x0]\n" + "str s19, [x25, #0x0]\n" + "str s23, [x24, #0x0]\n" + "str s27, [x23, #0x0]\n" + "str s31, [x22, #0x0]\n" "b 196f\n" "190:" // Height 6: Partial direct writeback: partial_2_8 "tbz x11, #1, 191f\n" "str d10, [x9], #0x8\n" - "str d14, [x25], #0x8\n" - "str d18, [x24], #0x8\n" - "str d22, [x23], #0x8\n" - "str d26, [x22], #0x8\n" - "str d30, [x21], #0x8\n" + "str d14, [x26], #0x8\n" + "str d18, [x25], #0x8\n" + "str d22, [x24], #0x8\n" + "str d26, [x23], #0x8\n" + "str d30, [x22], #0x8\n" "tbz x11, #0, 196f\n" "st1 { v10.s }[2], [x9]\n" - "st1 { v14.s }[2], [x25]\n" - "st1 { v18.s }[2], [x24]\n" - "st1 { v22.s }[2], [x23]\n" - "st1 { v26.s }[2], [x22]\n" - "st1 { v30.s }[2], [x21]\n" + "st1 { v14.s }[2], [x26]\n" + "st1 { v18.s }[2], [x25]\n" + "st1 { v22.s }[2], [x24]\n" + "st1 { v26.s }[2], [x23]\n" + "st1 { v30.s }[2], [x22]\n" "b 196f\n" "191:" // Height 6: Partial direct writeback: partial_1_8 "tbz x11, #0, 196f\n" "str s10, [x9, #0x0]\n" - "str s14, [x25, #0x0]\n" - "str s18, [x24, #0x0]\n" - "str s22, [x23, #0x0]\n" - "str s26, [x22, #0x0]\n" - "str s30, [x21, #0x0]\n" + "str s14, [x26, #0x0]\n" + "str s18, [x25, #0x0]\n" + "str s22, [x24, #0x0]\n" + "str s26, [x23, #0x0]\n" + "str s30, [x22, #0x0]\n" "b 196f\n" "192:" // Height 6: Partial direct writeback: partial_4_0 "tbz x11, #2, 194f\n" "st1 { v8.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" - "st1 { v24.4s }, [x22], #0x10\n" - "st1 { v28.4s }, [x21], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" + "st1 { v24.4s }, [x23], #0x10\n" + "st1 { v28.4s }, [x22], #0x10\n" "tbz x11, #1, 193f\n" "str d9, [x9], #0x8\n" - "str d13, [x25], #0x8\n" - "str d17, [x24], #0x8\n" - "str d21, [x23], #0x8\n" - "str d25, [x22], #0x8\n" - "str d29, [x21], #0x8\n" + "str d13, [x26], #0x8\n" + "str d17, [x25], #0x8\n" + "str d21, [x24], #0x8\n" + "str d25, [x23], #0x8\n" + "str d29, [x22], #0x8\n" "tbz x11, #0, 196f\n" "st1 { v9.s }[2], [x9]\n" - "st1 { v13.s }[2], [x25]\n" - "st1 { v17.s }[2], [x24]\n" - "st1 { v21.s }[2], [x23]\n" - "st1 { v25.s }[2], [x22]\n" - "st1 { v29.s }[2], [x21]\n" + "st1 { v13.s }[2], [x26]\n" + "st1 { v17.s }[2], [x25]\n" + "st1 { v21.s }[2], [x24]\n" + "st1 { v25.s }[2], [x23]\n" + "st1 { v29.s }[2], [x22]\n" "b 196f\n" "193:" // Height 6: Partial direct writeback: partial_1_4 "tbz x11, #0, 196f\n" "str s9, [x9, #0x0]\n" - "str s13, [x25, #0x0]\n" - "str s17, [x24, #0x0]\n" - "str s21, [x23, #0x0]\n" - "str s25, [x22, #0x0]\n" - "str s29, [x21, #0x0]\n" + "str s13, [x26, #0x0]\n" + "str s17, [x25, #0x0]\n" + "str s21, [x24, #0x0]\n" + "str s25, [x23, #0x0]\n" + "str s29, [x22, #0x0]\n" "b 196f\n" "194:" // Height 6: Partial direct writeback: partial_2_0 "tbz x11, #1, 195f\n" "str d8, [x9], #0x8\n" - "str d12, [x25], #0x8\n" - "str d16, [x24], #0x8\n" - "str d20, [x23], #0x8\n" - "str d24, [x22], #0x8\n" - "str d28, [x21], #0x8\n" + "str d12, [x26], #0x8\n" + "str d16, [x25], #0x8\n" + "str d20, [x24], #0x8\n" + "str d24, [x23], #0x8\n" + "str d28, [x22], #0x8\n" "tbz x11, #0, 196f\n" "st1 { v8.s }[2], [x9]\n" - "st1 { v12.s }[2], [x25]\n" - "st1 { v16.s }[2], [x24]\n" - "st1 { v20.s }[2], [x23]\n" - "st1 { v24.s }[2], [x22]\n" - "st1 { v28.s }[2], [x21]\n" + "st1 { v12.s }[2], [x26]\n" + "st1 { v16.s }[2], [x25]\n" + "st1 { v20.s }[2], [x24]\n" + "st1 { v24.s }[2], [x23]\n" + "st1 { v28.s }[2], [x22]\n" "b 196f\n" "195:" // Height 6: Partial direct writeback: partial_1_0 "str s8, [x9, #0x0]\n" - "str s12, [x25, #0x0]\n" - "str s16, [x24, #0x0]\n" - "str s20, [x23, #0x0]\n" - "str s24, [x22, #0x0]\n" - "str s28, [x21, #0x0]\n" + "str s12, [x26, #0x0]\n" + "str s16, [x25, #0x0]\n" + "str s20, [x24, #0x0]\n" + "str s24, [x23, #0x0]\n" + "str s28, [x22, #0x0]\n" "196:" // Height 6: Partial direct writeback: Done "b 198f\n" "197:" // Height 6: Full writeback @@ -3314,26 +3317,26 @@ void a64_hybrid_fp32_mla_6x16 ( "str q10, [x9, #0x20]\n" "str q11, [x9, #0x30]\n" "add x9, x9, #0x40\n" - "str q12, [x25, #0x0]\n" - "str q13, [x25, #0x10]\n" - "str q14, [x25, #0x20]\n" - "str q15, [x25, #0x30]\n" - "str q16, [x24, #0x0]\n" - "str q17, [x24, #0x10]\n" - "str q18, [x24, #0x20]\n" - "str q19, [x24, #0x30]\n" - "str q20, [x23, #0x0]\n" - "str q21, [x23, #0x10]\n" - "str q22, [x23, #0x20]\n" - "str q23, [x23, #0x30]\n" - "str q24, [x22, #0x0]\n" - "str q25, [x22, #0x10]\n" - "str q26, [x22, #0x20]\n" - "str q27, [x22, #0x30]\n" - "str q28, [x21, #0x0]\n" - "str q29, [x21, #0x10]\n" - "str q30, [x21, #0x20]\n" - "str q31, [x21, #0x30]\n" + "str q12, [x26, #0x0]\n" + "str q13, [x26, #0x10]\n" + "str q14, [x26, #0x20]\n" + "str q15, [x26, #0x30]\n" + "str q16, [x25, #0x0]\n" + "str q17, [x25, #0x10]\n" + "str q18, [x25, #0x20]\n" + "str q19, [x25, #0x30]\n" + "str q20, [x24, #0x0]\n" + "str q21, [x24, #0x10]\n" + "str q22, [x24, #0x20]\n" + "str q23, [x24, #0x30]\n" + "str q24, [x23, #0x0]\n" + "str q25, [x23, #0x10]\n" + "str q26, [x23, #0x20]\n" + "str q27, [x23, #0x30]\n" + "str q28, [x22, #0x0]\n" + "str q29, [x22, #0x10]\n" + "str q30, [x22, #0x20]\n" + "str q31, [x22, #0x30]\n" "198:" // Height 6: Writeback done "subs x11, x11, #0x10\n" "bgt 167b\n" @@ -3349,8 +3352,8 @@ void a64_hybrid_fp32_mla_6x16 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "200:" // 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)) + : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr) + : [args_ptr] "r" (&ka), [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_bias] "I" (offsetof(KernelArgs, bias)), [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_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |