diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp | 2163 |
1 files changed, 1082 insertions, 1081 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp index 0771829d37..867bcded1f 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8qs_mmla_6x16/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, 2023 Arm Limited. + * Copyright (c) 2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,18 +47,18 @@ void a64_hybrid_s8qs_mmla_6x16 ( size_t output_offset = {}; size_t input_initial_col = {}; size_t input_offset = {}; + void *output_ptr = {}; } 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; } @@ -97,9 +97,9 @@ void a64_hybrid_s8qs_mmla_6x16 ( "mov x14, %x[col_bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x11, %x[output_ptr]\n" - "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "2:" // Height 1: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" @@ -113,8 +113,8 @@ void a64_hybrid_s8qs_mmla_6x16 ( "mov x28, #0x0\n" "4:" // 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, 5f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -129,115 +129,115 @@ void a64_hybrid_s8qs_mmla_6x16 ( "cmp x27, #0x10\n" "blt 9f\n" "ldr q1, [x26, #0x0]\n" - "ldr q7, [x9, #0x0]\n" + "ldr q7, [x10, #0x0]\n" "cmp x27, #0x20\n" - "ldr q6, [x9, #0x10]\n" + "ldr q6, [x10, #0x10]\n" "blt 8f\n" "7:" // Height 1: Multiply loop: Main loop head + "sub x27, x27, #0x10\n" + "add x26, x26, #0x10\n" + "cmp x27, #0x20\n" + "prfm pldl1keep, [x26, #0x80]\n" "trn1 v18.2d, v1.2d, v21.2d\n" + "trn2 v1.2d, v1.2d, v21.2d\n" ".inst 0x4e87a648 // smmla v8.4s, v18.16b, v7.16b\n" - "ldr q17, [x9, #0x20]\n" + "ldr q17, [x10, #0x20]\n" ".inst 0x4e86a64c // smmla v12.4s, v18.16b, v6.16b\n" - "ldr q16, [x9, #0x30]\n" + "ldr q16, [x10, #0x30]\n" ".inst 0x4e91a649 // smmla v9.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x40]\n" + "ldr q17, [x10, #0x40]\n" ".inst 0x4e90a64d // smmla v13.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x50]\n" + "ldr q16, [x10, #0x50]\n" ".inst 0x4e91a64a // smmla v10.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x60]\n" + "ldr q17, [x10, #0x60]\n" ".inst 0x4e90a64e // smmla v14.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x70]\n" - "trn2 v1.2d, v1.2d, v21.2d\n" + "ldr q16, [x10, #0x70]\n" ".inst 0x4e91a64b // smmla v11.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x80]\n" + "ldr q17, [x10, #0x80]\n" ".inst 0x4e90a64f // smmla v15.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x90]\n" + "ldr q16, [x10, #0x90]\n" ".inst 0x4e91a428 // smmla v8.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xa0]\n" + "ldr q17, [x10, #0xa0]\n" ".inst 0x4e90a42c // smmla v12.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xb0]\n" + "ldr q16, [x10, #0xb0]\n" ".inst 0x4e91a429 // smmla v9.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xc0]\n" + "ldr q17, [x10, #0xc0]\n" ".inst 0x4e90a42d // smmla v13.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xd0]\n" + "ldr q16, [x10, #0xd0]\n" ".inst 0x4e91a42a // smmla v10.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xe0]\n" + "ldr q17, [x10, #0xe0]\n" ".inst 0x4e90a42e // smmla v14.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xf0]\n" - "sub x27, x27, #0x10\n" - "add x26, x26, #0x10\n" - "cmp x27, #0x20\n" + "ldr q16, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e91a42b // smmla v11.4s, v1.16b, v17.16b\n" + "ldr q7, [x10, #0x0]\n" ".inst 0x4e90a42f // smmla v15.4s, v1.16b, v16.16b\n" "ldr q1, [x26, #0x0]\n" - "add x9, x9, #0x100\n" - "ldr q7, [x9, #0x0]\n" - "ldr q6, [x9, #0x10]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "ldr q6, [x10, #0x10]\n" "bge 7b\n" "8:" // Height 1: Multiply loop: Single iteration only - "trn1 v18.2d, v1.2d, v19.2d\n" + "add x26, x26, #0x10\n" + "sub x27, x27, #0x10\n" + "prfm pldl1keep, [x26, #0x80]\n" + "trn1 v18.2d, v1.2d, v16.2d\n" + "trn2 v1.2d, v1.2d, v16.2d\n" ".inst 0x4e87a648 // smmla v8.4s, v18.16b, v7.16b\n" - "ldr q17, [x9, #0x20]\n" + "ldr q17, [x10, #0x20]\n" ".inst 0x4e86a64c // smmla v12.4s, v18.16b, v6.16b\n" - "ldr q16, [x9, #0x30]\n" + "ldr q16, [x10, #0x30]\n" ".inst 0x4e91a649 // smmla v9.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x40]\n" + "ldr q17, [x10, #0x40]\n" ".inst 0x4e90a64d // smmla v13.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x50]\n" + "ldr q16, [x10, #0x50]\n" ".inst 0x4e91a64a // smmla v10.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x60]\n" + "ldr q17, [x10, #0x60]\n" ".inst 0x4e90a64e // smmla v14.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x70]\n" - "trn2 v1.2d, v1.2d, v19.2d\n" + "ldr q16, [x10, #0x70]\n" ".inst 0x4e91a64b // smmla v11.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x80]\n" + "ldr q17, [x10, #0x80]\n" ".inst 0x4e90a64f // smmla v15.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x90]\n" + "ldr q16, [x10, #0x90]\n" ".inst 0x4e91a428 // smmla v8.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xa0]\n" + "ldr q17, [x10, #0xa0]\n" ".inst 0x4e90a42c // smmla v12.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xb0]\n" + "ldr q16, [x10, #0xb0]\n" ".inst 0x4e91a429 // smmla v9.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xc0]\n" + "ldr q17, [x10, #0xc0]\n" ".inst 0x4e90a42d // smmla v13.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xd0]\n" + "ldr q16, [x10, #0xd0]\n" ".inst 0x4e91a42a // smmla v10.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xe0]\n" + "ldr q17, [x10, #0xe0]\n" ".inst 0x4e90a42e // smmla v14.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xf0]\n" - "add x26, x26, #0x10\n" - "sub x27, x27, #0x10\n" + "ldr q16, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e91a42b // smmla v11.4s, v1.16b, v17.16b\n" ".inst 0x4e90a42f // smmla v15.4s, v1.16b, v16.16b\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x9, x9, #0x100\n" "9:" // Height 1: Multiply loop: Main loop skip "cbz x27, 16f\n" "cmp x27, #0x8\n" "blt 11f\n" "10:" // Height 1: Multiply loop: Odd block loop "ldr d18, [x26], #0x8\n" - "ldr q17, [x9, #0x0]\n" - "trn1 v18.2d, v18.2d, v16.2d\n" - "ldr q31, [x9, #0x10]\n" - ".inst 0x4e91a648 // smmla v8.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x20]\n" - ".inst 0x4e9fa64c // smmla v12.4s, v18.16b, v31.16b\n" - "ldr q16, [x9, #0x30]\n" + "ldr q19, [x10, #0x0]\n" + "sub x27, x27, #0x8\n" + "ldr q16, [x10, #0x10]\n" + "cmp x27, #0x8\n" + "trn1 v18.2d, v18.2d, v17.2d\n" + ".inst 0x4e93a648 // smmla v8.4s, v18.16b, v19.16b\n" + "ldr q17, [x10, #0x20]\n" + ".inst 0x4e90a64c // smmla v12.4s, v18.16b, v16.16b\n" + "ldr q16, [x10, #0x30]\n" ".inst 0x4e91a649 // smmla v9.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x40]\n" + "ldr q17, [x10, #0x40]\n" ".inst 0x4e90a64d // smmla v13.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x50]\n" + "ldr q16, [x10, #0x50]\n" ".inst 0x4e91a64a // smmla v10.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x60]\n" + "ldr q17, [x10, #0x60]\n" ".inst 0x4e90a64e // smmla v14.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x70]\n" - "sub x27, x27, #0x8\n" - "cmp x27, #0x8\n" + "ldr q16, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e91a64b // smmla v11.4s, v18.16b, v17.16b\n" ".inst 0x4e90a64f // smmla v15.4s, v18.16b, v16.16b\n" - "add x9, x9, #0x80\n" "bge 10b\n" "11:" // Height 1: Multiply loop: Skip odd blocks "cbz x27, 16f\n" @@ -261,24 +261,24 @@ void a64_hybrid_s8qs_mmla_6x16 ( "14:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 "ldr b1, [x26, #0x0]\n" "15:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q17, [x9, #0x0]\n" - "ldr q19, [x9, #0x10]\n" + "ldr q17, [x10, #0x0]\n" + "ldr q19, [x10, #0x10]\n" "trn1 v18.2d, v1.2d, v16.2d\n" ".inst 0x4e91a648 // smmla v8.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x20]\n" + "ldr q17, [x10, #0x20]\n" ".inst 0x4e93a64c // smmla v12.4s, v18.16b, v19.16b\n" - "ldr q16, [x9, #0x30]\n" + "ldr q16, [x10, #0x30]\n" ".inst 0x4e91a649 // smmla v9.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x40]\n" + "ldr q17, [x10, #0x40]\n" ".inst 0x4e90a64d // smmla v13.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x50]\n" + "ldr q16, [x10, #0x50]\n" ".inst 0x4e91a64a // smmla v10.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x60]\n" + "ldr q17, [x10, #0x60]\n" ".inst 0x4e90a64e // smmla v14.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x70]\n" + "ldr q16, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e91a64b // smmla v11.4s, v18.16b, v17.16b\n" ".inst 0x4e90a64f // smmla v15.4s, v18.16b, v16.16b\n" - "add x9, x9, #0x80\n" "16:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -292,11 +292,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr q16, [x14, #0x30]\n" "uzp1 v10.2d, v10.2d, v14.2d\n" "uzp1 v11.2d, v11.2d, v15.2d\n" - "mov v15.16b, v8.16b\n" - "prfm pstl1keep, [x11, #0x0]\n" - "add v15.4s, v15.4s, v19.4s\n" + "prfm pstl1keep, [x9, #0x0]\n" "add x14, x14, #0x40\n" + "mov v15.16b, v8.16b\n" "add v9.4s, v9.4s, v18.4s\n" + "add v15.4s, v15.4s, v19.4s\n" "add v10.4s, v10.4s, v17.4s\n" "add v11.4s, v11.4s, v16.4s\n" "tbz %x[flags], #4, 17f\n" @@ -312,9 +312,9 @@ void a64_hybrid_s8qs_mmla_6x16 ( "add x13, x13, #0x40\n" "b 18f\n" "17:" // Height 1: per layer parameters - "add x20, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x20]\n" + "add x21, %x[qp], %[per_layer_right_shift]\n" "add x20, %x[qp], %[per_layer_mul]\n" + "ld1r { v0.4s }, [x21]\n" "ld1r { v4.4s }, [x20]\n" "mov v1.16b, v0.16b\n" "mov v5.16b, v4.16b\n" @@ -328,34 +328,34 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sqrdmulh v10.4s, v10.4s, v6.4s\n" "sqrdmulh v11.4s, v11.4s, v7.4s\n" "tbz %x[flags], #5, 19f\n" - "and v17.16b, v15.16b, v0.16b\n" - "and v16.16b, v9.16b, v1.16b\n" - "and v25.16b, v10.16b, v2.16b\n" - "and v18.16b, v11.16b, v3.16b\n" + "and v18.16b, v15.16b, v0.16b\n" + "and v17.16b, v9.16b, v1.16b\n" + "and v26.16b, v10.16b, v2.16b\n" + "and v16.16b, v11.16b, v3.16b\n" + "sshr v18.4s, v18.4s, #0x1f\n" "sshr v17.4s, v17.4s, #0x1f\n" + "sshr v26.4s, v26.4s, #0x1f\n" "sshr v16.4s, v16.4s, #0x1f\n" - "sshr v25.4s, v25.4s, #0x1f\n" - "sshr v18.4s, v18.4s, #0x1f\n" - "sqadd v15.4s, v15.4s, v17.4s\n" - "sqadd v9.4s, v9.4s, v16.4s\n" - "sqadd v10.4s, v10.4s, v25.4s\n" - "sqadd v11.4s, v11.4s, v18.4s\n" + "sqadd v15.4s, v15.4s, v18.4s\n" + "sqadd v9.4s, v9.4s, v17.4s\n" + "sqadd v10.4s, v10.4s, v26.4s\n" + "sqadd v11.4s, v11.4s, v16.4s\n" "19:" // Height 1: no shift correction - "add x20, %x[qp], %[c_offset]\n" - "ld1r { v18.4s }, [x20]\n" + "add x21, %x[qp], %[c_offset]\n" "srshl v15.4s, v15.4s, v0.4s\n" "srshl v9.4s, v9.4s, v1.4s\n" - "srshl v10.4s, v10.4s, v2.4s\n" - "srshl v11.4s, v11.4s, v3.4s\n" "add x20, %x[qp], %[maxval]\n" + "ld1r { v18.4s }, [x21]\n" "ld1r { v17.4s }, [x20]\n" - "add v15.4s, v15.4s, v18.4s\n" - "add v9.4s, v9.4s, v18.4s\n" + "srshl v10.4s, v10.4s, v2.4s\n" + "srshl v11.4s, v11.4s, v3.4s\n" "add x20, %x[qp], %[minval]\n" + "cmp x11, #0x10\n" "ld1r { v16.4s }, [x20]\n" + "add v15.4s, v15.4s, v18.4s\n" + "add v9.4s, v9.4s, v18.4s\n" "add v10.4s, v10.4s, v18.4s\n" "add v11.4s, v11.4s, v18.4s\n" - "cmp x10, #0x10\n" "smin v15.4s, v15.4s, v17.4s\n" "smin v9.4s, v9.4s, v17.4s\n" "smin v10.4s, v10.4s, v17.4s\n" @@ -368,65 +368,65 @@ void a64_hybrid_s8qs_mmla_6x16 ( "uzp1 v16.8h, v10.8h, v11.8h\n" "uzp1 v15.16b, v15.16b, v16.16b\n" "bge 28f\n" - "tbz x10, #3, 23f\n" - "str d15, [x11], #0x8\n" - "tbz x10, #2, 21f\n" - "st1 { v15.s }[2], [x11], #0x4\n" - "tbz x10, #1, 20f\n" - "st1 { v15.h }[6], [x11], #0x2\n" - "tbz x10, #0, 27f\n" - "st1 { v15.b }[14], [x11]\n" + "tbz x11, #3, 23f\n" + "str d15, [x9], #0x8\n" + "tbz x11, #2, 21f\n" + "st1 { v15.s }[2], [x9], #0x4\n" + "tbz x11, #1, 20f\n" + "st1 { v15.h }[6], [x9], #0x2\n" + "tbz x11, #0, 27f\n" + "st1 { v15.b }[14], [x9]\n" "b 27f\n" "20:" // Height 1: Partial direct writeback: partial_1_12 - "tbz x10, #0, 27f\n" - "st1 { v15.b }[12], [x11]\n" + "tbz x11, #0, 27f\n" + "st1 { v15.b }[12], [x9]\n" "b 27f\n" "21:" // Height 1: Partial direct writeback: partial_2_8 - "tbz x10, #1, 22f\n" - "st1 { v15.h }[4], [x11], #0x2\n" - "tbz x10, #0, 27f\n" - "st1 { v15.b }[10], [x11]\n" + "tbz x11, #1, 22f\n" + "st1 { v15.h }[4], [x9], #0x2\n" + "tbz x11, #0, 27f\n" + "st1 { v15.b }[10], [x9]\n" "b 27f\n" "22:" // Height 1: Partial direct writeback: partial_1_8 - "tbz x10, #0, 27f\n" - "st1 { v15.b }[8], [x11]\n" + "tbz x11, #0, 27f\n" + "st1 { v15.b }[8], [x9]\n" "b 27f\n" "23:" // Height 1: Partial direct writeback: partial_4_0 - "tbz x10, #2, 25f\n" - "str s15, [x11], #0x4\n" - "tbz x10, #1, 24f\n" - "st1 { v15.h }[2], [x11], #0x2\n" - "tbz x10, #0, 27f\n" - "st1 { v15.b }[6], [x11]\n" + "tbz x11, #2, 25f\n" + "str s15, [x9], #0x4\n" + "tbz x11, #1, 24f\n" + "st1 { v15.h }[2], [x9], #0x2\n" + "tbz x11, #0, 27f\n" + "st1 { v15.b }[6], [x9]\n" "b 27f\n" "24:" // Height 1: Partial direct writeback: partial_1_4 - "tbz x10, #0, 27f\n" - "st1 { v15.b }[4], [x11]\n" + "tbz x11, #0, 27f\n" + "st1 { v15.b }[4], [x9]\n" "b 27f\n" "25:" // Height 1: Partial direct writeback: partial_2_0 - "tbz x10, #1, 26f\n" - "str h15, [x11], #0x2\n" - "tbz x10, #0, 27f\n" - "st1 { v15.b }[2], [x11]\n" + "tbz x11, #1, 26f\n" + "str h15, [x9], #0x2\n" + "tbz x11, #0, 27f\n" + "st1 { v15.b }[2], [x9]\n" "b 27f\n" "26:" // Height 1: Partial direct writeback: partial_1_0 - "str b15, [x11, #0x0]\n" + "str b15, [x9, #0x0]\n" "27:" // Height 1: Partial direct writeback: Done "b 29f\n" "28:" // Height 1: Full writeback - "str q15, [x11, #0x0]\n" - "add x11, x11, #0x10\n" + "str q15, [x9, #0x0]\n" + "add x9, x9, #0x10\n" "29:" // Height 1: Writeback done - "subs x10, x10, #0x10\n" + "subs x11, x11, #0x10\n" "bgt 2b\n" "b 176f\n" "30:" // Height 2 "mov x14, %x[col_bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x11, %x[output_ptr]\n" - "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "31:" // Height 2: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" @@ -440,8 +440,8 @@ void a64_hybrid_s8qs_mmla_6x16 ( "mov x28, #0x0\n" "33:" // 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, 34f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -461,120 +461,120 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr q1, [x26, #0x0]\n" "ldr q2, [x25, #0x0]\n" "cmp x27, #0x20\n" - "ldr q7, [x9, #0x0]\n" - "ldr q6, [x9, #0x10]\n" + "ldr q7, [x10, #0x0]\n" + "ldr q6, [x10, #0x10]\n" "blt 37f\n" "36:" // Height 2: Multiply loop: Main loop head "trn1 v18.2d, v1.2d, v2.2d\n" + "trn2 v1.2d, v1.2d, v2.2d\n" + "sub x27, x27, #0x10\n" + "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" + "cmp x27, #0x20\n" + "prfm pldl1keep, [x26, #0x80]\n" + "ldr q2, [x25, #0x0]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4e87a648 // smmla v8.4s, v18.16b, v7.16b\n" - "ldr q17, [x9, #0x20]\n" + "ldr q16, [x10, #0x20]\n" ".inst 0x4e86a64c // smmla v12.4s, v18.16b, v6.16b\n" - "ldr q16, [x9, #0x30]\n" - ".inst 0x4e91a649 // smmla v9.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x40]\n" - ".inst 0x4e90a64d // smmla v13.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x50]\n" - ".inst 0x4e91a64a // smmla v10.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x60]\n" - ".inst 0x4e90a64e // smmla v14.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x70]\n" - "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x4e91a64b // smmla v11.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x80]\n" - ".inst 0x4e90a64f // smmla v15.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x90]\n" + "ldr q17, [x10, #0x30]\n" + ".inst 0x4e90a649 // smmla v9.4s, v18.16b, v16.16b\n" + "ldr q16, [x10, #0x40]\n" + ".inst 0x4e91a64d // smmla v13.4s, v18.16b, v17.16b\n" + "ldr q17, [x10, #0x50]\n" + ".inst 0x4e90a64a // smmla v10.4s, v18.16b, v16.16b\n" + "ldr q16, [x10, #0x60]\n" + ".inst 0x4e91a64e // smmla v14.4s, v18.16b, v17.16b\n" + "ldr q31, [x10, #0x70]\n" + ".inst 0x4e90a64b // smmla v11.4s, v18.16b, v16.16b\n" + "ldr q17, [x10, #0x80]\n" + ".inst 0x4e9fa64f // smmla v15.4s, v18.16b, v31.16b\n" + "ldr q16, [x10, #0x90]\n" ".inst 0x4e91a428 // smmla v8.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xa0]\n" + "ldr q17, [x10, #0xa0]\n" ".inst 0x4e90a42c // smmla v12.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xb0]\n" + "ldr q16, [x10, #0xb0]\n" ".inst 0x4e91a429 // smmla v9.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xc0]\n" + "ldr q17, [x10, #0xc0]\n" ".inst 0x4e90a42d // smmla v13.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xd0]\n" + "ldr q16, [x10, #0xd0]\n" ".inst 0x4e91a42a // smmla v10.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xe0]\n" + "ldr q17, [x10, #0xe0]\n" ".inst 0x4e90a42e // smmla v14.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xf0]\n" - "sub x27, x27, #0x10\n" - "add x26, x26, #0x10\n" - "add x25, x25, #0x10\n" - "ldr q2, [x25, #0x0]\n" - "cmp x27, #0x20\n" + "ldr q16, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e91a42b // smmla v11.4s, v1.16b, v17.16b\n" - "add x9, x9, #0x100\n" - "ldr q7, [x9, #0x0]\n" + "ldr q7, [x10, #0x0]\n" ".inst 0x4e90a42f // smmla v15.4s, v1.16b, v16.16b\n" "ldr q1, [x26, #0x0]\n" - "ldr q6, [x9, #0x10]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "ldr q6, [x10, #0x10]\n" "bge 36b\n" "37:" // Height 2: Multiply loop: Single iteration only "trn1 v18.2d, v1.2d, v2.2d\n" + "trn2 v1.2d, v1.2d, v2.2d\n" + "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" + "sub x27, x27, #0x10\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4e87a648 // smmla v8.4s, v18.16b, v7.16b\n" - "ldr q17, [x9, #0x20]\n" + "ldr q17, [x10, #0x20]\n" ".inst 0x4e86a64c // smmla v12.4s, v18.16b, v6.16b\n" - "ldr q16, [x9, #0x30]\n" + "ldr q16, [x10, #0x30]\n" ".inst 0x4e91a649 // smmla v9.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x40]\n" + "ldr q17, [x10, #0x40]\n" ".inst 0x4e90a64d // smmla v13.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x50]\n" + "ldr q16, [x10, #0x50]\n" ".inst 0x4e91a64a // smmla v10.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x60]\n" + "ldr q17, [x10, #0x60]\n" ".inst 0x4e90a64e // smmla v14.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x70]\n" - "trn2 v1.2d, v1.2d, v2.2d\n" + "ldr q16, [x10, #0x70]\n" ".inst 0x4e91a64b // smmla v11.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x80]\n" + "ldr q17, [x10, #0x80]\n" ".inst 0x4e90a64f // smmla v15.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x90]\n" + "ldr q16, [x10, #0x90]\n" ".inst 0x4e91a428 // smmla v8.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xa0]\n" + "ldr q17, [x10, #0xa0]\n" ".inst 0x4e90a42c // smmla v12.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xb0]\n" + "ldr q16, [x10, #0xb0]\n" ".inst 0x4e91a429 // smmla v9.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xc0]\n" + "ldr q17, [x10, #0xc0]\n" ".inst 0x4e90a42d // smmla v13.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xd0]\n" + "ldr q16, [x10, #0xd0]\n" ".inst 0x4e91a42a // smmla v10.4s, v1.16b, v17.16b\n" - "ldr q17, [x9, #0xe0]\n" + "ldr q17, [x10, #0xe0]\n" ".inst 0x4e90a42e // smmla v14.4s, v1.16b, v16.16b\n" - "ldr q16, [x9, #0xf0]\n" - "add x26, x26, #0x10\n" - "add x25, x25, #0x10\n" + "ldr q16, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e91a42b // smmla v11.4s, v1.16b, v17.16b\n" ".inst 0x4e90a42f // smmla v15.4s, v1.16b, v16.16b\n" - "sub x27, x27, #0x10\n" - "prfm pldl1keep, [x26, #0x80]\n" - "prfm pldl1keep, [x25, #0x80]\n" - "add x9, x9, #0x100\n" "38:" // Height 2: Multiply loop: Main loop skip "cbz x27, 45f\n" "cmp x27, #0x8\n" "blt 40f\n" "39:" // Height 2: Multiply loop: Odd block loop - "ldr d17, [x26], #0x8\n" - "ldr d16, [x25], #0x8\n" - "trn1 v18.2d, v17.2d, v16.2d\n" + "ldr d19, [x26], #0x8\n" + "ldr d18, [x25], #0x8\n" "sub x27, x27, #0x8\n" - "ldr q17, [x9, #0x0]\n" - "ldr q16, [x9, #0x10]\n" + "ldr q17, [x10, #0x0]\n" + "ldr q16, [x10, #0x10]\n" + "cmp x27, #0x8\n" + "trn1 v18.2d, v19.2d, v18.2d\n" ".inst 0x4e91a648 // smmla v8.4s, v18.16b, v17.16b\n" + "ldr q17, [x10, #0x20]\n" ".inst 0x4e90a64c // smmla v12.4s, v18.16b, v16.16b\n" - "ldr q17, [x9, #0x20]\n" - "ldr q16, [x9, #0x30]\n" + "ldr q16, [x10, #0x30]\n" ".inst 0x4e91a649 // smmla v9.4s, v18.16b, v17.16b\n" + "ldr q17, [x10, #0x40]\n" ".inst 0x4e90a64d // smmla v13.4s, v18.16b, v16.16b\n" - "ldr q17, [x9, #0x40]\n" - "ldr q16, [x9, #0x50]\n" + "ldr q16, [x10, #0x50]\n" ".inst 0x4e91a64a // smmla v10.4s, v18.16b, v17.16b\n" + "ldr q17, [x10, #0x60]\n" ".inst 0x4e90a64e // smmla v14.4s, v18.16b, v16.16b\n" - "ldr q17, [x9, #0x60]\n" - "ldr q16, [x9, #0x70]\n" - "cmp x27, #0x8\n" + "ldr q16, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e91a64b // smmla v11.4s, v18.16b, v17.16b\n" ".inst 0x4e90a64f // smmla v15.4s, v18.16b, v16.16b\n" - "add x9, x9, #0x80\n" "bge 39b\n" "40:" // Height 2: Multiply loop: Skip odd blocks "cbz x27, 45f\n" @@ -605,24 +605,24 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr b1, [x26, #0x0]\n" "ldr b2, [x25, #0x0]\n" "44:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q17, [x9, #0x0]\n" - "ldr q16, [x9, #0x10]\n" + "ldr q17, [x10, #0x0]\n" + "ldr q16, [x10, #0x10]\n" "trn1 v18.2d, v1.2d, v2.2d\n" ".inst 0x4e91a648 // smmla v8.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x20]\n" + "ldr q17, [x10, #0x20]\n" ".inst 0x4e90a64c // smmla v12.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x30]\n" + "ldr q16, [x10, #0x30]\n" ".inst 0x4e91a649 // smmla v9.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x40]\n" + "ldr q17, [x10, #0x40]\n" ".inst 0x4e90a64d // smmla v13.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x50]\n" + "ldr q16, [x10, #0x50]\n" ".inst 0x4e91a64a // smmla v10.4s, v18.16b, v17.16b\n" - "ldr q17, [x9, #0x60]\n" + "ldr q17, [x10, #0x60]\n" ".inst 0x4e90a64e // smmla v14.4s, v18.16b, v16.16b\n" - "ldr q16, [x9, #0x70]\n" + "ldr q16, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e91a64b // smmla v11.4s, v18.16b, v17.16b\n" ".inst 0x4e90a64f // smmla v15.4s, v18.16b, v16.16b\n" - "add x9, x9, #0x80\n" "45:" // Height 2: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -639,18 +639,18 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "uzp1 v13.2d, v10.2d, v14.2d\n" "uzp2 v10.2d, v10.2d, v14.2d\n" - "add x25, x11, x20\n" + "prfm pstl1keep, [x9, #0x0]\n" "uzp1 v14.2d, v11.2d, v15.2d\n" "uzp2 v11.2d, v11.2d, v15.2d\n" - "prfm pstl1keep, [x11, #0x0]\n" - "prfm pstl1keep, [x25, #0x0]\n" - "mov v15.16b, v17.16b\n" - "add v15.4s, v15.4s, v19.4s\n" "add x14, x14, #0x40\n" + "mov v15.16b, v17.16b\n" "add v12.4s, v12.4s, v18.4s\n" + "add x26, x9, x20\n" + "prfm pstl1keep, [x26, #0x0]\n" "add v13.4s, v13.4s, v5.4s\n" - "add v14.4s, v14.4s, v16.4s\n" "add v8.4s, v8.4s, v19.4s\n" + "add v15.4s, v15.4s, v19.4s\n" + "add v14.4s, v14.4s, v16.4s\n" "add v9.4s, v9.4s, v18.4s\n" "add v10.4s, v10.4s, v5.4s\n" "add v11.4s, v11.4s, v16.4s\n" @@ -667,9 +667,9 @@ void a64_hybrid_s8qs_mmla_6x16 ( "add x13, x13, #0x40\n" "b 47f\n" "46:" // Height 2: per layer parameters - "add x20, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x20]\n" + "add x21, %x[qp], %[per_layer_right_shift]\n" "add x20, %x[qp], %[per_layer_mul]\n" + "ld1r { v0.4s }, [x21]\n" "ld1r { v4.4s }, [x20]\n" "mov v1.16b, v0.16b\n" "mov v5.16b, v4.16b\n" @@ -696,11 +696,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v17.4s, v17.4s, #0x1f\n" "sshr v16.4s, v16.4s, #0x1f\n" "sqadd v15.4s, v15.4s, v19.4s\n" + "and v19.16b, v8.16b, v0.16b\n" "sqadd v12.4s, v12.4s, v18.4s\n" + "and v18.16b, v9.16b, v1.16b\n" "sqadd v13.4s, v13.4s, v17.4s\n" "sqadd v14.4s, v14.4s, v16.4s\n" - "and v19.16b, v8.16b, v0.16b\n" - "and v18.16b, v9.16b, v1.16b\n" "and v17.16b, v10.16b, v2.16b\n" "and v16.16b, v11.16b, v3.16b\n" "sshr v19.4s, v19.4s, #0x1f\n" @@ -712,21 +712,21 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sqadd v10.4s, v10.4s, v17.4s\n" "sqadd v11.4s, v11.4s, v16.4s\n" "48:" // Height 2: no shift correction - "add x20, %x[qp], %[c_offset]\n" - "ld1r { v18.4s }, [x20]\n" + "add x21, %x[qp], %[c_offset]\n" "srshl v15.4s, v15.4s, v0.4s\n" "srshl v12.4s, v12.4s, v1.4s\n" - "srshl v13.4s, v13.4s, v2.4s\n" - "srshl v14.4s, v14.4s, v3.4s\n" "add x20, %x[qp], %[maxval]\n" + "ld1r { v18.4s }, [x21]\n" "ld1r { v17.4s }, [x20]\n" + "srshl v13.4s, v13.4s, v2.4s\n" + "srshl v14.4s, v14.4s, v3.4s\n" "srshl v8.4s, v8.4s, v0.4s\n" "srshl v9.4s, v9.4s, v1.4s\n" "add x20, %x[qp], %[minval]\n" + "cmp x11, #0x10\n" "ld1r { v16.4s }, [x20]\n" "srshl v10.4s, v10.4s, v2.4s\n" "srshl v11.4s, v11.4s, v3.4s\n" - "cmp x10, #0x10\n" "add v15.4s, v15.4s, v18.4s\n" "add v12.4s, v12.4s, v18.4s\n" "add v13.4s, v13.4s, v18.4s\n" @@ -758,81 +758,81 @@ void a64_hybrid_s8qs_mmla_6x16 ( "uzp1 v15.16b, v15.16b, v17.16b\n" "uzp1 v8.16b, v8.16b, v16.16b\n" "bge 57f\n" - "tbz x10, #3, 52f\n" - "str d15, [x11], #0x8\n" - "str d8, [x25], #0x8\n" - "tbz x10, #2, 50f\n" - "st1 { v15.s }[2], [x11], #0x4\n" - "st1 { v8.s }[2], [x25], #0x4\n" - "tbz x10, #1, 49f\n" - "st1 { v15.h }[6], [x11], #0x2\n" - "st1 { v8.h }[6], [x25], #0x2\n" - "tbz x10, #0, 56f\n" - "st1 { v15.b }[14], [x11]\n" - "st1 { v8.b }[14], [x25]\n" + "tbz x11, #3, 52f\n" + "str d15, [x9], #0x8\n" + "str d8, [x26], #0x8\n" + "tbz x11, #2, 50f\n" + "st1 { v15.s }[2], [x9], #0x4\n" + "st1 { v8.s }[2], [x26], #0x4\n" + "tbz x11, #1, 49f\n" + "st1 { v15.h }[6], [x9], #0x2\n" + "st1 { v8.h }[6], [x26], #0x2\n" + "tbz x11, #0, 56f\n" + "st1 { v15.b }[14], [x9]\n" + "st1 { v8.b }[14], [x26]\n" "b 56f\n" "49:" // Height 2: Partial direct writeback: partial_1_12 - "tbz x10, #0, 56f\n" - "st1 { v15.b }[12], [x11]\n" - "st1 { v8.b }[12], [x25]\n" + "tbz x11, #0, 56f\n" + "st1 { v15.b }[12], [x9]\n" + "st1 { v8.b }[12], [x26]\n" "b 56f\n" "50:" // Height 2: Partial direct writeback: partial_2_8 - "tbz x10, #1, 51f\n" - "st1 { v15.h }[4], [x11], #0x2\n" - "st1 { v8.h }[4], [x25], #0x2\n" - "tbz x10, #0, 56f\n" - "st1 { v15.b }[10], [x11]\n" - "st1 { v8.b }[10], [x25]\n" + "tbz x11, #1, 51f\n" + "st1 { v15.h }[4], [x9], #0x2\n" + "st1 { v8.h }[4], [x26], #0x2\n" + "tbz x11, #0, 56f\n" + "st1 { v15.b }[10], [x9]\n" + "st1 { v8.b }[10], [x26]\n" "b 56f\n" "51:" // Height 2: Partial direct writeback: partial_1_8 - "tbz x10, #0, 56f\n" - "st1 { v15.b }[8], [x11]\n" - "st1 { v8.b }[8], [x25]\n" + "tbz x11, #0, 56f\n" + "st1 { v15.b }[8], [x9]\n" + "st1 { v8.b }[8], [x26]\n" "b 56f\n" "52:" // Height 2: Partial direct writeback: partial_4_0 - "tbz x10, #2, 54f\n" - "str s15, [x11], #0x4\n" - "str s8, [x25], #0x4\n" - "tbz x10, #1, 53f\n" - "st1 { v15.h }[2], [x11], #0x2\n" - "st1 { v8.h }[2], [x25], #0x2\n" - "tbz x10, #0, 56f\n" - "st1 { v15.b }[6], [x11]\n" - "st1 { v8.b }[6], [x25]\n" + "tbz x11, #2, 54f\n" + "str s15, [x9], #0x4\n" + "str s8, [x26], #0x4\n" + "tbz x11, #1, 53f\n" + "st1 { v15.h }[2], [x9], #0x2\n" + "st1 { v8.h }[2], [x26], #0x2\n" + "tbz x11, #0, 56f\n" + "st1 { v15.b }[6], [x9]\n" + "st1 { v8.b }[6], [x26]\n" "b 56f\n" "53:" // Height 2: Partial direct writeback: partial_1_4 - "tbz x10, #0, 56f\n" - "st1 { v15.b }[4], [x11]\n" - "st1 { v8.b }[4], [x25]\n" + "tbz x11, #0, 56f\n" + "st1 { v15.b }[4], [x9]\n" + "st1 { v8.b }[4], [x26]\n" "b 56f\n" "54:" // Height 2: Partial direct writeback: partial_2_0 - "tbz x10, #1, 55f\n" - "str h15, [x11], #0x2\n" - "str h8, [x25], #0x2\n" - "tbz x10, #0, 56f\n" - "st1 { v15.b }[2], [x11]\n" - "st1 { v8.b }[2], [x25]\n" + "tbz x11, #1, 55f\n" + "str h15, [x9], #0x2\n" + "str h8, [x26], #0x2\n" + "tbz x11, #0, 56f\n" + "st1 { v15.b }[2], [x9]\n" + "st1 { v8.b }[2], [x26]\n" "b 56f\n" "55:" // Height 2: Partial direct writeback: partial_1_0 - "str b15, [x11, #0x0]\n" - "str b8, [x25, #0x0]\n" + "str b15, [x9, #0x0]\n" + "str b8, [x26, #0x0]\n" "56:" // Height 2: Partial direct writeback: Done "b 58f\n" "57:" // Height 2: Full writeback - "str q15, [x11, #0x0]\n" - "add x11, x11, #0x10\n" - "str q8, [x25, #0x0]\n" + "str q15, [x9, #0x0]\n" + "add x9, x9, #0x10\n" + "str q8, [x26, #0x0]\n" "58:" // Height 2: Writeback done - "subs x10, x10, #0x10\n" + "subs x11, x11, #0x10\n" "bgt 31b\n" "b 176f\n" "59:" // Height 3 "mov x14, %x[col_bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x11, %x[output_ptr]\n" - "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "60:" // Height 3: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" @@ -854,8 +854,8 @@ void a64_hybrid_s8qs_mmla_6x16 ( "mov x28, #0x0\n" "62:" // 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, 63f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -879,130 +879,130 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr q2, [x25, #0x0]\n" "cmp x27, #0x20\n" "ldr q3, [x24, #0x0]\n" - "ldr q7, [x9, #0x0]\n" - "ldr q6, [x9, #0x10]\n" + "ldr q7, [x10, #0x0]\n" + "ldr q6, [x10, #0x10]\n" "blt 66f\n" "65:" // Height 3: Multiply loop: Main loop head "trn1 v27.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" + "sub x27, x27, #0x10\n" + "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" + "add x24, x24, #0x10\n" + "cmp x27, #0x20\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "trn1 v26.2d, v3.2d, v24.2d\n" ".inst 0x4e87a768 // smmla v8.4s, v27.16b, v7.16b\n" - "trn1 v26.2d, v3.2d, v28.2d\n" - ".inst 0x4e87a750 // smmla v16.4s, v26.16b, v7.16b\n" - "ldr q25, [x9, #0x20]\n" ".inst 0x4e86a76c // smmla v12.4s, v27.16b, v6.16b\n" + "prfm pldl1keep, [x24, #0x80]\n" + "trn2 v3.2d, v3.2d, v24.2d\n" + ".inst 0x4e87a750 // smmla v16.4s, v26.16b, v7.16b\n" + "ldr q25, [x10, #0x20]\n" ".inst 0x4e86a754 // smmla v20.4s, v26.16b, v6.16b\n" - "ldr q24, [x9, #0x30]\n" + "ldr q24, [x10, #0x30]\n" ".inst 0x4e99a769 // smmla v9.4s, v27.16b, v25.16b\n" - "trn2 v3.2d, v3.2d, v28.2d\n" ".inst 0x4e99a751 // smmla v17.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x40]\n" + "ldr q25, [x10, #0x40]\n" ".inst 0x4e98a76d // smmla v13.4s, v27.16b, v24.16b\n" - "sub x27, x27, #0x10\n" ".inst 0x4e98a755 // smmla v21.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x50]\n" + "ldr q24, [x10, #0x50]\n" ".inst 0x4e99a76a // smmla v10.4s, v27.16b, v25.16b\n" - "add x26, x26, #0x10\n" ".inst 0x4e99a752 // smmla v18.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x60]\n" + "ldr q25, [x10, #0x60]\n" ".inst 0x4e98a76e // smmla v14.4s, v27.16b, v24.16b\n" - "add x25, x25, #0x10\n" ".inst 0x4e98a756 // smmla v22.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x70]\n" + "ldr q24, [x10, #0x70]\n" ".inst 0x4e99a76b // smmla v11.4s, v27.16b, v25.16b\n" - "add x24, x24, #0x10\n" ".inst 0x4e99a753 // smmla v19.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x80]\n" + "ldr q25, [x10, #0x80]\n" ".inst 0x4e98a76f // smmla v15.4s, v27.16b, v24.16b\n" - "cmp x27, #0x20\n" ".inst 0x4e98a757 // smmla v23.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x90]\n" + "ldr q24, [x10, #0x90]\n" "ldr q2, [x25, #0x0]\n" ".inst 0x4e99a428 // smmla v8.4s, v1.16b, v25.16b\n" ".inst 0x4e99a470 // smmla v16.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xa0]\n" + "ldr q25, [x10, #0xa0]\n" ".inst 0x4e98a42c // smmla v12.4s, v1.16b, v24.16b\n" - "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4e98a474 // smmla v20.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xb0]\n" + "ldr q24, [x10, #0xb0]\n" ".inst 0x4e99a429 // smmla v9.4s, v1.16b, v25.16b\n" - "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4e99a471 // smmla v17.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xc0]\n" + "ldr q25, [x10, #0xc0]\n" ".inst 0x4e98a42d // smmla v13.4s, v1.16b, v24.16b\n" - "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4e98a475 // smmla v21.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xd0]\n" + "ldr q24, [x10, #0xd0]\n" ".inst 0x4e99a42a // smmla v10.4s, v1.16b, v25.16b\n" ".inst 0x4e99a472 // smmla v18.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xe0]\n" + "ldr q25, [x10, #0xe0]\n" ".inst 0x4e98a42e // smmla v14.4s, v1.16b, v24.16b\n" ".inst 0x4e98a476 // smmla v22.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xf0]\n" - "add x9, x9, #0x100\n" + "ldr q24, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e99a42b // smmla v11.4s, v1.16b, v25.16b\n" ".inst 0x4e99a473 // smmla v19.4s, v3.16b, v25.16b\n" - "ldr q7, [x9, #0x0]\n" + "ldr q7, [x10, #0x0]\n" ".inst 0x4e98a42f // smmla v15.4s, v1.16b, v24.16b\n" "ldr q1, [x26, #0x0]\n" ".inst 0x4e98a477 // smmla v23.4s, v3.16b, v24.16b\n" "ldr q3, [x24, #0x0]\n" - "ldr q6, [x9, #0x10]\n" + "ldr q6, [x10, #0x10]\n" "bge 65b\n" "66:" // Height 3: Multiply loop: Single iteration only "trn1 v27.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" + "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" + "add x24, x24, #0x10\n" + "prfm pldl1keep, [x26, #0x80]\n" + "sub x27, x27, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "trn1 v26.2d, v3.2d, v24.2d\n" ".inst 0x4e87a768 // smmla v8.4s, v27.16b, v7.16b\n" - "trn1 v26.2d, v3.2d, v25.2d\n" - ".inst 0x4e87a750 // smmla v16.4s, v26.16b, v7.16b\n" - "ldr q24, [x9, #0x20]\n" ".inst 0x4e86a76c // smmla v12.4s, v27.16b, v6.16b\n" + "trn2 v3.2d, v3.2d, v24.2d\n" + ".inst 0x4e87a750 // smmla v16.4s, v26.16b, v7.16b\n" + "ldr q24, [x10, #0x20]\n" ".inst 0x4e86a754 // smmla v20.4s, v26.16b, v6.16b\n" - "ldr q0, [x9, #0x30]\n" + "ldr q4, [x10, #0x30]\n" ".inst 0x4e98a769 // smmla v9.4s, v27.16b, v24.16b\n" - "trn2 v3.2d, v3.2d, v25.2d\n" ".inst 0x4e98a751 // smmla v17.4s, v26.16b, v24.16b\n" - "ldr q25, [x9, #0x40]\n" - ".inst 0x4e80a76d // smmla v13.4s, v27.16b, v0.16b\n" - "add x26, x26, #0x10\n" - ".inst 0x4e80a755 // smmla v21.4s, v26.16b, v0.16b\n" - "ldr q24, [x9, #0x50]\n" + "ldr q25, [x10, #0x40]\n" + ".inst 0x4e84a76d // smmla v13.4s, v27.16b, v4.16b\n" + ".inst 0x4e84a755 // smmla v21.4s, v26.16b, v4.16b\n" + "ldr q24, [x10, #0x50]\n" ".inst 0x4e99a76a // smmla v10.4s, v27.16b, v25.16b\n" - "add x25, x25, #0x10\n" ".inst 0x4e99a752 // smmla v18.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x60]\n" + "ldr q25, [x10, #0x60]\n" ".inst 0x4e98a76e // smmla v14.4s, v27.16b, v24.16b\n" - "add x24, x24, #0x10\n" ".inst 0x4e98a756 // smmla v22.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x70]\n" + "ldr q24, [x10, #0x70]\n" ".inst 0x4e99a76b // smmla v11.4s, v27.16b, v25.16b\n" - "sub x27, x27, #0x10\n" ".inst 0x4e99a753 // smmla v19.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x80]\n" + "ldr q25, [x10, #0x80]\n" ".inst 0x4e98a76f // smmla v15.4s, v27.16b, v24.16b\n" - "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4e98a757 // smmla v23.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x90]\n" + "ldr q24, [x10, #0x90]\n" ".inst 0x4e99a428 // smmla v8.4s, v1.16b, v25.16b\n" - "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4e99a470 // smmla v16.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xa0]\n" + "ldr q25, [x10, #0xa0]\n" ".inst 0x4e98a42c // smmla v12.4s, v1.16b, v24.16b\n" - "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4e98a474 // smmla v20.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xb0]\n" + "ldr q24, [x10, #0xb0]\n" ".inst 0x4e99a429 // smmla v9.4s, v1.16b, v25.16b\n" ".inst 0x4e99a471 // smmla v17.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xc0]\n" + "ldr q25, [x10, #0xc0]\n" ".inst 0x4e98a42d // smmla v13.4s, v1.16b, v24.16b\n" ".inst 0x4e98a475 // smmla v21.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xd0]\n" + "ldr q24, [x10, #0xd0]\n" ".inst 0x4e99a42a // smmla v10.4s, v1.16b, v25.16b\n" ".inst 0x4e99a472 // smmla v18.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xe0]\n" + "ldr q25, [x10, #0xe0]\n" ".inst 0x4e98a42e // smmla v14.4s, v1.16b, v24.16b\n" ".inst 0x4e98a476 // smmla v22.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xf0]\n" - "add x9, x9, #0x100\n" + "ldr q24, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e99a42b // smmla v11.4s, v1.16b, v25.16b\n" ".inst 0x4e99a473 // smmla v19.4s, v3.16b, v25.16b\n" ".inst 0x4e98a42f // smmla v15.4s, v1.16b, v24.16b\n" @@ -1012,35 +1012,35 @@ void a64_hybrid_s8qs_mmla_6x16 ( "cmp x27, #0x8\n" "blt 69f\n" "68:" // Height 3: Multiply loop: Odd block loop - "ldr d25, [x26], #0x8\n" - "ldr d24, [x25], #0x8\n" - "trn1 v27.2d, v25.2d, v24.2d\n" - "ldr d24, [x24], #0x8\n" - "ldr q25, [x9, #0x0]\n" - "trn1 v26.2d, v24.2d, v26.2d\n" + "ldr d29, [x26], #0x8\n" + "ldr d27, [x25], #0x8\n" + "sub x27, x27, #0x8\n" + "ldr d26, [x24], #0x8\n" + "ldr q25, [x10, #0x0]\n" + "cmp x27, #0x8\n" + "ldr q24, [x10, #0x10]\n" + "trn1 v27.2d, v29.2d, v27.2d\n" + "trn1 v26.2d, v26.2d, v28.2d\n" ".inst 0x4e99a768 // smmla v8.4s, v27.16b, v25.16b\n" - "ldr q24, [x9, #0x10]\n" ".inst 0x4e99a750 // smmla v16.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x20]\n" + "ldr q25, [x10, #0x20]\n" ".inst 0x4e98a76c // smmla v12.4s, v27.16b, v24.16b\n" ".inst 0x4e98a754 // smmla v20.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x30]\n" + "ldr q24, [x10, #0x30]\n" ".inst 0x4e99a769 // smmla v9.4s, v27.16b, v25.16b\n" - "sub x27, x27, #0x8\n" ".inst 0x4e99a751 // smmla v17.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x40]\n" + "ldr q25, [x10, #0x40]\n" ".inst 0x4e98a76d // smmla v13.4s, v27.16b, v24.16b\n" - "cmp x27, #0x8\n" ".inst 0x4e98a755 // smmla v21.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x50]\n" + "ldr q24, [x10, #0x50]\n" ".inst 0x4e99a76a // smmla v10.4s, v27.16b, v25.16b\n" ".inst 0x4e99a752 // smmla v18.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x60]\n" + "ldr q25, [x10, #0x60]\n" ".inst 0x4e98a76e // smmla v14.4s, v27.16b, v24.16b\n" ".inst 0x4e98a756 // smmla v22.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x70]\n" + "ldr q24, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e99a76b // smmla v11.4s, v27.16b, v25.16b\n" - "add x9, x9, #0x80\n" ".inst 0x4e99a753 // smmla v19.4s, v26.16b, v25.16b\n" ".inst 0x4e98a76f // smmla v15.4s, v27.16b, v24.16b\n" ".inst 0x4e98a757 // smmla v23.4s, v26.16b, v24.16b\n" @@ -1081,29 +1081,29 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr b2, [x25, #0x0]\n" "ldr b3, [x24, #0x0]\n" "73:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q25, [x9, #0x0]\n" - "ldr q28, [x9, #0x10]\n" + "ldr q25, [x10, #0x0]\n" + "ldr q28, [x10, #0x10]\n" "trn1 v27.2d, v1.2d, v2.2d\n" "trn1 v26.2d, v3.2d, v24.2d\n" ".inst 0x4e99a768 // smmla v8.4s, v27.16b, v25.16b\n" - ".inst 0x4e99a750 // smmla v16.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x20]\n" ".inst 0x4e9ca76c // smmla v12.4s, v27.16b, v28.16b\n" + ".inst 0x4e99a750 // smmla v16.4s, v26.16b, v25.16b\n" + "ldr q25, [x10, #0x20]\n" ".inst 0x4e9ca754 // smmla v20.4s, v26.16b, v28.16b\n" - "ldr q24, [x9, #0x30]\n" + "ldr q24, [x10, #0x30]\n" ".inst 0x4e99a769 // smmla v9.4s, v27.16b, v25.16b\n" ".inst 0x4e99a751 // smmla v17.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x40]\n" + "ldr q25, [x10, #0x40]\n" ".inst 0x4e98a76d // smmla v13.4s, v27.16b, v24.16b\n" ".inst 0x4e98a755 // smmla v21.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x50]\n" + "ldr q24, [x10, #0x50]\n" ".inst 0x4e99a76a // smmla v10.4s, v27.16b, v25.16b\n" ".inst 0x4e99a752 // smmla v18.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x60]\n" + "ldr q25, [x10, #0x60]\n" ".inst 0x4e98a76e // smmla v14.4s, v27.16b, v24.16b\n" ".inst 0x4e98a756 // smmla v22.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x70]\n" - "add x9, x9, #0x80\n" + "ldr q24, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e99a76b // smmla v11.4s, v27.16b, v25.16b\n" ".inst 0x4e99a753 // smmla v19.4s, v26.16b, v25.16b\n" ".inst 0x4e98a76f // smmla v15.4s, v27.16b, v24.16b\n" @@ -1124,23 +1124,23 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "uzp1 v13.2d, v10.2d, v14.2d\n" "uzp2 v10.2d, v10.2d, v14.2d\n" - "add x25, x11, x20\n" + "prfm pstl1keep, [x9, #0x0]\n" "uzp1 v14.2d, v11.2d, v15.2d\n" "uzp2 v11.2d, v11.2d, v15.2d\n" - "add x24, x25, x20\n" - "prfm pstl1keep, [x11, #0x0]\n" + "add x14, x14, #0x40\n" "uzp1 v16.2d, v16.2d, v20.2d\n" "uzp1 v17.2d, v17.2d, v21.2d\n" - "prfm pstl1keep, [x25, #0x0]\n" - "prfm pstl1keep, [x24, #0x0]\n" "uzp1 v18.2d, v18.2d, v22.2d\n" "uzp1 v19.2d, v19.2d, v23.2d\n" - "add x14, x14, #0x40\n" + "add x26, x9, x20\n" "mov v23.16b, v26.16b\n" - "add v23.4s, v23.4s, v28.4s\n" + "add x25, x26, x20\n" + "prfm pstl1keep, [x26, #0x0]\n" "add v12.4s, v12.4s, v27.4s\n" + "prfm pstl1keep, [x25, #0x0]\n" "add v13.4s, v13.4s, v25.4s\n" "add v14.4s, v14.4s, v24.4s\n" + "add v23.4s, v23.4s, v28.4s\n" "add v8.4s, v8.4s, v28.4s\n" "add v9.4s, v9.4s, v27.4s\n" "add v10.4s, v10.4s, v25.4s\n" @@ -1162,9 +1162,9 @@ void a64_hybrid_s8qs_mmla_6x16 ( "add x13, x13, #0x40\n" "b 76f\n" "75:" // Height 3: per layer parameters - "add x20, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x20]\n" + "add x21, %x[qp], %[per_layer_right_shift]\n" "add x20, %x[qp], %[per_layer_mul]\n" + "ld1r { v0.4s }, [x21]\n" "ld1r { v4.4s }, [x20]\n" "mov v1.16b, v0.16b\n" "mov v5.16b, v4.16b\n" @@ -1195,11 +1195,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v21.4s, v21.4s, #0x1f\n" "sshr v20.4s, v20.4s, #0x1f\n" "sqadd v23.4s, v23.4s, v24.4s\n" + "and v24.16b, v8.16b, v0.16b\n" "sqadd v12.4s, v12.4s, v22.4s\n" + "and v22.16b, v9.16b, v1.16b\n" "sqadd v13.4s, v13.4s, v21.4s\n" "sqadd v14.4s, v14.4s, v20.4s\n" - "and v24.16b, v8.16b, v0.16b\n" - "and v22.16b, v9.16b, v1.16b\n" "and v21.16b, v10.16b, v2.16b\n" "and v20.16b, v11.16b, v3.16b\n" "sshr v24.4s, v24.4s, #0x1f\n" @@ -1207,11 +1207,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v21.4s, v21.4s, #0x1f\n" "sshr v20.4s, v20.4s, #0x1f\n" "sqadd v8.4s, v8.4s, v24.4s\n" + "and v24.16b, v16.16b, v0.16b\n" "sqadd v9.4s, v9.4s, v22.4s\n" + "and v22.16b, v17.16b, v1.16b\n" "sqadd v10.4s, v10.4s, v21.4s\n" "sqadd v11.4s, v11.4s, v20.4s\n" - "and v24.16b, v16.16b, v0.16b\n" - "and v22.16b, v17.16b, v1.16b\n" "and v21.16b, v18.16b, v2.16b\n" "and v20.16b, v19.16b, v3.16b\n" "sshr v24.4s, v24.4s, #0x1f\n" @@ -1223,21 +1223,21 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sqadd v18.4s, v18.4s, v21.4s\n" "sqadd v19.4s, v19.4s, v20.4s\n" "77:" // Height 3: no shift correction - "add x20, %x[qp], %[c_offset]\n" - "ld1r { v22.4s }, [x20]\n" + "add x21, %x[qp], %[c_offset]\n" "srshl v23.4s, v23.4s, v0.4s\n" "srshl v12.4s, v12.4s, v1.4s\n" - "srshl v13.4s, v13.4s, v2.4s\n" - "srshl v14.4s, v14.4s, v3.4s\n" "add x20, %x[qp], %[maxval]\n" + "ld1r { v22.4s }, [x21]\n" "ld1r { v21.4s }, [x20]\n" + "srshl v13.4s, v13.4s, v2.4s\n" + "srshl v14.4s, v14.4s, v3.4s\n" "srshl v8.4s, v8.4s, v0.4s\n" "srshl v9.4s, v9.4s, v1.4s\n" "add x20, %x[qp], %[minval]\n" + "cmp x11, #0x10\n" "ld1r { v20.4s }, [x20]\n" "srshl v10.4s, v10.4s, v2.4s\n" "srshl v11.4s, v11.4s, v3.4s\n" - "cmp x10, #0x10\n" "srshl v16.4s, v16.4s, v0.4s\n" "srshl v17.4s, v17.4s, v1.4s\n" "srshl v18.4s, v18.4s, v2.4s\n" @@ -1288,97 +1288,97 @@ void a64_hybrid_s8qs_mmla_6x16 ( "uzp1 v8.16b, v8.16b, v20.16b\n" "uzp1 v16.16b, v16.16b, v17.16b\n" "bge 86f\n" - "tbz x10, #3, 81f\n" - "str d23, [x11], #0x8\n" - "str d8, [x25], #0x8\n" - "str d16, [x24], #0x8\n" - "tbz x10, #2, 79f\n" - "st1 { v23.s }[2], [x11], #0x4\n" - "st1 { v8.s }[2], [x25], #0x4\n" - "st1 { v16.s }[2], [x24], #0x4\n" - "tbz x10, #1, 78f\n" - "st1 { v23.h }[6], [x11], #0x2\n" - "st1 { v8.h }[6], [x25], #0x2\n" - "st1 { v16.h }[6], [x24], #0x2\n" - "tbz x10, #0, 85f\n" - "st1 { v23.b }[14], [x11]\n" - "st1 { v8.b }[14], [x25]\n" - "st1 { v16.b }[14], [x24]\n" + "tbz x11, #3, 81f\n" + "str d23, [x9], #0x8\n" + "str d8, [x26], #0x8\n" + "str d16, [x25], #0x8\n" + "tbz x11, #2, 79f\n" + "st1 { v23.s }[2], [x9], #0x4\n" + "st1 { v8.s }[2], [x26], #0x4\n" + "st1 { v16.s }[2], [x25], #0x4\n" + "tbz x11, #1, 78f\n" + "st1 { v23.h }[6], [x9], #0x2\n" + "st1 { v8.h }[6], [x26], #0x2\n" + "st1 { v16.h }[6], [x25], #0x2\n" + "tbz x11, #0, 85f\n" + "st1 { v23.b }[14], [x9]\n" + "st1 { v8.b }[14], [x26]\n" + "st1 { v16.b }[14], [x25]\n" "b 85f\n" "78:" // Height 3: Partial direct writeback: partial_1_12 - "tbz x10, #0, 85f\n" - "st1 { v23.b }[12], [x11]\n" - "st1 { v8.b }[12], [x25]\n" - "st1 { v16.b }[12], [x24]\n" + "tbz x11, #0, 85f\n" + "st1 { v23.b }[12], [x9]\n" + "st1 { v8.b }[12], [x26]\n" + "st1 { v16.b }[12], [x25]\n" "b 85f\n" "79:" // Height 3: Partial direct writeback: partial_2_8 - "tbz x10, #1, 80f\n" - "st1 { v23.h }[4], [x11], #0x2\n" - "st1 { v8.h }[4], [x25], #0x2\n" - "st1 { v16.h }[4], [x24], #0x2\n" - "tbz x10, #0, 85f\n" - "st1 { v23.b }[10], [x11]\n" - "st1 { v8.b }[10], [x25]\n" - "st1 { v16.b }[10], [x24]\n" + "tbz x11, #1, 80f\n" + "st1 { v23.h }[4], [x9], #0x2\n" + "st1 { v8.h }[4], [x26], #0x2\n" + "st1 { v16.h }[4], [x25], #0x2\n" + "tbz x11, #0, 85f\n" + "st1 { v23.b }[10], [x9]\n" + "st1 { v8.b }[10], [x26]\n" + "st1 { v16.b }[10], [x25]\n" "b 85f\n" "80:" // Height 3: Partial direct writeback: partial_1_8 - "tbz x10, #0, 85f\n" - "st1 { v23.b }[8], [x11]\n" - "st1 { v8.b }[8], [x25]\n" - "st1 { v16.b }[8], [x24]\n" + "tbz x11, #0, 85f\n" + "st1 { v23.b }[8], [x9]\n" + "st1 { v8.b }[8], [x26]\n" + "st1 { v16.b }[8], [x25]\n" "b 85f\n" "81:" // Height 3: Partial direct writeback: partial_4_0 - "tbz x10, #2, 83f\n" - "str s23, [x11], #0x4\n" - "str s8, [x25], #0x4\n" - "str s16, [x24], #0x4\n" - "tbz x10, #1, 82f\n" - "st1 { v23.h }[2], [x11], #0x2\n" - "st1 { v8.h }[2], [x25], #0x2\n" - "st1 { v16.h }[2], [x24], #0x2\n" - "tbz x10, #0, 85f\n" - "st1 { v23.b }[6], [x11]\n" - "st1 { v8.b }[6], [x25]\n" - "st1 { v16.b }[6], [x24]\n" + "tbz x11, #2, 83f\n" + "str s23, [x9], #0x4\n" + "str s8, [x26], #0x4\n" + "str s16, [x25], #0x4\n" + "tbz x11, #1, 82f\n" + "st1 { v23.h }[2], [x9], #0x2\n" + "st1 { v8.h }[2], [x26], #0x2\n" + "st1 { v16.h }[2], [x25], #0x2\n" + "tbz x11, #0, 85f\n" + "st1 { v23.b }[6], [x9]\n" + "st1 { v8.b }[6], [x26]\n" + "st1 { v16.b }[6], [x25]\n" "b 85f\n" "82:" // Height 3: Partial direct writeback: partial_1_4 - "tbz x10, #0, 85f\n" - "st1 { v23.b }[4], [x11]\n" - "st1 { v8.b }[4], [x25]\n" - "st1 { v16.b }[4], [x24]\n" + "tbz x11, #0, 85f\n" + "st1 { v23.b }[4], [x9]\n" + "st1 { v8.b }[4], [x26]\n" + "st1 { v16.b }[4], [x25]\n" "b 85f\n" "83:" // Height 3: Partial direct writeback: partial_2_0 - "tbz x10, #1, 84f\n" - "str h23, [x11], #0x2\n" - "str h8, [x25], #0x2\n" - "str h16, [x24], #0x2\n" - "tbz x10, #0, 85f\n" - "st1 { v23.b }[2], [x11]\n" - "st1 { v8.b }[2], [x25]\n" - "st1 { v16.b }[2], [x24]\n" + "tbz x11, #1, 84f\n" + "str h23, [x9], #0x2\n" + "str h8, [x26], #0x2\n" + "str h16, [x25], #0x2\n" + "tbz x11, #0, 85f\n" + "st1 { v23.b }[2], [x9]\n" + "st1 { v8.b }[2], [x26]\n" + "st1 { v16.b }[2], [x25]\n" "b 85f\n" "84:" // Height 3: Partial direct writeback: partial_1_0 - "str b23, [x11, #0x0]\n" - "str b8, [x25, #0x0]\n" - "str b16, [x24, #0x0]\n" + "str b23, [x9, #0x0]\n" + "str b8, [x26, #0x0]\n" + "str b16, [x25, #0x0]\n" "85:" // Height 3: Partial direct writeback: Done "b 87f\n" "86:" // Height 3: Full writeback - "str q23, [x11, #0x0]\n" - "add x11, x11, #0x10\n" - "str q8, [x25, #0x0]\n" - "str q16, [x24, #0x0]\n" + "str q23, [x9, #0x0]\n" + "add x9, x9, #0x10\n" + "str q8, [x26, #0x0]\n" + "str q16, [x25, #0x0]\n" "87:" // Height 3: Writeback done - "subs x10, x10, #0x10\n" + "subs x11, x11, #0x10\n" "bgt 60b\n" "b 176f\n" "88:" // Height 4 "mov x14, %x[col_bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x11, %x[output_ptr]\n" - "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "89:" // Height 4: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" @@ -1400,8 +1400,8 @@ void a64_hybrid_s8qs_mmla_6x16 ( "mov x28, #0x0\n" "91:" // 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, 92f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1429,135 +1429,135 @@ void a64_hybrid_s8qs_mmla_6x16 ( "cmp x27, #0x20\n" "ldr q3, [x24, #0x0]\n" "ldr q4, [x23, #0x0]\n" - "ldr q7, [x9, #0x0]\n" - "ldr q6, [x9, #0x10]\n" + "ldr q7, [x10, #0x0]\n" + "ldr q6, [x10, #0x10]\n" "blt 95f\n" "94:" // Height 4: Multiply loop: Main loop head "trn1 v27.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x4e87a768 // smmla v8.4s, v27.16b, v7.16b\n" "sub x27, x27, #0x10\n" + "add x26, x26, #0x10\n" "trn1 v26.2d, v3.2d, v4.2d\n" - ".inst 0x4e87a750 // smmla v16.4s, v26.16b, v7.16b\n" - "ldr q25, [x9, #0x20]\n" + "trn2 v3.2d, v3.2d, v4.2d\n" + "add x25, x25, #0x10\n" + "add x24, x24, #0x10\n" + "add x23, x23, #0x10\n" + "cmp x27, #0x20\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "ldr q4, [x23, #0x0]\n" + ".inst 0x4e87a768 // smmla v8.4s, v27.16b, v7.16b\n" ".inst 0x4e86a76c // smmla v12.4s, v27.16b, v6.16b\n" + "prfm pldl1keep, [x24, #0x80]\n" + ".inst 0x4e87a750 // smmla v16.4s, v26.16b, v7.16b\n" + "ldr q25, [x10, #0x20]\n" ".inst 0x4e86a754 // smmla v20.4s, v26.16b, v6.16b\n" - "ldr q24, [x9, #0x30]\n" + "ldr q24, [x10, #0x30]\n" + "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4e99a769 // smmla v9.4s, v27.16b, v25.16b\n" - "trn2 v3.2d, v3.2d, v4.2d\n" ".inst 0x4e99a751 // smmla v17.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x40]\n" + "ldr q25, [x10, #0x40]\n" ".inst 0x4e98a76d // smmla v13.4s, v27.16b, v24.16b\n" - "add x26, x26, #0x10\n" ".inst 0x4e98a755 // smmla v21.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x50]\n" + "ldr q24, [x10, #0x50]\n" ".inst 0x4e99a76a // smmla v10.4s, v27.16b, v25.16b\n" - "add x25, x25, #0x10\n" ".inst 0x4e99a752 // smmla v18.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x60]\n" + "ldr q25, [x10, #0x60]\n" ".inst 0x4e98a76e // smmla v14.4s, v27.16b, v24.16b\n" - "add x24, x24, #0x10\n" ".inst 0x4e98a756 // smmla v22.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x70]\n" + "ldr q24, [x10, #0x70]\n" ".inst 0x4e99a76b // smmla v11.4s, v27.16b, v25.16b\n" - "add x23, x23, #0x10\n" - "ldr q4, [x23, #0x0]\n" ".inst 0x4e99a753 // smmla v19.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x80]\n" + "ldr q25, [x10, #0x80]\n" ".inst 0x4e98a76f // smmla v15.4s, v27.16b, v24.16b\n" ".inst 0x4e98a757 // smmla v23.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x90]\n" + "ldr q24, [x10, #0x90]\n" "ldr q2, [x25, #0x0]\n" ".inst 0x4e99a428 // smmla v8.4s, v1.16b, v25.16b\n" ".inst 0x4e99a470 // smmla v16.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xa0]\n" + "ldr q25, [x10, #0xa0]\n" ".inst 0x4e98a42c // smmla v12.4s, v1.16b, v24.16b\n" - "cmp x27, #0x20\n" ".inst 0x4e98a474 // smmla v20.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xb0]\n" + "ldr q24, [x10, #0xb0]\n" ".inst 0x4e99a429 // smmla v9.4s, v1.16b, v25.16b\n" - "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4e99a471 // smmla v17.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xc0]\n" + "ldr q25, [x10, #0xc0]\n" ".inst 0x4e98a42d // smmla v13.4s, v1.16b, v24.16b\n" - "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4e98a475 // smmla v21.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xd0]\n" + "ldr q24, [x10, #0xd0]\n" ".inst 0x4e99a42a // smmla v10.4s, v1.16b, v25.16b\n" - "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4e99a472 // smmla v18.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xe0]\n" + "ldr q25, [x10, #0xe0]\n" ".inst 0x4e98a42e // smmla v14.4s, v1.16b, v24.16b\n" - "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4e98a476 // smmla v22.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xf0]\n" - "add x9, x9, #0x100\n" + "ldr q24, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e99a42b // smmla v11.4s, v1.16b, v25.16b\n" ".inst 0x4e99a473 // smmla v19.4s, v3.16b, v25.16b\n" - "ldr q7, [x9, #0x0]\n" + "ldr q7, [x10, #0x0]\n" ".inst 0x4e98a42f // smmla v15.4s, v1.16b, v24.16b\n" "ldr q1, [x26, #0x0]\n" ".inst 0x4e98a477 // smmla v23.4s, v3.16b, v24.16b\n" "ldr q3, [x24, #0x0]\n" - "ldr q6, [x9, #0x10]\n" + "ldr q6, [x10, #0x10]\n" "bge 94b\n" "95:" // Height 4: Multiply loop: Single iteration only "trn1 v27.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x4e87a768 // smmla v8.4s, v27.16b, v7.16b\n" "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" "trn1 v26.2d, v3.2d, v4.2d\n" - ".inst 0x4e87a750 // smmla v16.4s, v26.16b, v7.16b\n" - "ldr q25, [x9, #0x20]\n" + "trn2 v3.2d, v3.2d, v4.2d\n" + "add x24, x24, #0x10\n" + "add x23, x23, #0x10\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "sub x27, x27, #0x10\n" + ".inst 0x4e87a768 // smmla v8.4s, v27.16b, v7.16b\n" ".inst 0x4e86a76c // smmla v12.4s, v27.16b, v6.16b\n" + "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" + ".inst 0x4e87a750 // smmla v16.4s, v26.16b, v7.16b\n" + "ldr q25, [x10, #0x20]\n" ".inst 0x4e86a754 // smmla v20.4s, v26.16b, v6.16b\n" - "ldr q24, [x9, #0x30]\n" + "ldr q24, [x10, #0x30]\n" ".inst 0x4e99a769 // smmla v9.4s, v27.16b, v25.16b\n" - "trn2 v3.2d, v3.2d, v4.2d\n" ".inst 0x4e99a751 // smmla v17.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x40]\n" + "ldr q25, [x10, #0x40]\n" ".inst 0x4e98a76d // smmla v13.4s, v27.16b, v24.16b\n" - "add x25, x25, #0x10\n" ".inst 0x4e98a755 // smmla v21.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x50]\n" + "ldr q24, [x10, #0x50]\n" ".inst 0x4e99a76a // smmla v10.4s, v27.16b, v25.16b\n" - "add x24, x24, #0x10\n" ".inst 0x4e99a752 // smmla v18.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x60]\n" + "ldr q25, [x10, #0x60]\n" ".inst 0x4e98a76e // smmla v14.4s, v27.16b, v24.16b\n" - "add x23, x23, #0x10\n" ".inst 0x4e98a756 // smmla v22.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x70]\n" + "ldr q24, [x10, #0x70]\n" ".inst 0x4e99a76b // smmla v11.4s, v27.16b, v25.16b\n" - "sub x27, x27, #0x10\n" ".inst 0x4e99a753 // smmla v19.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x80]\n" + "ldr q25, [x10, #0x80]\n" ".inst 0x4e98a76f // smmla v15.4s, v27.16b, v24.16b\n" - "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4e98a757 // smmla v23.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x90]\n" + "ldr q24, [x10, #0x90]\n" ".inst 0x4e99a428 // smmla v8.4s, v1.16b, v25.16b\n" - "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4e99a470 // smmla v16.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xa0]\n" + "ldr q25, [x10, #0xa0]\n" ".inst 0x4e98a42c // smmla v12.4s, v1.16b, v24.16b\n" - "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4e98a474 // smmla v20.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xb0]\n" + "ldr q24, [x10, #0xb0]\n" ".inst 0x4e99a429 // smmla v9.4s, v1.16b, v25.16b\n" - "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4e99a471 // smmla v17.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xc0]\n" + "ldr q25, [x10, #0xc0]\n" ".inst 0x4e98a42d // smmla v13.4s, v1.16b, v24.16b\n" ".inst 0x4e98a475 // smmla v21.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xd0]\n" + "ldr q24, [x10, #0xd0]\n" ".inst 0x4e99a42a // smmla v10.4s, v1.16b, v25.16b\n" ".inst 0x4e99a472 // smmla v18.4s, v3.16b, v25.16b\n" - "ldr q25, [x9, #0xe0]\n" + "ldr q25, [x10, #0xe0]\n" ".inst 0x4e98a42e // smmla v14.4s, v1.16b, v24.16b\n" ".inst 0x4e98a476 // smmla v22.4s, v3.16b, v24.16b\n" - "ldr q24, [x9, #0xf0]\n" - "add x9, x9, #0x100\n" + "ldr q24, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e99a42b // smmla v11.4s, v1.16b, v25.16b\n" ".inst 0x4e99a473 // smmla v19.4s, v3.16b, v25.16b\n" ".inst 0x4e98a42f // smmla v15.4s, v1.16b, v24.16b\n" @@ -1567,35 +1567,35 @@ void a64_hybrid_s8qs_mmla_6x16 ( "cmp x27, #0x8\n" "blt 98f\n" "97:" // Height 4: Multiply loop: Odd block loop - "ldr d25, [x26], #0x8\n" - "ldr d24, [x25], #0x8\n" - "trn1 v27.2d, v25.2d, v24.2d\n" + "ldr d29, [x26], #0x8\n" + "ldr d27, [x25], #0x8\n" "sub x27, x27, #0x8\n" - "ldr d25, [x24], #0x8\n" - "ldr d24, [x23], #0x8\n" - "trn1 v26.2d, v25.2d, v24.2d\n" + "ldr d28, [x24], #0x8\n" + "ldr d26, [x23], #0x8\n" "cmp x27, #0x8\n" - "ldr q25, [x9, #0x0]\n" - "ldr q24, [x9, #0x10]\n" + "ldr q25, [x10, #0x0]\n" + "ldr q24, [x10, #0x10]\n" + "trn1 v27.2d, v29.2d, v27.2d\n" + "trn1 v26.2d, v28.2d, v26.2d\n" ".inst 0x4e99a768 // smmla v8.4s, v27.16b, v25.16b\n" ".inst 0x4e99a750 // smmla v16.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x20]\n" + "ldr q25, [x10, #0x20]\n" ".inst 0x4e98a76c // smmla v12.4s, v27.16b, v24.16b\n" ".inst 0x4e98a754 // smmla v20.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x30]\n" + "ldr q24, [x10, #0x30]\n" ".inst 0x4e99a769 // smmla v9.4s, v27.16b, v25.16b\n" ".inst 0x4e99a751 // smmla v17.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x40]\n" + "ldr q25, [x10, #0x40]\n" ".inst 0x4e98a76d // smmla v13.4s, v27.16b, v24.16b\n" ".inst 0x4e98a755 // smmla v21.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x50]\n" + "ldr q24, [x10, #0x50]\n" ".inst 0x4e99a76a // smmla v10.4s, v27.16b, v25.16b\n" ".inst 0x4e99a752 // smmla v18.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x60]\n" + "ldr q25, [x10, #0x60]\n" ".inst 0x4e98a76e // smmla v14.4s, v27.16b, v24.16b\n" ".inst 0x4e98a756 // smmla v22.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x70]\n" - "add x9, x9, #0x80\n" + "ldr q24, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e99a76b // smmla v11.4s, v27.16b, v25.16b\n" ".inst 0x4e99a753 // smmla v19.4s, v26.16b, v25.16b\n" ".inst 0x4e98a76f // smmla v15.4s, v27.16b, v24.16b\n" @@ -1644,29 +1644,29 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr b3, [x24, #0x0]\n" "ldr b4, [x23, #0x0]\n" "102:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q25, [x9, #0x0]\n" - "ldr q24, [x9, #0x10]\n" + "ldr q25, [x10, #0x0]\n" + "ldr q24, [x10, #0x10]\n" "trn1 v27.2d, v1.2d, v2.2d\n" "trn1 v26.2d, v3.2d, v4.2d\n" ".inst 0x4e99a768 // smmla v8.4s, v27.16b, v25.16b\n" ".inst 0x4e99a750 // smmla v16.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x20]\n" + "ldr q25, [x10, #0x20]\n" ".inst 0x4e98a76c // smmla v12.4s, v27.16b, v24.16b\n" ".inst 0x4e98a754 // smmla v20.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x30]\n" + "ldr q24, [x10, #0x30]\n" ".inst 0x4e99a769 // smmla v9.4s, v27.16b, v25.16b\n" ".inst 0x4e99a751 // smmla v17.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x40]\n" + "ldr q25, [x10, #0x40]\n" ".inst 0x4e98a76d // smmla v13.4s, v27.16b, v24.16b\n" ".inst 0x4e98a755 // smmla v21.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x50]\n" + "ldr q24, [x10, #0x50]\n" ".inst 0x4e99a76a // smmla v10.4s, v27.16b, v25.16b\n" ".inst 0x4e99a752 // smmla v18.4s, v26.16b, v25.16b\n" - "ldr q25, [x9, #0x60]\n" + "ldr q25, [x10, #0x60]\n" ".inst 0x4e98a76e // smmla v14.4s, v27.16b, v24.16b\n" ".inst 0x4e98a756 // smmla v22.4s, v26.16b, v24.16b\n" - "ldr q24, [x9, #0x70]\n" - "add x9, x9, #0x80\n" + "ldr q24, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e99a76b // smmla v11.4s, v27.16b, v25.16b\n" ".inst 0x4e99a753 // smmla v19.4s, v26.16b, v25.16b\n" ".inst 0x4e98a76f // smmla v15.4s, v27.16b, v24.16b\n" @@ -1687,29 +1687,29 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "uzp1 v13.2d, v10.2d, v14.2d\n" "uzp2 v10.2d, v10.2d, v14.2d\n" - "add x25, x11, x20\n" + "prfm pstl1keep, [x9, #0x0]\n" "uzp1 v14.2d, v11.2d, v15.2d\n" "uzp2 v11.2d, v11.2d, v15.2d\n" - "add x24, x25, x20\n" - "add x23, x24, x20\n" + "add x14, x14, #0x40\n" "uzp1 v15.2d, v16.2d, v20.2d\n" "uzp2 v16.2d, v16.2d, v20.2d\n" - "prfm pstl1keep, [x11, #0x0]\n" - "prfm pstl1keep, [x25, #0x0]\n" "uzp1 v20.2d, v17.2d, v21.2d\n" "uzp2 v17.2d, v17.2d, v21.2d\n" - "prfm pstl1keep, [x24, #0x0]\n" - "prfm pstl1keep, [x23, #0x0]\n" + "add x26, x9, x20\n" "uzp1 v21.2d, v18.2d, v22.2d\n" "uzp2 v18.2d, v18.2d, v22.2d\n" - "add x14, x14, #0x40\n" + "add x25, x26, x20\n" + "prfm pstl1keep, [x26, #0x0]\n" "uzp1 v22.2d, v19.2d, v23.2d\n" "uzp2 v19.2d, v19.2d, v23.2d\n" + "add x24, x25, x20\n" + "prfm pstl1keep, [x25, #0x0]\n" "mov v23.16b, v26.16b\n" - "add v23.4s, v23.4s, v28.4s\n" + "prfm pstl1keep, [x24, #0x0]\n" "add v12.4s, v12.4s, v27.4s\n" "add v13.4s, v13.4s, v25.4s\n" "add v14.4s, v14.4s, v24.4s\n" + "add v23.4s, v23.4s, v28.4s\n" "add v8.4s, v8.4s, v28.4s\n" "add v9.4s, v9.4s, v27.4s\n" "add v10.4s, v10.4s, v25.4s\n" @@ -1735,9 +1735,9 @@ void a64_hybrid_s8qs_mmla_6x16 ( "add x13, x13, #0x40\n" "b 105f\n" "104:" // Height 4: per layer parameters - "add x20, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x20]\n" + "add x21, %x[qp], %[per_layer_right_shift]\n" "add x20, %x[qp], %[per_layer_mul]\n" + "ld1r { v0.4s }, [x21]\n" "ld1r { v4.4s }, [x20]\n" "mov v1.16b, v0.16b\n" "mov v5.16b, v4.16b\n" @@ -1772,11 +1772,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v25.4s, v25.4s, #0x1f\n" "sshr v24.4s, v24.4s, #0x1f\n" "sqadd v23.4s, v23.4s, v27.4s\n" + "and v27.16b, v8.16b, v0.16b\n" "sqadd v12.4s, v12.4s, v26.4s\n" + "and v26.16b, v9.16b, v1.16b\n" "sqadd v13.4s, v13.4s, v25.4s\n" "sqadd v14.4s, v14.4s, v24.4s\n" - "and v27.16b, v8.16b, v0.16b\n" - "and v26.16b, v9.16b, v1.16b\n" "and v25.16b, v10.16b, v2.16b\n" "and v24.16b, v11.16b, v3.16b\n" "sshr v27.4s, v27.4s, #0x1f\n" @@ -1784,11 +1784,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v25.4s, v25.4s, #0x1f\n" "sshr v24.4s, v24.4s, #0x1f\n" "sqadd v8.4s, v8.4s, v27.4s\n" + "and v27.16b, v15.16b, v0.16b\n" "sqadd v9.4s, v9.4s, v26.4s\n" + "and v26.16b, v20.16b, v1.16b\n" "sqadd v10.4s, v10.4s, v25.4s\n" "sqadd v11.4s, v11.4s, v24.4s\n" - "and v27.16b, v15.16b, v0.16b\n" - "and v26.16b, v20.16b, v1.16b\n" "and v25.16b, v21.16b, v2.16b\n" "and v24.16b, v22.16b, v3.16b\n" "sshr v27.4s, v27.4s, #0x1f\n" @@ -1796,11 +1796,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v25.4s, v25.4s, #0x1f\n" "sshr v24.4s, v24.4s, #0x1f\n" "sqadd v15.4s, v15.4s, v27.4s\n" + "and v27.16b, v16.16b, v0.16b\n" "sqadd v20.4s, v20.4s, v26.4s\n" + "and v26.16b, v17.16b, v1.16b\n" "sqadd v21.4s, v21.4s, v25.4s\n" "sqadd v22.4s, v22.4s, v24.4s\n" - "and v27.16b, v16.16b, v0.16b\n" - "and v26.16b, v17.16b, v1.16b\n" "and v25.16b, v18.16b, v2.16b\n" "and v24.16b, v19.16b, v3.16b\n" "sshr v27.4s, v27.4s, #0x1f\n" @@ -1812,21 +1812,21 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sqadd v18.4s, v18.4s, v25.4s\n" "sqadd v19.4s, v19.4s, v24.4s\n" "106:" // Height 4: no shift correction - "add x20, %x[qp], %[c_offset]\n" - "ld1r { v26.4s }, [x20]\n" + "add x21, %x[qp], %[c_offset]\n" "srshl v23.4s, v23.4s, v0.4s\n" "srshl v12.4s, v12.4s, v1.4s\n" - "srshl v13.4s, v13.4s, v2.4s\n" - "srshl v14.4s, v14.4s, v3.4s\n" "add x20, %x[qp], %[maxval]\n" + "ld1r { v26.4s }, [x21]\n" "ld1r { v25.4s }, [x20]\n" + "srshl v13.4s, v13.4s, v2.4s\n" + "srshl v14.4s, v14.4s, v3.4s\n" "srshl v8.4s, v8.4s, v0.4s\n" "srshl v9.4s, v9.4s, v1.4s\n" "add x20, %x[qp], %[minval]\n" + "cmp x11, #0x10\n" "ld1r { v24.4s }, [x20]\n" "srshl v10.4s, v10.4s, v2.4s\n" "srshl v11.4s, v11.4s, v3.4s\n" - "cmp x10, #0x10\n" "srshl v15.4s, v15.4s, v0.4s\n" "srshl v20.4s, v20.4s, v1.4s\n" "srshl v21.4s, v21.4s, v2.4s\n" @@ -1896,113 +1896,113 @@ void a64_hybrid_s8qs_mmla_6x16 ( "uzp1 v15.16b, v15.16b, v20.16b\n" "uzp1 v16.16b, v16.16b, v17.16b\n" "bge 115f\n" - "tbz x10, #3, 110f\n" - "str d23, [x11], #0x8\n" - "str d8, [x25], #0x8\n" - "str d15, [x24], #0x8\n" - "str d16, [x23], #0x8\n" - "tbz x10, #2, 108f\n" - "st1 { v23.s }[2], [x11], #0x4\n" - "st1 { v8.s }[2], [x25], #0x4\n" - "st1 { v15.s }[2], [x24], #0x4\n" - "st1 { v16.s }[2], [x23], #0x4\n" - "tbz x10, #1, 107f\n" - "st1 { v23.h }[6], [x11], #0x2\n" - "st1 { v8.h }[6], [x25], #0x2\n" - "st1 { v15.h }[6], [x24], #0x2\n" - "st1 { v16.h }[6], [x23], #0x2\n" - "tbz x10, #0, 114f\n" - "st1 { v23.b }[14], [x11]\n" - "st1 { v8.b }[14], [x25]\n" - "st1 { v15.b }[14], [x24]\n" - "st1 { v16.b }[14], [x23]\n" + "tbz x11, #3, 110f\n" + "str d23, [x9], #0x8\n" + "str d8, [x26], #0x8\n" + "str d15, [x25], #0x8\n" + "str d16, [x24], #0x8\n" + "tbz x11, #2, 108f\n" + "st1 { v23.s }[2], [x9], #0x4\n" + "st1 { v8.s }[2], [x26], #0x4\n" + "st1 { v15.s }[2], [x25], #0x4\n" + "st1 { v16.s }[2], [x24], #0x4\n" + "tbz x11, #1, 107f\n" + "st1 { v23.h }[6], [x9], #0x2\n" + "st1 { v8.h }[6], [x26], #0x2\n" + "st1 { v15.h }[6], [x25], #0x2\n" + "st1 { v16.h }[6], [x24], #0x2\n" + "tbz x11, #0, 114f\n" + "st1 { v23.b }[14], [x9]\n" + "st1 { v8.b }[14], [x26]\n" + "st1 { v15.b }[14], [x25]\n" + "st1 { v16.b }[14], [x24]\n" "b 114f\n" "107:" // Height 4: Partial direct writeback: partial_1_12 - "tbz x10, #0, 114f\n" - "st1 { v23.b }[12], [x11]\n" - "st1 { v8.b }[12], [x25]\n" - "st1 { v15.b }[12], [x24]\n" - "st1 { v16.b }[12], [x23]\n" + "tbz x11, #0, 114f\n" + "st1 { v23.b }[12], [x9]\n" + "st1 { v8.b }[12], [x26]\n" + "st1 { v15.b }[12], [x25]\n" + "st1 { v16.b }[12], [x24]\n" "b 114f\n" "108:" // Height 4: Partial direct writeback: partial_2_8 - "tbz x10, #1, 109f\n" - "st1 { v23.h }[4], [x11], #0x2\n" - "st1 { v8.h }[4], [x25], #0x2\n" - "st1 { v15.h }[4], [x24], #0x2\n" - "st1 { v16.h }[4], [x23], #0x2\n" - "tbz x10, #0, 114f\n" - "st1 { v23.b }[10], [x11]\n" - "st1 { v8.b }[10], [x25]\n" - "st1 { v15.b }[10], [x24]\n" - "st1 { v16.b }[10], [x23]\n" + "tbz x11, #1, 109f\n" + "st1 { v23.h }[4], [x9], #0x2\n" + "st1 { v8.h }[4], [x26], #0x2\n" + "st1 { v15.h }[4], [x25], #0x2\n" + "st1 { v16.h }[4], [x24], #0x2\n" + "tbz x11, #0, 114f\n" + "st1 { v23.b }[10], [x9]\n" + "st1 { v8.b }[10], [x26]\n" + "st1 { v15.b }[10], [x25]\n" + "st1 { v16.b }[10], [x24]\n" "b 114f\n" "109:" // Height 4: Partial direct writeback: partial_1_8 - "tbz x10, #0, 114f\n" - "st1 { v23.b }[8], [x11]\n" - "st1 { v8.b }[8], [x25]\n" - "st1 { v15.b }[8], [x24]\n" - "st1 { v16.b }[8], [x23]\n" + "tbz x11, #0, 114f\n" + "st1 { v23.b }[8], [x9]\n" + "st1 { v8.b }[8], [x26]\n" + "st1 { v15.b }[8], [x25]\n" + "st1 { v16.b }[8], [x24]\n" "b 114f\n" "110:" // Height 4: Partial direct writeback: partial_4_0 - "tbz x10, #2, 112f\n" - "str s23, [x11], #0x4\n" - "str s8, [x25], #0x4\n" - "str s15, [x24], #0x4\n" - "str s16, [x23], #0x4\n" - "tbz x10, #1, 111f\n" - "st1 { v23.h }[2], [x11], #0x2\n" - "st1 { v8.h }[2], [x25], #0x2\n" - "st1 { v15.h }[2], [x24], #0x2\n" - "st1 { v16.h }[2], [x23], #0x2\n" - "tbz x10, #0, 114f\n" - "st1 { v23.b }[6], [x11]\n" - "st1 { v8.b }[6], [x25]\n" - "st1 { v15.b }[6], [x24]\n" - "st1 { v16.b }[6], [x23]\n" + "tbz x11, #2, 112f\n" + "str s23, [x9], #0x4\n" + "str s8, [x26], #0x4\n" + "str s15, [x25], #0x4\n" + "str s16, [x24], #0x4\n" + "tbz x11, #1, 111f\n" + "st1 { v23.h }[2], [x9], #0x2\n" + "st1 { v8.h }[2], [x26], #0x2\n" + "st1 { v15.h }[2], [x25], #0x2\n" + "st1 { v16.h }[2], [x24], #0x2\n" + "tbz x11, #0, 114f\n" + "st1 { v23.b }[6], [x9]\n" + "st1 { v8.b }[6], [x26]\n" + "st1 { v15.b }[6], [x25]\n" + "st1 { v16.b }[6], [x24]\n" "b 114f\n" "111:" // Height 4: Partial direct writeback: partial_1_4 - "tbz x10, #0, 114f\n" - "st1 { v23.b }[4], [x11]\n" - "st1 { v8.b }[4], [x25]\n" - "st1 { v15.b }[4], [x24]\n" - "st1 { v16.b }[4], [x23]\n" + "tbz x11, #0, 114f\n" + "st1 { v23.b }[4], [x9]\n" + "st1 { v8.b }[4], [x26]\n" + "st1 { v15.b }[4], [x25]\n" + "st1 { v16.b }[4], [x24]\n" "b 114f\n" "112:" // Height 4: Partial direct writeback: partial_2_0 - "tbz x10, #1, 113f\n" - "str h23, [x11], #0x2\n" - "str h8, [x25], #0x2\n" - "str h15, [x24], #0x2\n" - "str h16, [x23], #0x2\n" - "tbz x10, #0, 114f\n" - "st1 { v23.b }[2], [x11]\n" - "st1 { v8.b }[2], [x25]\n" - "st1 { v15.b }[2], [x24]\n" - "st1 { v16.b }[2], [x23]\n" + "tbz x11, #1, 113f\n" + "str h23, [x9], #0x2\n" + "str h8, [x26], #0x2\n" + "str h15, [x25], #0x2\n" + "str h16, [x24], #0x2\n" + "tbz x11, #0, 114f\n" + "st1 { v23.b }[2], [x9]\n" + "st1 { v8.b }[2], [x26]\n" + "st1 { v15.b }[2], [x25]\n" + "st1 { v16.b }[2], [x24]\n" "b 114f\n" "113:" // Height 4: Partial direct writeback: partial_1_0 - "str b23, [x11, #0x0]\n" - "str b8, [x25, #0x0]\n" - "str b15, [x24, #0x0]\n" - "str b16, [x23, #0x0]\n" + "str b23, [x9, #0x0]\n" + "str b8, [x26, #0x0]\n" + "str b15, [x25, #0x0]\n" + "str b16, [x24, #0x0]\n" "114:" // Height 4: Partial direct writeback: Done "b 116f\n" "115:" // Height 4: Full writeback - "str q23, [x11, #0x0]\n" - "add x11, x11, #0x10\n" - "str q8, [x25, #0x0]\n" - "str q15, [x24, #0x0]\n" - "str q16, [x23, #0x0]\n" + "str q23, [x9, #0x0]\n" + "add x9, x9, #0x10\n" + "str q8, [x26, #0x0]\n" + "str q15, [x25, #0x0]\n" + "str q16, [x24, #0x0]\n" "116:" // Height 4: Writeback done - "subs x10, x10, #0x10\n" + "subs x11, x11, #0x10\n" "bgt 89b\n" "b 176f\n" "117:" // Height 5 "mov x14, %x[col_bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "mov x11, %x[output_ptr]\n" - "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" - "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "118:" // Height 5: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" @@ -2032,8 +2032,8 @@ void a64_hybrid_s8qs_mmla_6x16 ( "mov x28, #0x0\n" "120:" // 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, 121f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -2065,91 +2065,91 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr q3, [x24, #0x0]\n" "ldr q4, [x23, #0x0]\n" "ldr q5, [x22, #0x0]\n" - "ldr q7, [x9, #0x0]\n" + "ldr q7, [x10, #0x0]\n" "blt 124f\n" "123:" // Height 5: Multiply loop: Main loop head "trn1 v0.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x4e87a408 // smmla v8.4s, v0.16b, v7.16b\n" + "sub x27, x27, #0x10\n" "trn1 v2.2d, v3.2d, v4.2d\n" "trn2 v3.2d, v3.2d, v4.2d\n" - ".inst 0x4e87a450 // smmla v16.4s, v2.16b, v7.16b\n" - "sub x27, x27, #0x10\n" + "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" + "add x24, x24, #0x10\n" + "add x23, x23, #0x10\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "trn1 v4.2d, v5.2d, v6.2d\n" "trn2 v5.2d, v5.2d, v6.2d\n" - "ldr q6, [x9, #0x10]\n" + "ldr q6, [x10, #0x10]\n" + ".inst 0x4e87a408 // smmla v8.4s, v0.16b, v7.16b\n" + ".inst 0x4e87a450 // smmla v16.4s, v2.16b, v7.16b\n" + "add x22, x22, #0x10\n" + "cmp x27, #0x20\n" + "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4e87a498 // smmla v24.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x20]\n" + "ldr q7, [x10, #0x20]\n" ".inst 0x4e86a40c // smmla v12.4s, v0.16b, v6.16b\n" ".inst 0x4e86a454 // smmla v20.4s, v2.16b, v6.16b\n" - "add x26, x26, #0x10\n" ".inst 0x4e86a49c // smmla v28.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x30]\n" + "ldr q6, [x10, #0x30]\n" ".inst 0x4e87a409 // smmla v9.4s, v0.16b, v7.16b\n" - "add x25, x25, #0x10\n" ".inst 0x4e87a451 // smmla v17.4s, v2.16b, v7.16b\n" ".inst 0x4e87a499 // smmla v25.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x40]\n" - "add x24, x24, #0x10\n" + "ldr q7, [x10, #0x40]\n" ".inst 0x4e86a40d // smmla v13.4s, v0.16b, v6.16b\n" ".inst 0x4e86a455 // smmla v21.4s, v2.16b, v6.16b\n" - "add x23, x23, #0x10\n" - "add x22, x22, #0x10\n" ".inst 0x4e86a49d // smmla v29.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x50]\n" + "ldr q6, [x10, #0x50]\n" ".inst 0x4e87a40a // smmla v10.4s, v0.16b, v7.16b\n" - "cmp x27, #0x20\n" ".inst 0x4e87a452 // smmla v18.4s, v2.16b, v7.16b\n" ".inst 0x4e87a49a // smmla v26.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x60]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "ldr q7, [x10, #0x60]\n" ".inst 0x4e86a40e // smmla v14.4s, v0.16b, v6.16b\n" ".inst 0x4e86a456 // smmla v22.4s, v2.16b, v6.16b\n" - "prfm pldl1keep, [x25, #0x80]\n" - "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4e86a49e // smmla v30.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x70]\n" + "ldr q6, [x10, #0x70]\n" ".inst 0x4e87a40b // smmla v11.4s, v0.16b, v7.16b\n" - "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4e87a453 // smmla v19.4s, v2.16b, v7.16b\n" ".inst 0x4e87a49b // smmla v27.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x80]\n" - "prfm pldl1keep, [x22, #0x80]\n" + "ldr q7, [x10, #0x80]\n" ".inst 0x4e86a40f // smmla v15.4s, v0.16b, v6.16b\n" ".inst 0x4e86a457 // smmla v23.4s, v2.16b, v6.16b\n" "ldr q2, [x25, #0x0]\n" ".inst 0x4e86a49f // smmla v31.4s, v4.16b, v6.16b\n" - "ldr q0, [x9, #0x90]\n" + "ldr q0, [x10, #0x90]\n" "ldr q4, [x23, #0x0]\n" ".inst 0x4e87a428 // smmla v8.4s, v1.16b, v7.16b\n" ".inst 0x4e87a470 // smmla v16.4s, v3.16b, v7.16b\n" ".inst 0x4e87a4b8 // smmla v24.4s, v5.16b, v7.16b\n" - "ldr q6, [x9, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4e80a42c // smmla v12.4s, v1.16b, v0.16b\n" ".inst 0x4e80a474 // smmla v20.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4bc // smmla v28.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xb0]\n" + "ldr q0, [x10, #0xb0]\n" ".inst 0x4e86a429 // smmla v9.4s, v1.16b, v6.16b\n" ".inst 0x4e86a471 // smmla v17.4s, v3.16b, v6.16b\n" ".inst 0x4e86a4b9 // smmla v25.4s, v5.16b, v6.16b\n" - "ldr q6, [x9, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4e80a42d // smmla v13.4s, v1.16b, v0.16b\n" ".inst 0x4e80a475 // smmla v21.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4bd // smmla v29.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xd0]\n" + "ldr q0, [x10, #0xd0]\n" ".inst 0x4e86a42a // smmla v10.4s, v1.16b, v6.16b\n" ".inst 0x4e86a472 // smmla v18.4s, v3.16b, v6.16b\n" ".inst 0x4e86a4ba // smmla v26.4s, v5.16b, v6.16b\n" - "ldr q6, [x9, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4e80a42e // smmla v14.4s, v1.16b, v0.16b\n" ".inst 0x4e80a476 // smmla v22.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4be // smmla v30.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xf0]\n" - "add x9, x9, #0x100\n" + "ldr q0, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e86a42b // smmla v11.4s, v1.16b, v6.16b\n" ".inst 0x4e86a473 // smmla v19.4s, v3.16b, v6.16b\n" ".inst 0x4e86a4bb // smmla v27.4s, v5.16b, v6.16b\n" - "ldr q7, [x9, #0x0]\n" + "ldr q7, [x10, #0x0]\n" ".inst 0x4e80a42f // smmla v15.4s, v1.16b, v0.16b\n" "ldr q1, [x26, #0x0]\n" ".inst 0x4e80a477 // smmla v23.4s, v3.16b, v0.16b\n" @@ -2160,79 +2160,79 @@ void a64_hybrid_s8qs_mmla_6x16 ( "124:" // Height 5: Multiply loop: Single iteration only "trn1 v0.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x4e87a408 // smmla v8.4s, v0.16b, v7.16b\n" + "add x26, x26, #0x10\n" "trn1 v2.2d, v3.2d, v4.2d\n" "trn2 v3.2d, v3.2d, v4.2d\n" - ".inst 0x4e87a450 // smmla v16.4s, v2.16b, v7.16b\n" - "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" + "add x24, x24, #0x10\n" + "add x23, x23, #0x10\n" + "add x22, x22, #0x10\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" "trn1 v4.2d, v5.2d, v6.2d\n" "trn2 v5.2d, v5.2d, v6.2d\n" - "ldr q6, [x9, #0x10]\n" + "ldr q6, [x10, #0x10]\n" + ".inst 0x4e87a408 // smmla v8.4s, v0.16b, v7.16b\n" + ".inst 0x4e87a450 // smmla v16.4s, v2.16b, v7.16b\n" + "sub x27, x27, #0x10\n" + "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4e87a498 // smmla v24.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x20]\n" + "ldr q7, [x10, #0x20]\n" ".inst 0x4e86a40c // smmla v12.4s, v0.16b, v6.16b\n" ".inst 0x4e86a454 // smmla v20.4s, v2.16b, v6.16b\n" - "add x25, x25, #0x10\n" ".inst 0x4e86a49c // smmla v28.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x30]\n" + "ldr q6, [x10, #0x30]\n" ".inst 0x4e87a409 // smmla v9.4s, v0.16b, v7.16b\n" - "add x24, x24, #0x10\n" ".inst 0x4e87a451 // smmla v17.4s, v2.16b, v7.16b\n" ".inst 0x4e87a499 // smmla v25.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x40]\n" - "add x23, x23, #0x10\n" + "ldr q7, [x10, #0x40]\n" ".inst 0x4e86a40d // smmla v13.4s, v0.16b, v6.16b\n" ".inst 0x4e86a455 // smmla v21.4s, v2.16b, v6.16b\n" - "add x22, x22, #0x10\n" - "sub x27, x27, #0x10\n" ".inst 0x4e86a49d // smmla v29.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x50]\n" + "ldr q6, [x10, #0x50]\n" ".inst 0x4e87a40a // smmla v10.4s, v0.16b, v7.16b\n" - "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4e87a452 // smmla v18.4s, v2.16b, v7.16b\n" ".inst 0x4e87a49a // smmla v26.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x60]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "ldr q7, [x10, #0x60]\n" ".inst 0x4e86a40e // smmla v14.4s, v0.16b, v6.16b\n" ".inst 0x4e86a456 // smmla v22.4s, v2.16b, v6.16b\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4e86a49e // smmla v30.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x70]\n" + "ldr q6, [x10, #0x70]\n" ".inst 0x4e87a40b // smmla v11.4s, v0.16b, v7.16b\n" - "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4e87a453 // smmla v19.4s, v2.16b, v7.16b\n" ".inst 0x4e87a49b // smmla v27.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x80]\n" + "ldr q7, [x10, #0x80]\n" ".inst 0x4e86a40f // smmla v15.4s, v0.16b, v6.16b\n" ".inst 0x4e86a457 // smmla v23.4s, v2.16b, v6.16b\n" ".inst 0x4e86a49f // smmla v31.4s, v4.16b, v6.16b\n" - "ldr q0, [x9, #0x90]\n" + "ldr q0, [x10, #0x90]\n" ".inst 0x4e87a428 // smmla v8.4s, v1.16b, v7.16b\n" ".inst 0x4e87a470 // smmla v16.4s, v3.16b, v7.16b\n" ".inst 0x4e87a4b8 // smmla v24.4s, v5.16b, v7.16b\n" - "ldr q2, [x9, #0xa0]\n" + "ldr q2, [x10, #0xa0]\n" ".inst 0x4e80a42c // smmla v12.4s, v1.16b, v0.16b\n" ".inst 0x4e80a474 // smmla v20.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4bc // smmla v28.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xb0]\n" + "ldr q0, [x10, #0xb0]\n" ".inst 0x4e82a429 // smmla v9.4s, v1.16b, v2.16b\n" ".inst 0x4e82a471 // smmla v17.4s, v3.16b, v2.16b\n" ".inst 0x4e82a4b9 // smmla v25.4s, v5.16b, v2.16b\n" - "ldr q2, [x9, #0xc0]\n" + "ldr q2, [x10, #0xc0]\n" ".inst 0x4e80a42d // smmla v13.4s, v1.16b, v0.16b\n" ".inst 0x4e80a475 // smmla v21.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4bd // smmla v29.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xd0]\n" + "ldr q0, [x10, #0xd0]\n" ".inst 0x4e82a42a // smmla v10.4s, v1.16b, v2.16b\n" ".inst 0x4e82a472 // smmla v18.4s, v3.16b, v2.16b\n" ".inst 0x4e82a4ba // smmla v26.4s, v5.16b, v2.16b\n" - "ldr q2, [x9, #0xe0]\n" + "ldr q2, [x10, #0xe0]\n" ".inst 0x4e80a42e // smmla v14.4s, v1.16b, v0.16b\n" ".inst 0x4e80a476 // smmla v22.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4be // smmla v30.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xf0]\n" - "add x9, x9, #0x100\n" + "ldr q0, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e82a42b // smmla v11.4s, v1.16b, v2.16b\n" ".inst 0x4e82a473 // smmla v19.4s, v3.16b, v2.16b\n" ".inst 0x4e82a4bb // smmla v27.4s, v5.16b, v2.16b\n" @@ -2244,44 +2244,44 @@ void a64_hybrid_s8qs_mmla_6x16 ( "cmp x27, #0x8\n" "blt 127f\n" "126:" // Height 5: Multiply loop: Odd block loop - "ldr d1, [x26], #0x8\n" - "ldr d0, [x25], #0x8\n" - "trn1 v4.2d, v1.2d, v0.2d\n" - "ldr d1, [x24], #0x8\n" - "ldr d0, [x23], #0x8\n" - "trn1 v3.2d, v1.2d, v0.2d\n" + "ldr d6, [x26], #0x8\n" + "ldr d4, [x25], #0x8\n" "sub x27, x27, #0x8\n" + "ldr d3, [x24], #0x8\n" + "ldr d2, [x23], #0x8\n" + "cmp x27, #0x8\n" "ldr d0, [x22], #0x8\n" - "ldr q1, [x9, #0x0]\n" - "trn1 v2.2d, v0.2d, v2.2d\n" + "ldr q1, [x10, #0x0]\n" + "trn1 v4.2d, v6.2d, v4.2d\n" + "trn1 v3.2d, v3.2d, v2.2d\n" + "trn1 v2.2d, v0.2d, v5.2d\n" + "ldr q0, [x10, #0x10]\n" ".inst 0x4e81a488 // smmla v8.4s, v4.16b, v1.16b\n" - "ldr q0, [x9, #0x10]\n" ".inst 0x4e81a470 // smmla v16.4s, v3.16b, v1.16b\n" ".inst 0x4e81a458 // smmla v24.4s, v2.16b, v1.16b\n" - "ldr q1, [x9, #0x20]\n" + "ldr q1, [x10, #0x20]\n" ".inst 0x4e80a48c // smmla v12.4s, v4.16b, v0.16b\n" ".inst 0x4e80a474 // smmla v20.4s, v3.16b, v0.16b\n" - "cmp x27, #0x8\n" ".inst 0x4e80a45c // smmla v28.4s, v2.16b, v0.16b\n" - "ldr q0, [x9, #0x30]\n" + "ldr q0, [x10, #0x30]\n" ".inst 0x4e81a489 // smmla v9.4s, v4.16b, v1.16b\n" ".inst 0x4e81a471 // smmla v17.4s, v3.16b, v1.16b\n" ".inst 0x4e81a459 // smmla v25.4s, v2.16b, v1.16b\n" - "ldr q1, [x9, #0x40]\n" + "ldr q1, [x10, #0x40]\n" ".inst 0x4e80a48d // smmla v13.4s, v4.16b, v0.16b\n" ".inst 0x4e80a475 // smmla v21.4s, v3.16b, v0.16b\n" ".inst 0x4e80a45d // smmla v29.4s, v2.16b, v0.16b\n" - "ldr q0, [x9, #0x50]\n" + "ldr q0, [x10, #0x50]\n" ".inst 0x4e81a48a // smmla v10.4s, v4.16b, v1.16b\n" ".inst 0x4e81a472 // smmla v18.4s, v3.16b, v1.16b\n" ".inst 0x4e81a45a // smmla v26.4s, v2.16b, v1.16b\n" - "ldr q1, [x9, #0x60]\n" + "ldr q1, [x10, #0x60]\n" ".inst 0x4e80a48e // smmla v14.4s, v4.16b, v0.16b\n" ".inst 0x4e80a476 // smmla v22.4s, v3.16b, v0.16b\n" ".inst 0x4e80a45e // smmla v30.4s, v2.16b, v0.16b\n" - "ldr q0, [x9, #0x70]\n" + "ldr q0, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e81a48b // smmla v11.4s, v4.16b, v1.16b\n" - "add x9, x9, #0x80\n" ".inst 0x4e81a473 // smmla v19.4s, v3.16b, v1.16b\n" ".inst 0x4e81a45b // smmla v27.4s, v2.16b, v1.16b\n" ".inst 0x4e80a48f // smmla v15.4s, v4.16b, v0.16b\n" @@ -2338,36 +2338,36 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr b4, [x23, #0x0]\n" "ldr b5, [x22, #0x0]\n" "131:" // Height 5: Multiply loop: Ragged operand read: Done - "ldr q7, [x9, #0x0]\n" + "ldr q7, [x10, #0x0]\n" "trn1 v6.2d, v1.2d, v2.2d\n" "trn1 v3.2d, v3.2d, v4.2d\n" "trn1 v2.2d, v5.2d, v0.2d\n" - "ldr q0, [x9, #0x10]\n" + "ldr q0, [x10, #0x10]\n" ".inst 0x4e87a4c8 // smmla v8.4s, v6.16b, v7.16b\n" ".inst 0x4e87a470 // smmla v16.4s, v3.16b, v7.16b\n" ".inst 0x4e87a458 // smmla v24.4s, v2.16b, v7.16b\n" - "ldr q1, [x9, #0x20]\n" + "ldr q1, [x10, #0x20]\n" ".inst 0x4e80a4cc // smmla v12.4s, v6.16b, v0.16b\n" ".inst 0x4e80a474 // smmla v20.4s, v3.16b, v0.16b\n" ".inst 0x4e80a45c // smmla v28.4s, v2.16b, v0.16b\n" - "ldr q0, [x9, #0x30]\n" + "ldr q0, [x10, #0x30]\n" ".inst 0x4e81a4c9 // smmla v9.4s, v6.16b, v1.16b\n" ".inst 0x4e81a471 // smmla v17.4s, v3.16b, v1.16b\n" ".inst 0x4e81a459 // smmla v25.4s, v2.16b, v1.16b\n" - "ldr q1, [x9, #0x40]\n" + "ldr q1, [x10, #0x40]\n" ".inst 0x4e80a4cd // smmla v13.4s, v6.16b, v0.16b\n" ".inst 0x4e80a475 // smmla v21.4s, v3.16b, v0.16b\n" ".inst 0x4e80a45d // smmla v29.4s, v2.16b, v0.16b\n" - "ldr q0, [x9, #0x50]\n" + "ldr q0, [x10, #0x50]\n" ".inst 0x4e81a4ca // smmla v10.4s, v6.16b, v1.16b\n" ".inst 0x4e81a472 // smmla v18.4s, v3.16b, v1.16b\n" ".inst 0x4e81a45a // smmla v26.4s, v2.16b, v1.16b\n" - "ldr q1, [x9, #0x60]\n" + "ldr q1, [x10, #0x60]\n" ".inst 0x4e80a4ce // smmla v14.4s, v6.16b, v0.16b\n" ".inst 0x4e80a476 // smmla v22.4s, v3.16b, v0.16b\n" ".inst 0x4e80a45e // smmla v30.4s, v2.16b, v0.16b\n" - "ldr q0, [x9, #0x70]\n" - "add x9, x9, #0x80\n" + "ldr q0, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e81a4cb // smmla v11.4s, v6.16b, v1.16b\n" ".inst 0x4e81a473 // smmla v19.4s, v3.16b, v1.16b\n" ".inst 0x4e81a45b // smmla v27.4s, v2.16b, v1.16b\n" @@ -2388,37 +2388,37 @@ void a64_hybrid_s8qs_mmla_6x16 ( "uzp1 v12.2d, v9.2d, v13.2d\n" "uzp2 v9.2d, v9.2d, v13.2d\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x11, x20\n" "uzp1 v13.2d, v10.2d, v14.2d\n" "uzp2 v10.2d, v10.2d, v14.2d\n" + "prfm pstl1keep, [x9, #0x0]\n" "uzp1 v14.2d, v11.2d, v15.2d\n" "uzp2 v11.2d, v11.2d, v15.2d\n" - "add x24, x25, x20\n" - "add x23, x24, x20\n" + "add x14, x14, #0x40\n" "uzp1 v15.2d, v16.2d, v20.2d\n" "uzp2 v16.2d, v16.2d, v20.2d\n" - "add x22, x23, x20\n" - "prfm pstl1keep, [x11, #0x0]\n" + "add x26, x9, x20\n" "uzp1 v20.2d, v17.2d, v21.2d\n" "uzp2 v17.2d, v17.2d, v21.2d\n" - "prfm pstl1keep, [x25, #0x0]\n" - "prfm pstl1keep, [x24, #0x0]\n" + "add x25, x26, x20\n" + "prfm pstl1keep, [x26, #0x0]\n" "uzp1 v21.2d, v18.2d, v22.2d\n" "uzp2 v18.2d, v18.2d, v22.2d\n" - "prfm pstl1keep, [x23, #0x0]\n" - "prfm pstl1keep, [x22, #0x0]\n" + "add x24, x25, x20\n" + "prfm pstl1keep, [x25, #0x0]\n" "uzp1 v22.2d, v19.2d, v23.2d\n" "uzp2 v19.2d, v19.2d, v23.2d\n" - "add x14, x14, #0x40\n" + "add x23, x24, x20\n" + "prfm pstl1keep, [x24, #0x0]\n" "uzp1 v24.2d, v24.2d, v28.2d\n" "uzp1 v25.2d, v25.2d, v29.2d\n" + "prfm pstl1keep, [x23, #0x0]\n" "uzp1 v26.2d, v26.2d, v30.2d\n" "uzp1 v27.2d, v27.2d, v31.2d\n" "mov v31.16b, v2.16b\n" - "add v31.4s, v31.4s, v4.4s\n" "add v12.4s, v12.4s, v3.4s\n" "add v13.4s, v13.4s, v1.4s\n" "add v14.4s, v14.4s, v0.4s\n" + "add v31.4s, v31.4s, v4.4s\n" "add v8.4s, v8.4s, v4.4s\n" "add v9.4s, v9.4s, v3.4s\n" "add v10.4s, v10.4s, v1.4s\n" @@ -2448,9 +2448,9 @@ void a64_hybrid_s8qs_mmla_6x16 ( "add x13, x13, #0x40\n" "b 134f\n" "133:" // Height 5: per layer parameters - "add x20, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x20]\n" + "add x21, %x[qp], %[per_layer_right_shift]\n" "add x20, %x[qp], %[per_layer_mul]\n" + "ld1r { v0.4s }, [x21]\n" "ld1r { v4.4s }, [x20]\n" "mov v1.16b, v0.16b\n" "mov v5.16b, v4.16b\n" @@ -2489,11 +2489,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v28.4s, v28.4s, #0x1f\n" "sshr v23.4s, v23.4s, #0x1f\n" "sqadd v31.4s, v31.4s, v30.4s\n" + "and v30.16b, v8.16b, v0.16b\n" "sqadd v12.4s, v12.4s, v29.4s\n" + "and v29.16b, v9.16b, v1.16b\n" "sqadd v13.4s, v13.4s, v28.4s\n" "sqadd v14.4s, v14.4s, v23.4s\n" - "and v30.16b, v8.16b, v0.16b\n" - "and v29.16b, v9.16b, v1.16b\n" "and v28.16b, v10.16b, v2.16b\n" "and v23.16b, v11.16b, v3.16b\n" "sshr v30.4s, v30.4s, #0x1f\n" @@ -2501,11 +2501,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v28.4s, v28.4s, #0x1f\n" "sshr v23.4s, v23.4s, #0x1f\n" "sqadd v8.4s, v8.4s, v30.4s\n" + "and v30.16b, v15.16b, v0.16b\n" "sqadd v9.4s, v9.4s, v29.4s\n" + "and v29.16b, v20.16b, v1.16b\n" "sqadd v10.4s, v10.4s, v28.4s\n" "sqadd v11.4s, v11.4s, v23.4s\n" - "and v30.16b, v15.16b, v0.16b\n" - "and v29.16b, v20.16b, v1.16b\n" "and v28.16b, v21.16b, v2.16b\n" "and v23.16b, v22.16b, v3.16b\n" "sshr v30.4s, v30.4s, #0x1f\n" @@ -2513,11 +2513,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v28.4s, v28.4s, #0x1f\n" "sshr v23.4s, v23.4s, #0x1f\n" "sqadd v15.4s, v15.4s, v30.4s\n" + "and v30.16b, v16.16b, v0.16b\n" "sqadd v20.4s, v20.4s, v29.4s\n" + "and v29.16b, v17.16b, v1.16b\n" "sqadd v21.4s, v21.4s, v28.4s\n" "sqadd v22.4s, v22.4s, v23.4s\n" - "and v30.16b, v16.16b, v0.16b\n" - "and v29.16b, v17.16b, v1.16b\n" "and v28.16b, v18.16b, v2.16b\n" "and v23.16b, v19.16b, v3.16b\n" "sshr v30.4s, v30.4s, #0x1f\n" @@ -2525,11 +2525,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v28.4s, v28.4s, #0x1f\n" "sshr v23.4s, v23.4s, #0x1f\n" "sqadd v16.4s, v16.4s, v30.4s\n" + "and v30.16b, v24.16b, v0.16b\n" "sqadd v17.4s, v17.4s, v29.4s\n" + "and v29.16b, v25.16b, v1.16b\n" "sqadd v18.4s, v18.4s, v28.4s\n" "sqadd v19.4s, v19.4s, v23.4s\n" - "and v30.16b, v24.16b, v0.16b\n" - "and v29.16b, v25.16b, v1.16b\n" "and v28.16b, v26.16b, v2.16b\n" "and v23.16b, v27.16b, v3.16b\n" "sshr v30.4s, v30.4s, #0x1f\n" @@ -2541,21 +2541,21 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sqadd v26.4s, v26.4s, v28.4s\n" "sqadd v27.4s, v27.4s, v23.4s\n" "135:" // Height 5: no shift correction - "add x20, %x[qp], %[c_offset]\n" - "ld1r { v29.4s }, [x20]\n" + "add x21, %x[qp], %[c_offset]\n" "srshl v31.4s, v31.4s, v0.4s\n" "srshl v12.4s, v12.4s, v1.4s\n" - "srshl v13.4s, v13.4s, v2.4s\n" - "srshl v14.4s, v14.4s, v3.4s\n" "add x20, %x[qp], %[maxval]\n" + "ld1r { v29.4s }, [x21]\n" "ld1r { v28.4s }, [x20]\n" + "srshl v13.4s, v13.4s, v2.4s\n" + "srshl v14.4s, v14.4s, v3.4s\n" "srshl v8.4s, v8.4s, v0.4s\n" "srshl v9.4s, v9.4s, v1.4s\n" "add x20, %x[qp], %[minval]\n" + "cmp x11, #0x10\n" "ld1r { v23.4s }, [x20]\n" "srshl v10.4s, v10.4s, v2.4s\n" "srshl v11.4s, v11.4s, v3.4s\n" - "cmp x10, #0x10\n" "srshl v15.4s, v15.4s, v0.4s\n" "srshl v20.4s, v20.4s, v1.4s\n" "srshl v21.4s, v21.4s, v2.4s\n" @@ -2644,132 +2644,133 @@ void a64_hybrid_s8qs_mmla_6x16 ( "uzp1 v16.16b, v16.16b, v18.16b\n" "uzp1 v24.16b, v24.16b, v17.16b\n" "bge 144f\n" - "tbz x10, #3, 139f\n" - "str d31, [x11], #0x8\n" - "str d8, [x25], #0x8\n" - "str d15, [x24], #0x8\n" - "str d16, [x23], #0x8\n" - "str d24, [x22], #0x8\n" - "tbz x10, #2, 137f\n" - "st1 { v31.s }[2], [x11], #0x4\n" - "st1 { v8.s }[2], [x25], #0x4\n" - "st1 { v15.s }[2], [x24], #0x4\n" - "st1 { v16.s }[2], [x23], #0x4\n" - "st1 { v24.s }[2], [x22], #0x4\n" - "tbz x10, #1, 136f\n" - "st1 { v31.h }[6], [x11], #0x2\n" - "st1 { v8.h }[6], [x25], #0x2\n" - "st1 { v15.h }[6], [x24], #0x2\n" - "st1 { v16.h }[6], [x23], #0x2\n" - "st1 { v24.h }[6], [x22], #0x2\n" - "tbz x10, #0, 143f\n" - "st1 { v31.b }[14], [x11]\n" - "st1 { v8.b }[14], [x25]\n" - "st1 { v15.b }[14], [x24]\n" - "st1 { v16.b }[14], [x23]\n" - "st1 { v24.b }[14], [x22]\n" + "tbz x11, #3, 139f\n" + "str d31, [x9], #0x8\n" + "str d8, [x26], #0x8\n" + "str d15, [x25], #0x8\n" + "str d16, [x24], #0x8\n" + "str d24, [x23], #0x8\n" + "tbz x11, #2, 137f\n" + "st1 { v31.s }[2], [x9], #0x4\n" + "st1 { v8.s }[2], [x26], #0x4\n" + "st1 { v15.s }[2], [x25], #0x4\n" + "st1 { v16.s }[2], [x24], #0x4\n" + "st1 { v24.s }[2], [x23], #0x4\n" + "tbz x11, #1, 136f\n" + "st1 { v31.h }[6], [x9], #0x2\n" + "st1 { v8.h }[6], [x26], #0x2\n" + "st1 { v15.h }[6], [x25], #0x2\n" + "st1 { v16.h }[6], [x24], #0x2\n" + "st1 { v24.h }[6], [x23], #0x2\n" + "tbz x11, #0, 143f\n" + "st1 { v31.b }[14], [x9]\n" + "st1 { v8.b }[14], [x26]\n" + "st1 { v15.b }[14], [x25]\n" + "st1 { v16.b }[14], [x24]\n" + "st1 { v24.b }[14], [x23]\n" "b 143f\n" "136:" // Height 5: Partial direct writeback: partial_1_12 - "tbz x10, #0, 143f\n" - "st1 { v31.b }[12], [x11]\n" - "st1 { v8.b }[12], [x25]\n" - "st1 { v15.b }[12], [x24]\n" - "st1 { v16.b }[12], [x23]\n" - "st1 { v24.b }[12], [x22]\n" + "tbz x11, #0, 143f\n" + "st1 { v31.b }[12], [x9]\n" + "st1 { v8.b }[12], [x26]\n" + "st1 { v15.b }[12], [x25]\n" + "st1 { v16.b }[12], [x24]\n" + "st1 { v24.b }[12], [x23]\n" "b 143f\n" "137:" // Height 5: Partial direct writeback: partial_2_8 - "tbz x10, #1, 138f\n" - "st1 { v31.h }[4], [x11], #0x2\n" - "st1 { v8.h }[4], [x25], #0x2\n" - "st1 { v15.h }[4], [x24], #0x2\n" - "st1 { v16.h }[4], [x23], #0x2\n" - "st1 { v24.h }[4], [x22], #0x2\n" - "tbz x10, #0, 143f\n" - "st1 { v31.b }[10], [x11]\n" - "st1 { v8.b }[10], [x25]\n" - "st1 { v15.b }[10], [x24]\n" - "st1 { v16.b }[10], [x23]\n" - "st1 { v24.b }[10], [x22]\n" + "tbz x11, #1, 138f\n" + "st1 { v31.h }[4], [x9], #0x2\n" + "st1 { v8.h }[4], [x26], #0x2\n" + "st1 { v15.h }[4], [x25], #0x2\n" + "st1 { v16.h }[4], [x24], #0x2\n" + "st1 { v24.h }[4], [x23], #0x2\n" + "tbz x11, #0, 143f\n" + "st1 { v31.b }[10], [x9]\n" + "st1 { v8.b }[10], [x26]\n" + "st1 { v15.b }[10], [x25]\n" + "st1 { v16.b }[10], [x24]\n" + "st1 { v24.b }[10], [x23]\n" "b 143f\n" "138:" // Height 5: Partial direct writeback: partial_1_8 - "tbz x10, #0, 143f\n" - "st1 { v31.b }[8], [x11]\n" - "st1 { v8.b }[8], [x25]\n" - "st1 { v15.b }[8], [x24]\n" - "st1 { v16.b }[8], [x23]\n" - "st1 { v24.b }[8], [x22]\n" + "tbz x11, #0, 143f\n" + "st1 { v31.b }[8], [x9]\n" + "st1 { v8.b }[8], [x26]\n" + "st1 { v15.b }[8], [x25]\n" + "st1 { v16.b }[8], [x24]\n" + "st1 { v24.b }[8], [x23]\n" "b 143f\n" "139:" // Height 5: Partial direct writeback: partial_4_0 - "tbz x10, #2, 141f\n" - "str s31, [x11], #0x4\n" - "str s8, [x25], #0x4\n" - "str s15, [x24], #0x4\n" - "str s16, [x23], #0x4\n" - "str s24, [x22], #0x4\n" - "tbz x10, #1, 140f\n" - "st1 { v31.h }[2], [x11], #0x2\n" - "st1 { v8.h }[2], [x25], #0x2\n" - "st1 { v15.h }[2], [x24], #0x2\n" - "st1 { v16.h }[2], [x23], #0x2\n" - "st1 { v24.h }[2], [x22], #0x2\n" - "tbz x10, #0, 143f\n" - "st1 { v31.b }[6], [x11]\n" - "st1 { v8.b }[6], [x25]\n" - "st1 { v15.b }[6], [x24]\n" - "st1 { v16.b }[6], [x23]\n" - "st1 { v24.b }[6], [x22]\n" + "tbz x11, #2, 141f\n" + "str s31, [x9], #0x4\n" + "str s8, [x26], #0x4\n" + "str s15, [x25], #0x4\n" + "str s16, [x24], #0x4\n" + "str s24, [x23], #0x4\n" + "tbz x11, #1, 140f\n" + "st1 { v31.h }[2], [x9], #0x2\n" + "st1 { v8.h }[2], [x26], #0x2\n" + "st1 { v15.h }[2], [x25], #0x2\n" + "st1 { v16.h }[2], [x24], #0x2\n" + "st1 { v24.h }[2], [x23], #0x2\n" + "tbz x11, #0, 143f\n" + "st1 { v31.b }[6], [x9]\n" + "st1 { v8.b }[6], [x26]\n" + "st1 { v15.b }[6], [x25]\n" + "st1 { v16.b }[6], [x24]\n" + "st1 { v24.b }[6], [x23]\n" "b 143f\n" "140:" // Height 5: Partial direct writeback: partial_1_4 - "tbz x10, #0, 143f\n" - "st1 { v31.b }[4], [x11]\n" - "st1 { v8.b }[4], [x25]\n" - "st1 { v15.b }[4], [x24]\n" - "st1 { v16.b }[4], [x23]\n" - "st1 { v24.b }[4], [x22]\n" + "tbz x11, #0, 143f\n" + "st1 { v31.b }[4], [x9]\n" + "st1 { v8.b }[4], [x26]\n" + "st1 { v15.b }[4], [x25]\n" + "st1 { v16.b }[4], [x24]\n" + "st1 { v24.b }[4], [x23]\n" "b 143f\n" "141:" // Height 5: Partial direct writeback: partial_2_0 - "tbz x10, #1, 142f\n" - "str h31, [x11], #0x2\n" - "str h8, [x25], #0x2\n" - "str h15, [x24], #0x2\n" - "str h16, [x23], #0x2\n" - "str h24, [x22], #0x2\n" - "tbz x10, #0, 143f\n" - "st1 { v31.b }[2], [x11]\n" - "st1 { v8.b }[2], [x25]\n" - "st1 { v15.b }[2], [x24]\n" - "st1 { v16.b }[2], [x23]\n" - "st1 { v24.b }[2], [x22]\n" + "tbz x11, #1, 142f\n" + "str h31, [x9], #0x2\n" + "str h8, [x26], #0x2\n" + "str h15, [x25], #0x2\n" + "str h16, [x24], #0x2\n" + "str h24, [x23], #0x2\n" + "tbz x11, #0, 143f\n" + "st1 { v31.b }[2], [x9]\n" + "st1 { v8.b }[2], [x26]\n" + "st1 { v15.b }[2], [x25]\n" + "st1 { v16.b }[2], [x24]\n" + "st1 { v24.b }[2], [x23]\n" "b 143f\n" "142:" // Height 5: Partial direct writeback: partial_1_0 - "str b31, [x11, #0x0]\n" - "str b8, [x25, #0x0]\n" - "str b15, [x24, #0x0]\n" - "str b16, [x23, #0x0]\n" - "str b24, [x22, #0x0]\n" + "str b31, [x9, #0x0]\n" + "str b8, [x26, #0x0]\n" + "str b15, [x25, #0x0]\n" + "str b16, [x24, #0x0]\n" + "str b24, [x23, #0x0]\n" "143:" // Height 5: Partial direct writeback: Done "b 145f\n" "144:" // Height 5: Full writeback - "str q31, [x11, #0x0]\n" - "add x11, x11, #0x10\n" - "str q8, [x25, #0x0]\n" - "str q15, [x24, #0x0]\n" - "str q16, [x23, #0x0]\n" - "str q24, [x22, #0x0]\n" + "str q31, [x9, #0x0]\n" + "add x9, x9, #0x10\n" + "str q8, [x26, #0x0]\n" + "str q15, [x25, #0x0]\n" + "str q16, [x24, #0x0]\n" + "str q24, [x23, #0x0]\n" "145:" // Height 5: Writeback done - "subs x10, x10, #0x10\n" + "subs x11, x11, #0x10\n" "bgt 118b\n" "b 176f\n" "146:" // Height 6 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "mov x20, #0x6\n" "mov x14, %x[col_bias]\n" "ldr x13, [%x[args_ptr], %[offsetof_multiplier_ptr]]\n" "ldr x12, [%x[args_ptr], %[offsetof_shift_ptr]]\n" - "ldr x10, [%x[args_ptr], %[offsetof_N]]\n" - "mov x11, %x[output_ptr]\n" - "madd %x[output_ptr], x21, x20, %x[output_ptr]\n" - "ldr x9, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" + "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" + "madd x20, x21, x20, x9\n" + "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n" "147:" // Height 6: Column loop "movi v8.4s, #0x0\n" "movi v9.4s, #0x0\n" @@ -2799,8 +2800,8 @@ void a64_hybrid_s8qs_mmla_6x16 ( "mov x28, #0x0\n" "149:" // 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, 150f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -2836,93 +2837,93 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr q4, [x23, #0x0]\n" "ldr q5, [x22, #0x0]\n" "ldr q6, [x21, #0x0]\n" - "ldr q7, [x9, #0x0]\n" + "ldr q7, [x10, #0x0]\n" "blt 153f\n" "152:" // Height 6: Multiply loop: Main loop head "trn1 v0.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x4e87a408 // smmla v8.4s, v0.16b, v7.16b\n" "sub x27, x27, #0x10\n" + "add x26, x26, #0x10\n" "trn1 v2.2d, v3.2d, v4.2d\n" "trn2 v3.2d, v3.2d, v4.2d\n" - ".inst 0x4e87a450 // smmla v16.4s, v2.16b, v7.16b\n" - "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" + "add x24, x24, #0x10\n" "trn1 v4.2d, v5.2d, v6.2d\n" "trn2 v5.2d, v5.2d, v6.2d\n" - "ldr q6, [x9, #0x10]\n" + "ldr q6, [x10, #0x10]\n" + "add x23, x23, #0x10\n" + ".inst 0x4e87a408 // smmla v8.4s, v0.16b, v7.16b\n" + "add x22, x22, #0x10\n" + "add x21, x21, #0x10\n" + "prfm pldl1keep, [x26, #0x80]\n" + ".inst 0x4e87a450 // smmla v16.4s, v2.16b, v7.16b\n" + "cmp x27, #0x20\n" + "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4e87a498 // smmla v24.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x20]\n" + "ldr q7, [x10, #0x20]\n" ".inst 0x4e86a40c // smmla v12.4s, v0.16b, v6.16b\n" + "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4e86a454 // smmla v20.4s, v2.16b, v6.16b\n" - "add x25, x25, #0x10\n" ".inst 0x4e86a49c // smmla v28.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x30]\n" + "ldr q6, [x10, #0x30]\n" + "prfm pldl1keep, [x22, #0x80]\n" + "prfm pldl1keep, [x21, #0x80]\n" ".inst 0x4e87a409 // smmla v9.4s, v0.16b, v7.16b\n" - "add x24, x24, #0x10\n" ".inst 0x4e87a451 // smmla v17.4s, v2.16b, v7.16b\n" ".inst 0x4e87a499 // smmla v25.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x40]\n" - "add x23, x23, #0x10\n" + "ldr q7, [x10, #0x40]\n" ".inst 0x4e86a40d // smmla v13.4s, v0.16b, v6.16b\n" ".inst 0x4e86a455 // smmla v21.4s, v2.16b, v6.16b\n" - "add x22, x22, #0x10\n" - "add x21, x21, #0x10\n" ".inst 0x4e86a49d // smmla v29.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x50]\n" + "ldr q6, [x10, #0x50]\n" ".inst 0x4e87a40a // smmla v10.4s, v0.16b, v7.16b\n" - "cmp x27, #0x20\n" ".inst 0x4e87a452 // smmla v18.4s, v2.16b, v7.16b\n" ".inst 0x4e87a49a // smmla v26.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x60]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "ldr q7, [x10, #0x60]\n" ".inst 0x4e86a40e // smmla v14.4s, v0.16b, v6.16b\n" ".inst 0x4e86a456 // smmla v22.4s, v2.16b, v6.16b\n" - "prfm pldl1keep, [x25, #0x80]\n" - "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4e86a49e // smmla v30.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x70]\n" + "ldr q6, [x10, #0x70]\n" ".inst 0x4e87a40b // smmla v11.4s, v0.16b, v7.16b\n" - "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4e87a453 // smmla v19.4s, v2.16b, v7.16b\n" ".inst 0x4e87a49b // smmla v27.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x80]\n" - "prfm pldl1keep, [x22, #0x80]\n" + "ldr q7, [x10, #0x80]\n" ".inst 0x4e86a40f // smmla v15.4s, v0.16b, v6.16b\n" ".inst 0x4e86a457 // smmla v23.4s, v2.16b, v6.16b\n" "ldr q2, [x25, #0x0]\n" - "prfm pldl1keep, [x21, #0x80]\n" ".inst 0x4e86a49f // smmla v31.4s, v4.16b, v6.16b\n" - "ldr q0, [x9, #0x90]\n" + "ldr q0, [x10, #0x90]\n" "ldr q4, [x23, #0x0]\n" ".inst 0x4e87a428 // smmla v8.4s, v1.16b, v7.16b\n" ".inst 0x4e87a470 // smmla v16.4s, v3.16b, v7.16b\n" ".inst 0x4e87a4b8 // smmla v24.4s, v5.16b, v7.16b\n" - "ldr q6, [x9, #0xa0]\n" + "ldr q6, [x10, #0xa0]\n" ".inst 0x4e80a42c // smmla v12.4s, v1.16b, v0.16b\n" ".inst 0x4e80a474 // smmla v20.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4bc // smmla v28.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xb0]\n" + "ldr q0, [x10, #0xb0]\n" ".inst 0x4e86a429 // smmla v9.4s, v1.16b, v6.16b\n" ".inst 0x4e86a471 // smmla v17.4s, v3.16b, v6.16b\n" ".inst 0x4e86a4b9 // smmla v25.4s, v5.16b, v6.16b\n" - "ldr q6, [x9, #0xc0]\n" + "ldr q6, [x10, #0xc0]\n" ".inst 0x4e80a42d // smmla v13.4s, v1.16b, v0.16b\n" ".inst 0x4e80a475 // smmla v21.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4bd // smmla v29.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xd0]\n" + "ldr q0, [x10, #0xd0]\n" ".inst 0x4e86a42a // smmla v10.4s, v1.16b, v6.16b\n" ".inst 0x4e86a472 // smmla v18.4s, v3.16b, v6.16b\n" ".inst 0x4e86a4ba // smmla v26.4s, v5.16b, v6.16b\n" - "ldr q6, [x9, #0xe0]\n" + "ldr q6, [x10, #0xe0]\n" ".inst 0x4e80a42e // smmla v14.4s, v1.16b, v0.16b\n" ".inst 0x4e80a476 // smmla v22.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4be // smmla v30.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xf0]\n" - "add x9, x9, #0x100\n" + "ldr q0, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e86a42b // smmla v11.4s, v1.16b, v6.16b\n" ".inst 0x4e86a473 // smmla v19.4s, v3.16b, v6.16b\n" ".inst 0x4e86a4bb // smmla v27.4s, v5.16b, v6.16b\n" - "ldr q7, [x9, #0x0]\n" + "ldr q7, [x10, #0x0]\n" ".inst 0x4e80a42f // smmla v15.4s, v1.16b, v0.16b\n" "ldr q1, [x26, #0x0]\n" ".inst 0x4e80a477 // smmla v23.4s, v3.16b, v0.16b\n" @@ -2934,81 +2935,81 @@ void a64_hybrid_s8qs_mmla_6x16 ( "153:" // Height 6: Multiply loop: Single iteration only "trn1 v0.2d, v1.2d, v2.2d\n" "trn2 v1.2d, v1.2d, v2.2d\n" - ".inst 0x4e87a408 // smmla v8.4s, v0.16b, v7.16b\n" "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" "trn1 v2.2d, v3.2d, v4.2d\n" "trn2 v3.2d, v3.2d, v4.2d\n" - ".inst 0x4e87a450 // smmla v16.4s, v2.16b, v7.16b\n" - "add x25, x25, #0x10\n" + "add x24, x24, #0x10\n" + "add x23, x23, #0x10\n" "trn1 v4.2d, v5.2d, v6.2d\n" "trn2 v5.2d, v5.2d, v6.2d\n" - "ldr q6, [x9, #0x10]\n" + "ldr q6, [x10, #0x10]\n" + "add x22, x22, #0x10\n" + ".inst 0x4e87a408 // smmla v8.4s, v0.16b, v7.16b\n" + "add x21, x21, #0x10\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + ".inst 0x4e87a450 // smmla v16.4s, v2.16b, v7.16b\n" + "prfm pldl1keep, [x24, #0x80]\n" + "sub x27, x27, #0x10\n" + "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4e87a498 // smmla v24.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x20]\n" + "ldr q7, [x10, #0x20]\n" ".inst 0x4e86a40c // smmla v12.4s, v0.16b, v6.16b\n" + "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4e86a454 // smmla v20.4s, v2.16b, v6.16b\n" - "add x24, x24, #0x10\n" ".inst 0x4e86a49c // smmla v28.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x30]\n" + "ldr q6, [x10, #0x30]\n" + "prfm pldl1keep, [x21, #0x80]\n" ".inst 0x4e87a409 // smmla v9.4s, v0.16b, v7.16b\n" - "add x23, x23, #0x10\n" ".inst 0x4e87a451 // smmla v17.4s, v2.16b, v7.16b\n" ".inst 0x4e87a499 // smmla v25.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x40]\n" - "add x22, x22, #0x10\n" + "ldr q7, [x10, #0x40]\n" ".inst 0x4e86a40d // smmla v13.4s, v0.16b, v6.16b\n" ".inst 0x4e86a455 // smmla v21.4s, v2.16b, v6.16b\n" - "add x21, x21, #0x10\n" - "sub x27, x27, #0x10\n" ".inst 0x4e86a49d // smmla v29.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x50]\n" + "ldr q6, [x10, #0x50]\n" ".inst 0x4e87a40a // smmla v10.4s, v0.16b, v7.16b\n" - "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4e87a452 // smmla v18.4s, v2.16b, v7.16b\n" ".inst 0x4e87a49a // smmla v26.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x60]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "ldr q7, [x10, #0x60]\n" ".inst 0x4e86a40e // smmla v14.4s, v0.16b, v6.16b\n" ".inst 0x4e86a456 // smmla v22.4s, v2.16b, v6.16b\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4e86a49e // smmla v30.4s, v4.16b, v6.16b\n" - "ldr q6, [x9, #0x70]\n" + "ldr q6, [x10, #0x70]\n" ".inst 0x4e87a40b // smmla v11.4s, v0.16b, v7.16b\n" - "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4e87a453 // smmla v19.4s, v2.16b, v7.16b\n" ".inst 0x4e87a49b // smmla v27.4s, v4.16b, v7.16b\n" - "ldr q7, [x9, #0x80]\n" - "prfm pldl1keep, [x21, #0x80]\n" + "ldr q7, [x10, #0x80]\n" ".inst 0x4e86a40f // smmla v15.4s, v0.16b, v6.16b\n" ".inst 0x4e86a457 // smmla v23.4s, v2.16b, v6.16b\n" ".inst 0x4e86a49f // smmla v31.4s, v4.16b, v6.16b\n" - "ldr q0, [x9, #0x90]\n" + "ldr q0, [x10, #0x90]\n" ".inst 0x4e87a428 // smmla v8.4s, v1.16b, v7.16b\n" ".inst 0x4e87a470 // smmla v16.4s, v3.16b, v7.16b\n" ".inst 0x4e87a4b8 // smmla v24.4s, v5.16b, v7.16b\n" - "ldr q2, [x9, #0xa0]\n" + "ldr q2, [x10, #0xa0]\n" ".inst 0x4e80a42c // smmla v12.4s, v1.16b, v0.16b\n" ".inst 0x4e80a474 // smmla v20.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4bc // smmla v28.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xb0]\n" + "ldr q0, [x10, #0xb0]\n" ".inst 0x4e82a429 // smmla v9.4s, v1.16b, v2.16b\n" ".inst 0x4e82a471 // smmla v17.4s, v3.16b, v2.16b\n" ".inst 0x4e82a4b9 // smmla v25.4s, v5.16b, v2.16b\n" - "ldr q2, [x9, #0xc0]\n" + "ldr q2, [x10, #0xc0]\n" ".inst 0x4e80a42d // smmla v13.4s, v1.16b, v0.16b\n" ".inst 0x4e80a475 // smmla v21.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4bd // smmla v29.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xd0]\n" + "ldr q0, [x10, #0xd0]\n" ".inst 0x4e82a42a // smmla v10.4s, v1.16b, v2.16b\n" ".inst 0x4e82a472 // smmla v18.4s, v3.16b, v2.16b\n" ".inst 0x4e82a4ba // smmla v26.4s, v5.16b, v2.16b\n" - "ldr q2, [x9, #0xe0]\n" + "ldr q2, [x10, #0xe0]\n" ".inst 0x4e80a42e // smmla v14.4s, v1.16b, v0.16b\n" ".inst 0x4e80a476 // smmla v22.4s, v3.16b, v0.16b\n" ".inst 0x4e80a4be // smmla v30.4s, v5.16b, v0.16b\n" - "ldr q0, [x9, #0xf0]\n" - "add x9, x9, #0x100\n" + "ldr q0, [x10, #0xf0]\n" + "add x10, x10, #0x100\n" ".inst 0x4e82a42b // smmla v11.4s, v1.16b, v2.16b\n" ".inst 0x4e82a473 // smmla v19.4s, v3.16b, v2.16b\n" ".inst 0x4e82a4bb // smmla v27.4s, v5.16b, v2.16b\n" @@ -3020,44 +3021,44 @@ void a64_hybrid_s8qs_mmla_6x16 ( "cmp x27, #0x8\n" "blt 156f\n" "155:" // Height 6: Multiply loop: Odd block loop - "ldr d1, [x26], #0x8\n" - "ldr d0, [x25], #0x8\n" - "trn1 v4.2d, v1.2d, v0.2d\n" + "ldr d6, [x26], #0x8\n" + "ldr d4, [x25], #0x8\n" "sub x27, x27, #0x8\n" - "ldr d1, [x24], #0x8\n" - "ldr d0, [x23], #0x8\n" - "trn1 v3.2d, v1.2d, v0.2d\n" + "ldr d5, [x24], #0x8\n" + "ldr d3, [x23], #0x8\n" "cmp x27, #0x8\n" - "ldr d1, [x22], #0x8\n" + "ldr d2, [x22], #0x8\n" "ldr d0, [x21], #0x8\n" - "trn1 v2.2d, v1.2d, v0.2d\n" - "ldr q1, [x9, #0x0]\n" - "ldr q0, [x9, #0x10]\n" + "ldr q1, [x10, #0x0]\n" + "trn1 v4.2d, v6.2d, v4.2d\n" + "trn1 v3.2d, v5.2d, v3.2d\n" + "trn1 v2.2d, v2.2d, v0.2d\n" + "ldr q0, [x10, #0x10]\n" ".inst 0x4e81a488 // smmla v8.4s, v4.16b, v1.16b\n" ".inst 0x4e81a470 // smmla v16.4s, v3.16b, v1.16b\n" ".inst 0x4e81a458 // smmla v24.4s, v2.16b, v1.16b\n" - "ldr q1, [x9, #0x20]\n" + "ldr q1, [x10, #0x20]\n" ".inst 0x4e80a48c // smmla v12.4s, v4.16b, v0.16b\n" ".inst 0x4e80a474 // smmla v20.4s, v3.16b, v0.16b\n" ".inst 0x4e80a45c // smmla v28.4s, v2.16b, v0.16b\n" - "ldr q0, [x9, #0x30]\n" + "ldr q0, [x10, #0x30]\n" ".inst 0x4e81a489 // smmla v9.4s, v4.16b, v1.16b\n" ".inst 0x4e81a471 // smmla v17.4s, v3.16b, v1.16b\n" ".inst 0x4e81a459 // smmla v25.4s, v2.16b, v1.16b\n" - "ldr q1, [x9, #0x40]\n" + "ldr q1, [x10, #0x40]\n" ".inst 0x4e80a48d // smmla v13.4s, v4.16b, v0.16b\n" ".inst 0x4e80a475 // smmla v21.4s, v3.16b, v0.16b\n" ".inst 0x4e80a45d // smmla v29.4s, v2.16b, v0.16b\n" - "ldr q0, [x9, #0x50]\n" + "ldr q0, [x10, #0x50]\n" ".inst 0x4e81a48a // smmla v10.4s, v4.16b, v1.16b\n" ".inst 0x4e81a472 // smmla v18.4s, v3.16b, v1.16b\n" ".inst 0x4e81a45a // smmla v26.4s, v2.16b, v1.16b\n" - "ldr q1, [x9, #0x60]\n" + "ldr q1, [x10, #0x60]\n" ".inst 0x4e80a48e // smmla v14.4s, v4.16b, v0.16b\n" ".inst 0x4e80a476 // smmla v22.4s, v3.16b, v0.16b\n" ".inst 0x4e80a45e // smmla v30.4s, v2.16b, v0.16b\n" - "ldr q0, [x9, #0x70]\n" - "add x9, x9, #0x80\n" + "ldr q0, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e81a48b // smmla v11.4s, v4.16b, v1.16b\n" ".inst 0x4e81a473 // smmla v19.4s, v3.16b, v1.16b\n" ".inst 0x4e81a45b // smmla v27.4s, v2.16b, v1.16b\n" @@ -3122,37 +3123,37 @@ void a64_hybrid_s8qs_mmla_6x16 ( "ldr b5, [x22, #0x0]\n" "ldr b6, [x21, #0x0]\n" "160:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q7, [x9, #0x0]\n" + "ldr q7, [x10, #0x0]\n" "trn1 v2.2d, v1.2d, v2.2d\n" "trn1 v4.2d, v3.2d, v4.2d\n" - ".inst 0x4e87a448 // smmla v8.4s, v2.16b, v7.16b\n" "trn1 v3.2d, v5.2d, v6.2d\n" - "ldr q0, [x9, #0x10]\n" + "ldr q0, [x10, #0x10]\n" + ".inst 0x4e87a448 // smmla v8.4s, v2.16b, v7.16b\n" ".inst 0x4e87a490 // smmla v16.4s, v4.16b, v7.16b\n" ".inst 0x4e87a478 // smmla v24.4s, v3.16b, v7.16b\n" - "ldr q1, [x9, #0x20]\n" + "ldr q1, [x10, #0x20]\n" ".inst 0x4e80a44c // smmla v12.4s, v2.16b, v0.16b\n" ".inst 0x4e80a494 // smmla v20.4s, v4.16b, v0.16b\n" ".inst 0x4e80a47c // smmla v28.4s, v3.16b, v0.16b\n" - "ldr q0, [x9, #0x30]\n" + "ldr q0, [x10, #0x30]\n" ".inst 0x4e81a449 // smmla v9.4s, v2.16b, v1.16b\n" ".inst 0x4e81a491 // smmla v17.4s, v4.16b, v1.16b\n" ".inst 0x4e81a479 // smmla v25.4s, v3.16b, v1.16b\n" - "ldr q1, [x9, #0x40]\n" + "ldr q1, [x10, #0x40]\n" ".inst 0x4e80a44d // smmla v13.4s, v2.16b, v0.16b\n" ".inst 0x4e80a495 // smmla v21.4s, v4.16b, v0.16b\n" ".inst 0x4e80a47d // smmla v29.4s, v3.16b, v0.16b\n" - "ldr q0, [x9, #0x50]\n" + "ldr q0, [x10, #0x50]\n" ".inst 0x4e81a44a // smmla v10.4s, v2.16b, v1.16b\n" ".inst 0x4e81a492 // smmla v18.4s, v4.16b, v1.16b\n" ".inst 0x4e81a47a // smmla v26.4s, v3.16b, v1.16b\n" - "ldr q1, [x9, #0x60]\n" + "ldr q1, [x10, #0x60]\n" ".inst 0x4e80a44e // smmla v14.4s, v2.16b, v0.16b\n" ".inst 0x4e80a496 // smmla v22.4s, v4.16b, v0.16b\n" ".inst 0x4e80a47e // smmla v30.4s, v3.16b, v0.16b\n" - "ldr q0, [x9, #0x70]\n" + "ldr q0, [x10, #0x70]\n" + "add x10, x10, #0x80\n" ".inst 0x4e81a44b // smmla v11.4s, v2.16b, v1.16b\n" - "add x9, x9, #0x80\n" ".inst 0x4e81a493 // smmla v19.4s, v4.16b, v1.16b\n" ".inst 0x4e81a47b // smmla v27.4s, v3.16b, v1.16b\n" ".inst 0x4e80a44f // smmla v15.4s, v2.16b, v0.16b\n" @@ -3172,32 +3173,32 @@ void a64_hybrid_s8qs_mmla_6x16 ( "uzp1 v12.2d, v9.2d, v13.2d\n" "uzp2 v9.2d, v9.2d, v13.2d\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x11, x20\n" - "add x24, x25, x20\n" "uzp1 v13.2d, v10.2d, v14.2d\n" "uzp2 v10.2d, v10.2d, v14.2d\n" + "prfm pstl1keep, [x9, #0x0]\n" "uzp1 v14.2d, v11.2d, v15.2d\n" - "add x23, x24, x20\n" - "add x22, x23, x20\n" "uzp2 v11.2d, v11.2d, v15.2d\n" + "add x14, x14, #0x40\n" "uzp1 v15.2d, v16.2d, v20.2d\n" - "add x21, x22, x20\n" - "prfm pstl1keep, [x11, #0x0]\n" "uzp2 v16.2d, v16.2d, v20.2d\n" + "add x26, x9, x20\n" + "add x25, x26, x20\n" "uzp1 v20.2d, v17.2d, v21.2d\n" - "prfm pstl1keep, [x25, #0x0]\n" - "prfm pstl1keep, [x24, #0x0]\n" "uzp2 v17.2d, v17.2d, v21.2d\n" + "add x24, x25, x20\n" + "prfm pstl1keep, [x26, #0x0]\n" "uzp1 v21.2d, v18.2d, v22.2d\n" - "prfm pstl1keep, [x23, #0x0]\n" - "prfm pstl1keep, [x22, #0x0]\n" "uzp2 v18.2d, v18.2d, v22.2d\n" + "add x23, x24, x20\n" + "prfm pstl1keep, [x25, #0x0]\n" "uzp1 v22.2d, v19.2d, v23.2d\n" - "prfm pstl1keep, [x21, #0x0]\n" - "add x14, x14, #0x40\n" "uzp2 v19.2d, v19.2d, v23.2d\n" + "add x22, x23, x20\n" + "prfm pstl1keep, [x24, #0x0]\n" "uzp1 v23.2d, v24.2d, v28.2d\n" "uzp2 v24.2d, v24.2d, v28.2d\n" + "prfm pstl1keep, [x23, #0x0]\n" + "prfm pstl1keep, [x22, #0x0]\n" "uzp1 v28.2d, v25.2d, v29.2d\n" "uzp2 v25.2d, v25.2d, v29.2d\n" "uzp1 v29.2d, v26.2d, v30.2d\n" @@ -3205,10 +3206,10 @@ void a64_hybrid_s8qs_mmla_6x16 ( "uzp1 v30.2d, v27.2d, v31.2d\n" "uzp2 v27.2d, v27.2d, v31.2d\n" "mov v31.16b, v2.16b\n" - "add v31.4s, v31.4s, v4.4s\n" "add v12.4s, v12.4s, v3.4s\n" "add v13.4s, v13.4s, v1.4s\n" "add v14.4s, v14.4s, v0.4s\n" + "add v31.4s, v31.4s, v4.4s\n" "add v8.4s, v8.4s, v4.4s\n" "add v9.4s, v9.4s, v3.4s\n" "add v10.4s, v10.4s, v1.4s\n" @@ -3242,9 +3243,9 @@ void a64_hybrid_s8qs_mmla_6x16 ( "add x13, x13, #0x40\n" "b 163f\n" "162:" // Height 6: per layer parameters - "add x20, %x[qp], %[per_layer_right_shift]\n" - "ld1r { v0.4s }, [x20]\n" + "add x21, %x[qp], %[per_layer_right_shift]\n" "add x20, %x[qp], %[per_layer_mul]\n" + "ld1r { v0.4s }, [x21]\n" "ld1r { v4.4s }, [x20]\n" "mov v1.16b, v0.16b\n" "mov v5.16b, v4.16b\n" @@ -3287,11 +3288,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v5.4s, v5.4s, #0x1f\n" "sshr v4.4s, v4.4s, #0x1f\n" "sqadd v31.4s, v31.4s, v7.4s\n" + "and v7.16b, v8.16b, v0.16b\n" "sqadd v12.4s, v12.4s, v6.4s\n" + "and v6.16b, v9.16b, v1.16b\n" "sqadd v13.4s, v13.4s, v5.4s\n" "sqadd v14.4s, v14.4s, v4.4s\n" - "and v7.16b, v8.16b, v0.16b\n" - "and v6.16b, v9.16b, v1.16b\n" "and v5.16b, v10.16b, v2.16b\n" "and v4.16b, v11.16b, v3.16b\n" "sshr v7.4s, v7.4s, #0x1f\n" @@ -3299,11 +3300,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v5.4s, v5.4s, #0x1f\n" "sshr v4.4s, v4.4s, #0x1f\n" "sqadd v8.4s, v8.4s, v7.4s\n" + "and v7.16b, v15.16b, v0.16b\n" "sqadd v9.4s, v9.4s, v6.4s\n" + "and v6.16b, v20.16b, v1.16b\n" "sqadd v10.4s, v10.4s, v5.4s\n" "sqadd v11.4s, v11.4s, v4.4s\n" - "and v7.16b, v15.16b, v0.16b\n" - "and v6.16b, v20.16b, v1.16b\n" "and v5.16b, v21.16b, v2.16b\n" "and v4.16b, v22.16b, v3.16b\n" "sshr v7.4s, v7.4s, #0x1f\n" @@ -3311,11 +3312,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v5.4s, v5.4s, #0x1f\n" "sshr v4.4s, v4.4s, #0x1f\n" "sqadd v15.4s, v15.4s, v7.4s\n" + "and v7.16b, v16.16b, v0.16b\n" "sqadd v20.4s, v20.4s, v6.4s\n" + "and v6.16b, v17.16b, v1.16b\n" "sqadd v21.4s, v21.4s, v5.4s\n" "sqadd v22.4s, v22.4s, v4.4s\n" - "and v7.16b, v16.16b, v0.16b\n" - "and v6.16b, v17.16b, v1.16b\n" "and v5.16b, v18.16b, v2.16b\n" "and v4.16b, v19.16b, v3.16b\n" "sshr v7.4s, v7.4s, #0x1f\n" @@ -3323,11 +3324,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v5.4s, v5.4s, #0x1f\n" "sshr v4.4s, v4.4s, #0x1f\n" "sqadd v16.4s, v16.4s, v7.4s\n" + "and v7.16b, v23.16b, v0.16b\n" "sqadd v17.4s, v17.4s, v6.4s\n" + "and v6.16b, v28.16b, v1.16b\n" "sqadd v18.4s, v18.4s, v5.4s\n" "sqadd v19.4s, v19.4s, v4.4s\n" - "and v7.16b, v23.16b, v0.16b\n" - "and v6.16b, v28.16b, v1.16b\n" "and v5.16b, v29.16b, v2.16b\n" "and v4.16b, v30.16b, v3.16b\n" "sshr v7.4s, v7.4s, #0x1f\n" @@ -3335,11 +3336,11 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sshr v5.4s, v5.4s, #0x1f\n" "sshr v4.4s, v4.4s, #0x1f\n" "sqadd v23.4s, v23.4s, v7.4s\n" + "and v7.16b, v24.16b, v0.16b\n" "sqadd v28.4s, v28.4s, v6.4s\n" + "and v6.16b, v25.16b, v1.16b\n" "sqadd v29.4s, v29.4s, v5.4s\n" "sqadd v30.4s, v30.4s, v4.4s\n" - "and v7.16b, v24.16b, v0.16b\n" - "and v6.16b, v25.16b, v1.16b\n" "and v5.16b, v26.16b, v2.16b\n" "and v4.16b, v27.16b, v3.16b\n" "sshr v7.4s, v7.4s, #0x1f\n" @@ -3351,21 +3352,21 @@ void a64_hybrid_s8qs_mmla_6x16 ( "sqadd v26.4s, v26.4s, v5.4s\n" "sqadd v27.4s, v27.4s, v4.4s\n" "164:" // Height 6: no shift correction - "add x20, %x[qp], %[c_offset]\n" - "ld1r { v6.4s }, [x20]\n" + "add x21, %x[qp], %[c_offset]\n" "srshl v31.4s, v31.4s, v0.4s\n" "srshl v12.4s, v12.4s, v1.4s\n" - "srshl v13.4s, v13.4s, v2.4s\n" - "srshl v14.4s, v14.4s, v3.4s\n" "add x20, %x[qp], %[maxval]\n" + "ld1r { v6.4s }, [x21]\n" "ld1r { v5.4s }, [x20]\n" + "srshl v13.4s, v13.4s, v2.4s\n" + "srshl v14.4s, v14.4s, v3.4s\n" "srshl v8.4s, v8.4s, v0.4s\n" "srshl v9.4s, v9.4s, v1.4s\n" "add x20, %x[qp], %[minval]\n" + "cmp x11, #0x10\n" "ld1r { v4.4s }, [x20]\n" "srshl v10.4s, v10.4s, v2.4s\n" "srshl v11.4s, v11.4s, v3.4s\n" - "cmp x10, #0x10\n" "srshl v15.4s, v15.4s, v0.4s\n" "srshl v20.4s, v20.4s, v1.4s\n" "srshl v21.4s, v21.4s, v2.4s\n" @@ -3473,136 +3474,136 @@ void a64_hybrid_s8qs_mmla_6x16 ( "uzp1 v23.16b, v23.16b, v18.16b\n" "uzp1 v24.16b, v24.16b, v17.16b\n" "bge 173f\n" - "tbz x10, #3, 168f\n" - "str d31, [x11], #0x8\n" - "str d8, [x25], #0x8\n" - "str d15, [x24], #0x8\n" - "str d16, [x23], #0x8\n" - "str d23, [x22], #0x8\n" - "str d24, [x21], #0x8\n" - "tbz x10, #2, 166f\n" - "st1 { v31.s }[2], [x11], #0x4\n" - "st1 { v8.s }[2], [x25], #0x4\n" - "st1 { v15.s }[2], [x24], #0x4\n" - "st1 { v16.s }[2], [x23], #0x4\n" - "st1 { v23.s }[2], [x22], #0x4\n" - "st1 { v24.s }[2], [x21], #0x4\n" - "tbz x10, #1, 165f\n" - "st1 { v31.h }[6], [x11], #0x2\n" - "st1 { v8.h }[6], [x25], #0x2\n" - "st1 { v15.h }[6], [x24], #0x2\n" - "st1 { v16.h }[6], [x23], #0x2\n" - "st1 { v23.h }[6], [x22], #0x2\n" - "st1 { v24.h }[6], [x21], #0x2\n" - "tbz x10, #0, 172f\n" - "st1 { v31.b }[14], [x11]\n" - "st1 { v8.b }[14], [x25]\n" - "st1 { v15.b }[14], [x24]\n" - "st1 { v16.b }[14], [x23]\n" - "st1 { v23.b }[14], [x22]\n" - "st1 { v24.b }[14], [x21]\n" + "tbz x11, #3, 168f\n" + "str d31, [x9], #0x8\n" + "str d8, [x26], #0x8\n" + "str d15, [x25], #0x8\n" + "str d16, [x24], #0x8\n" + "str d23, [x23], #0x8\n" + "str d24, [x22], #0x8\n" + "tbz x11, #2, 166f\n" + "st1 { v31.s }[2], [x9], #0x4\n" + "st1 { v8.s }[2], [x26], #0x4\n" + "st1 { v15.s }[2], [x25], #0x4\n" + "st1 { v16.s }[2], [x24], #0x4\n" + "st1 { v23.s }[2], [x23], #0x4\n" + "st1 { v24.s }[2], [x22], #0x4\n" + "tbz x11, #1, 165f\n" + "st1 { v31.h }[6], [x9], #0x2\n" + "st1 { v8.h }[6], [x26], #0x2\n" + "st1 { v15.h }[6], [x25], #0x2\n" + "st1 { v16.h }[6], [x24], #0x2\n" + "st1 { v23.h }[6], [x23], #0x2\n" + "st1 { v24.h }[6], [x22], #0x2\n" + "tbz x11, #0, 172f\n" + "st1 { v31.b }[14], [x9]\n" + "st1 { v8.b }[14], [x26]\n" + "st1 { v15.b }[14], [x25]\n" + "st1 { v16.b }[14], [x24]\n" + "st1 { v23.b }[14], [x23]\n" + "st1 { v24.b }[14], [x22]\n" "b 172f\n" "165:" // Height 6: Partial direct writeback: partial_1_12 - "tbz x10, #0, 172f\n" - "st1 { v31.b }[12], [x11]\n" - "st1 { v8.b }[12], [x25]\n" - "st1 { v15.b }[12], [x24]\n" - "st1 { v16.b }[12], [x23]\n" - "st1 { v23.b }[12], [x22]\n" - "st1 { v24.b }[12], [x21]\n" + "tbz x11, #0, 172f\n" + "st1 { v31.b }[12], [x9]\n" + "st1 { v8.b }[12], [x26]\n" + "st1 { v15.b }[12], [x25]\n" + "st1 { v16.b }[12], [x24]\n" + "st1 { v23.b }[12], [x23]\n" + "st1 { v24.b }[12], [x22]\n" "b 172f\n" "166:" // Height 6: Partial direct writeback: partial_2_8 - "tbz x10, #1, 167f\n" - "st1 { v31.h }[4], [x11], #0x2\n" - "st1 { v8.h }[4], [x25], #0x2\n" - "st1 { v15.h }[4], [x24], #0x2\n" - "st1 { v16.h }[4], [x23], #0x2\n" - "st1 { v23.h }[4], [x22], #0x2\n" - "st1 { v24.h }[4], [x21], #0x2\n" - "tbz x10, #0, 172f\n" - "st1 { v31.b }[10], [x11]\n" - "st1 { v8.b }[10], [x25]\n" - "st1 { v15.b }[10], [x24]\n" - "st1 { v16.b }[10], [x23]\n" - "st1 { v23.b }[10], [x22]\n" - "st1 { v24.b }[10], [x21]\n" + "tbz x11, #1, 167f\n" + "st1 { v31.h }[4], [x9], #0x2\n" + "st1 { v8.h }[4], [x26], #0x2\n" + "st1 { v15.h }[4], [x25], #0x2\n" + "st1 { v16.h }[4], [x24], #0x2\n" + "st1 { v23.h }[4], [x23], #0x2\n" + "st1 { v24.h }[4], [x22], #0x2\n" + "tbz x11, #0, 172f\n" + "st1 { v31.b }[10], [x9]\n" + "st1 { v8.b }[10], [x26]\n" + "st1 { v15.b }[10], [x25]\n" + "st1 { v16.b }[10], [x24]\n" + "st1 { v23.b }[10], [x23]\n" + "st1 { v24.b }[10], [x22]\n" "b 172f\n" "167:" // Height 6: Partial direct writeback: partial_1_8 - "tbz x10, #0, 172f\n" - "st1 { v31.b }[8], [x11]\n" - "st1 { v8.b }[8], [x25]\n" - "st1 { v15.b }[8], [x24]\n" - "st1 { v16.b }[8], [x23]\n" - "st1 { v23.b }[8], [x22]\n" - "st1 { v24.b }[8], [x21]\n" + "tbz x11, #0, 172f\n" + "st1 { v31.b }[8], [x9]\n" + "st1 { v8.b }[8], [x26]\n" + "st1 { v15.b }[8], [x25]\n" + "st1 { v16.b }[8], [x24]\n" + "st1 { v23.b }[8], [x23]\n" + "st1 { v24.b }[8], [x22]\n" "b 172f\n" "168:" // Height 6: Partial direct writeback: partial_4_0 - "tbz x10, #2, 170f\n" - "str s31, [x11], #0x4\n" - "str s8, [x25], #0x4\n" - "str s15, [x24], #0x4\n" - "str s16, [x23], #0x4\n" - "str s23, [x22], #0x4\n" - "str s24, [x21], #0x4\n" - "tbz x10, #1, 169f\n" - "st1 { v31.h }[2], [x11], #0x2\n" - "st1 { v8.h }[2], [x25], #0x2\n" - "st1 { v15.h }[2], [x24], #0x2\n" - "st1 { v16.h }[2], [x23], #0x2\n" - "st1 { v23.h }[2], [x22], #0x2\n" - "st1 { v24.h }[2], [x21], #0x2\n" - "tbz x10, #0, 172f\n" - "st1 { v31.b }[6], [x11]\n" - "st1 { v8.b }[6], [x25]\n" - "st1 { v15.b }[6], [x24]\n" - "st1 { v16.b }[6], [x23]\n" - "st1 { v23.b }[6], [x22]\n" - "st1 { v24.b }[6], [x21]\n" + "tbz x11, #2, 170f\n" + "str s31, [x9], #0x4\n" + "str s8, [x26], #0x4\n" + "str s15, [x25], #0x4\n" + "str s16, [x24], #0x4\n" + "str s23, [x23], #0x4\n" + "str s24, [x22], #0x4\n" + "tbz x11, #1, 169f\n" + "st1 { v31.h }[2], [x9], #0x2\n" + "st1 { v8.h }[2], [x26], #0x2\n" + "st1 { v15.h }[2], [x25], #0x2\n" + "st1 { v16.h }[2], [x24], #0x2\n" + "st1 { v23.h }[2], [x23], #0x2\n" + "st1 { v24.h }[2], [x22], #0x2\n" + "tbz x11, #0, 172f\n" + "st1 { v31.b }[6], [x9]\n" + "st1 { v8.b }[6], [x26]\n" + "st1 { v15.b }[6], [x25]\n" + "st1 { v16.b }[6], [x24]\n" + "st1 { v23.b }[6], [x23]\n" + "st1 { v24.b }[6], [x22]\n" "b 172f\n" "169:" // Height 6: Partial direct writeback: partial_1_4 - "tbz x10, #0, 172f\n" - "st1 { v31.b }[4], [x11]\n" - "st1 { v8.b }[4], [x25]\n" - "st1 { v15.b }[4], [x24]\n" - "st1 { v16.b }[4], [x23]\n" - "st1 { v23.b }[4], [x22]\n" - "st1 { v24.b }[4], [x21]\n" + "tbz x11, #0, 172f\n" + "st1 { v31.b }[4], [x9]\n" + "st1 { v8.b }[4], [x26]\n" + "st1 { v15.b }[4], [x25]\n" + "st1 { v16.b }[4], [x24]\n" + "st1 { v23.b }[4], [x23]\n" + "st1 { v24.b }[4], [x22]\n" "b 172f\n" "170:" // Height 6: Partial direct writeback: partial_2_0 - "tbz x10, #1, 171f\n" - "str h31, [x11], #0x2\n" - "str h8, [x25], #0x2\n" - "str h15, [x24], #0x2\n" - "str h16, [x23], #0x2\n" - "str h23, [x22], #0x2\n" - "str h24, [x21], #0x2\n" - "tbz x10, #0, 172f\n" - "st1 { v31.b }[2], [x11]\n" - "st1 { v8.b }[2], [x25]\n" - "st1 { v15.b }[2], [x24]\n" - "st1 { v16.b }[2], [x23]\n" - "st1 { v23.b }[2], [x22]\n" - "st1 { v24.b }[2], [x21]\n" + "tbz x11, #1, 171f\n" + "str h31, [x9], #0x2\n" + "str h8, [x26], #0x2\n" + "str h15, [x25], #0x2\n" + "str h16, [x24], #0x2\n" + "str h23, [x23], #0x2\n" + "str h24, [x22], #0x2\n" + "tbz x11, #0, 172f\n" + "st1 { v31.b }[2], [x9]\n" + "st1 { v8.b }[2], [x26]\n" + "st1 { v15.b }[2], [x25]\n" + "st1 { v16.b }[2], [x24]\n" + "st1 { v23.b }[2], [x23]\n" + "st1 { v24.b }[2], [x22]\n" "b 172f\n" "171:" // Height 6: Partial direct writeback: partial_1_0 - "str b31, [x11, #0x0]\n" - "str b8, [x25, #0x0]\n" - "str b15, [x24, #0x0]\n" - "str b16, [x23, #0x0]\n" - "str b23, [x22, #0x0]\n" - "str b24, [x21, #0x0]\n" + "str b31, [x9, #0x0]\n" + "str b8, [x26, #0x0]\n" + "str b15, [x25, #0x0]\n" + "str b16, [x24, #0x0]\n" + "str b23, [x23, #0x0]\n" + "str b24, [x22, #0x0]\n" "172:" // Height 6: Partial direct writeback: Done "b 174f\n" "173:" // Height 6: Full writeback - "str q31, [x11, #0x0]\n" - "add x11, x11, #0x10\n" - "str q8, [x25, #0x0]\n" - "str q15, [x24, #0x0]\n" - "str q16, [x23, #0x0]\n" - "str q23, [x22, #0x0]\n" - "str q24, [x21, #0x0]\n" + "str q31, [x9, #0x0]\n" + "add x9, x9, #0x10\n" + "str q8, [x26, #0x0]\n" + "str q15, [x25, #0x0]\n" + "str q16, [x24, #0x0]\n" + "str q23, [x23, #0x0]\n" + "str q24, [x22, #0x0]\n" "174:" // Height 6: Writeback done - "subs x10, x10, #0x10\n" + "subs x11, x11, #0x10\n" "bgt 147b\n" "subs %x[M], %x[M], #0x6\n" "beq 176f\n" @@ -3616,8 +3617,8 @@ void a64_hybrid_s8qs_mmla_6x16 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "176:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) - : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp) + : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr) + : [args_ptr] "r" (&ka), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [flags] "r" (flags), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_multiplier_ptr] "I" (offsetof(KernelArgs, multiplier_ptr)), [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_shift_ptr] "I" (offsetof(KernelArgs, shift_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp) : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |