diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp | 1489 |
1 files changed, 746 insertions, 743 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp index fc323ea4fc..4494e2ac13 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_bf16fp32_dot_6x16/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021, 2023 Arm Limited. + * Copyright (c) 2019-2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -48,18 +48,19 @@ void a64_hybrid_bf16fp32_dot_6x16 ( size_t output_offset = {}; size_t input_initial_col = {}; size_t input_offset = {}; + void *output_ptr = {}; + const float *bias = {}; } ka; unsigned long flags=0; - void *output_ptr; void *input_ptr; if (output_arg.is_indirect) { - output_ptr=(void *)(output_arg.indirect.ptr); + ka.output_ptr=(void *)(output_arg.indirect.ptr); ka.output_offset=output_arg.indirect.offset; flags |= 0x4; } else { - output_ptr=(void *)(output_arg.direct.base); + ka.output_ptr=(void *)(output_arg.direct.base); ka.output_offset=output_arg.direct.stride; } @@ -80,6 +81,7 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ka.string_lengths = string_lengths; ka.N = N; ka.B_ptr = B_ptr; + ka.bias = bias; switch(act.type) { default: case Activation::Type::None: @@ -102,10 +104,10 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "cmp %x[M], #0x2\n" "bgt 71f\n" "beq 36f\n" - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "2:" // Height 1: Column loop "cbz x12, 3f\n" "ldr q8, [x12, #0x0]\n" @@ -188,8 +190,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "mov x28, #0x0\n" "15:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" "tbz %x[flags], #3, 16f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -213,6 +215,10 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr q17, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" "ldr q16, [x10, #0x30]\n" + "sub x27, x27, #0x8\n" + "add x26, x26, #0x10\n" + "cmp x27, #0x10\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f40f22a // bfdot v10.4s, v17.8h, v0.h[0]\n" "ldr q17, [x10, #0x40]\n" ".inst 0x4f40f20b // bfdot v11.4s, v16.8h, v0.h[0]\n" @@ -237,22 +243,21 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr q17, [x10, #0xe0]\n" ".inst 0x4f60fa09 // bfdot v9.4s, v16.8h, v0.h[3]\n" "ldr q16, [x10, #0xf0]\n" - "sub x27, x27, #0x8\n" - "add x26, x26, #0x10\n" + "add x10, x10, #0x100\n" ".inst 0x4f60fa2a // bfdot v10.4s, v17.8h, v0.h[3]\n" + "ldr q6, [x10, #0x0]\n" ".inst 0x4f60fa0b // bfdot v11.4s, v16.8h, v0.h[3]\n" "ldr q0, [x26, #0x0]\n" - "cmp x27, #0x10\n" - "add x10, x10, #0x100\n" - "ldr q6, [x10, #0x0]\n" "ldr q7, [x10, #0x10]\n" - "prfm pldl1keep, [x26, #0x80]\n" "bge 18b\n" "19:" // Height 1: Multiply loop: Single iteration only ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" "ldr q17, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" "ldr q16, [x10, #0x30]\n" + "add x26, x26, #0x10\n" + "sub x27, x27, #0x8\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f40f22a // bfdot v10.4s, v17.8h, v0.h[0]\n" "ldr q17, [x10, #0x40]\n" ".inst 0x4f40f20b // bfdot v11.4s, v16.8h, v0.h[0]\n" @@ -277,29 +282,26 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr q17, [x10, #0xe0]\n" ".inst 0x4f60fa09 // bfdot v9.4s, v16.8h, v0.h[3]\n" "ldr q16, [x10, #0xf0]\n" - "add x26, x26, #0x10\n" - "sub x27, x27, #0x8\n" + "add x10, x10, #0x100\n" ".inst 0x4f60fa2a // bfdot v10.4s, v17.8h, v0.h[3]\n" ".inst 0x4f60fa0b // bfdot v11.4s, v16.8h, v0.h[3]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "add x10, x10, #0x100\n" "20:" // Height 1: Multiply loop: Main loop skip "cbz x27, 24f\n" "cmp x27, #0x2\n" "blt 22f\n" "21:" // Height 1: Multiply loop: Odd block loop "ldr s18, [x26], #0x4\n" - "ldr q16, [x10, #0x0]\n" - ".inst 0x4f52f208 // bfdot v8.4s, v16.8h, v18.h[0]\n" + "ldr q17, [x10, #0x0]\n" "sub x27, x27, #0x2\n" "ldr q16, [x10, #0x10]\n" + "cmp x27, #0x2\n" + ".inst 0x4f52f228 // bfdot v8.4s, v17.8h, v18.h[0]\n" "ldr q17, [x10, #0x20]\n" ".inst 0x4f52f209 // bfdot v9.4s, v16.8h, v18.h[0]\n" - "cmp x27, #0x2\n" "ldr q16, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f52f22a // bfdot v10.4s, v17.8h, v18.h[0]\n" ".inst 0x4f52f20b // bfdot v11.4s, v16.8h, v18.h[0]\n" - "add x10, x10, #0x40\n" "bge 21b\n" "22:" // Height 1: Multiply loop: Skip odd blocks "cbz x27, 24f\n" @@ -308,12 +310,12 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr q17, [x10, #0x0]\n" "ldr q16, [x10, #0x10]\n" ".inst 0x4f40f228 // bfdot v8.4s, v17.8h, v0.h[0]\n" - ".inst 0x4f40f209 // bfdot v9.4s, v16.8h, v0.h[0]\n" "ldr q17, [x10, #0x20]\n" + ".inst 0x4f40f209 // bfdot v9.4s, v16.8h, v0.h[0]\n" "ldr q16, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f22a // bfdot v10.4s, v17.8h, v0.h[0]\n" ".inst 0x4f40f20b // bfdot v11.4s, v16.8h, v0.h[0]\n" - "add x10, x10, #0x40\n" "24:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -321,9 +323,9 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "bne 15b\n" "prfm pstl1keep, [x9, #0x0]\n" "tbz %x[flags], #1, 25f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v17.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v8.4s, v8.4s, v17.4s\n" "fmin v9.4s, v9.4s, v17.4s\n" @@ -393,95 +395,95 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "bgt 2b\n" "b 212f\n" "36:" // Height 2 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "37:" // Height 2: Column loop "cbz x12, 38f\n" "ldr q8, [x12, #0x0]\n" "ldr q9, [x12, #0x10]\n" - "mov v12.16b, v8.16b\n" - "mov v13.16b, v9.16b\n" "ldr q10, [x12, #0x20]\n" "ldr q11, [x12, #0x30]\n" + "add x12, x12, #0x40\n" + "mov v12.16b, v8.16b\n" + "mov v13.16b, v9.16b\n" "mov v14.16b, v10.16b\n" "mov v15.16b, v11.16b\n" - "add x12, x12, #0x40\n" "b 49f\n" "38:" // Height 2: no bias "tbz %x[flags], #0, 48f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" "cmp x11, #0x10\n" - "add x25, x9, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" "bge 47f\n" "tbz x11, #3, 42f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" "ld1 { v9.4s }, [x9], #0x10\n" - "ld1 { v13.4s }, [x25], #0x10\n" + "ld1 { v13.4s }, [x26], #0x10\n" "tbz x11, #2, 40f\n" "ld1 { v10.4s }, [x9], #0x10\n" - "ld1 { v14.4s }, [x25], #0x10\n" + "ld1 { v14.4s }, [x26], #0x10\n" "tbz x11, #1, 39f\n" "ldr d11, [x9], #0x8\n" - "ldr d15, [x25], #0x8\n" + "ldr d15, [x26], #0x8\n" "mov x20, #0x38\n" "tbz x11, #0, 46f\n" "ld1 { v11.s }[2], [x9]\n" - "ld1 { v15.s }[2], [x25]\n" + "ld1 { v15.s }[2], [x26]\n" "b 46f\n" "39:" // Height 2: Partial accumulate: partial_1_12 "mov x20, #0x30\n" "tbz x11, #0, 46f\n" "ldr s11, [x9, #0x0]\n" - "ldr s15, [x25, #0x0]\n" + "ldr s15, [x26, #0x0]\n" "b 46f\n" "40:" // Height 2: Partial accumulate: partial_2_8 "tbz x11, #1, 41f\n" "ldr d10, [x9], #0x8\n" - "ldr d14, [x25], #0x8\n" + "ldr d14, [x26], #0x8\n" "mov x20, #0x28\n" "tbz x11, #0, 46f\n" "ld1 { v10.s }[2], [x9]\n" - "ld1 { v14.s }[2], [x25]\n" + "ld1 { v14.s }[2], [x26]\n" "b 46f\n" "41:" // Height 2: Partial accumulate: partial_1_8 "mov x20, #0x20\n" "tbz x11, #0, 46f\n" "ldr s10, [x9, #0x0]\n" - "ldr s14, [x25, #0x0]\n" + "ldr s14, [x26, #0x0]\n" "b 46f\n" "42:" // Height 2: Partial accumulate: partial_4_0 "tbz x11, #2, 44f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" "tbz x11, #1, 43f\n" "ldr d9, [x9], #0x8\n" - "ldr d13, [x25], #0x8\n" + "ldr d13, [x26], #0x8\n" "mov x20, #0x18\n" "tbz x11, #0, 46f\n" "ld1 { v9.s }[2], [x9]\n" - "ld1 { v13.s }[2], [x25]\n" + "ld1 { v13.s }[2], [x26]\n" "b 46f\n" "43:" // Height 2: Partial accumulate: partial_1_4 "mov x20, #0x10\n" "tbz x11, #0, 46f\n" "ldr s9, [x9, #0x0]\n" - "ldr s13, [x25, #0x0]\n" + "ldr s13, [x26, #0x0]\n" "b 46f\n" "44:" // Height 2: Partial accumulate: partial_2_0 "tbz x11, #1, 45f\n" "ldr d8, [x9], #0x8\n" - "ldr d12, [x25], #0x8\n" + "ldr d12, [x26], #0x8\n" "mov x20, #0x8\n" "tbz x11, #0, 46f\n" "ld1 { v8.s }[2], [x9]\n" - "ld1 { v12.s }[2], [x25]\n" + "ld1 { v12.s }[2], [x26]\n" "b 46f\n" "45:" // Height 2: Partial accumulate: partial_1_0 "ldr s8, [x9, #0x0]\n" - "ldr s12, [x25, #0x0]\n" + "ldr s12, [x26, #0x0]\n" "mov x20, #0x0\n" "46:" // Height 2: Partial accumulate: Done "sub x9, x9, x20\n" @@ -491,10 +493,10 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr q9, [x9, #0x10]\n" "ldr q10, [x9, #0x20]\n" "ldr q11, [x9, #0x30]\n" - "ldr q12, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "ldr q14, [x25, #0x20]\n" - "ldr q15, [x25, #0x30]\n" + "ldr q12, [x26, #0x0]\n" + "ldr q13, [x26, #0x10]\n" + "ldr q14, [x26, #0x20]\n" + "ldr q15, [x26, #0x30]\n" "b 49f\n" "48:" // Height 2: no accumulate "movi v8.16b, #0x0\n" @@ -509,8 +511,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "mov x28, #0x0\n" "50:" // 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, 51f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -542,22 +544,22 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" "ldr q16, [x10, #0x30]\n" "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" + "cmp x27, #0x10\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f40f22a // bfdot v10.4s, v17.8h, v0.h[0]\n" ".inst 0x4f41f22e // bfdot v14.4s, v17.8h, v1.h[0]\n" "ldr q17, [x10, #0x40]\n" - "add x25, x25, #0x10\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f40f20b // bfdot v11.4s, v16.8h, v0.h[0]\n" ".inst 0x4f41f20f // bfdot v15.4s, v16.8h, v1.h[0]\n" "ldr q16, [x10, #0x50]\n" - "cmp x27, #0x10\n" ".inst 0x4f60f228 // bfdot v8.4s, v17.8h, v0.h[1]\n" ".inst 0x4f61f22c // bfdot v12.4s, v17.8h, v1.h[1]\n" "ldr q17, [x10, #0x60]\n" - "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f60f209 // bfdot v9.4s, v16.8h, v0.h[1]\n" ".inst 0x4f61f20d // bfdot v13.4s, v16.8h, v1.h[1]\n" "ldr q16, [x10, #0x70]\n" - "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f60f22a // bfdot v10.4s, v17.8h, v0.h[1]\n" ".inst 0x4f61f22e // bfdot v14.4s, v17.8h, v1.h[1]\n" "ldr q17, [x10, #0x80]\n" @@ -601,18 +603,18 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" "ldr q16, [x10, #0x30]\n" "add x25, x25, #0x10\n" + "sub x27, x27, #0x8\n" + "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f40f22a // bfdot v10.4s, v17.8h, v0.h[0]\n" ".inst 0x4f41f22e // bfdot v14.4s, v17.8h, v1.h[0]\n" "ldr q17, [x10, #0x40]\n" - "sub x27, x27, #0x8\n" ".inst 0x4f40f20b // bfdot v11.4s, v16.8h, v0.h[0]\n" ".inst 0x4f41f20f // bfdot v15.4s, v16.8h, v1.h[0]\n" "ldr q16, [x10, #0x50]\n" - "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f60f228 // bfdot v8.4s, v17.8h, v0.h[1]\n" ".inst 0x4f61f22c // bfdot v12.4s, v17.8h, v1.h[1]\n" "ldr q17, [x10, #0x60]\n" - "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f60f209 // bfdot v9.4s, v16.8h, v0.h[1]\n" ".inst 0x4f61f20d // bfdot v13.4s, v16.8h, v1.h[1]\n" "ldr q16, [x10, #0x70]\n" @@ -653,18 +655,18 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr s19, [x26], #0x4\n" "ldr s18, [x25], #0x4\n" "sub x27, x27, #0x2\n" - "cmp x27, #0x2\n" "ldr q17, [x10, #0x0]\n" "ldr q16, [x10, #0x10]\n" + "cmp x27, #0x2\n" ".inst 0x4f53f228 // bfdot v8.4s, v17.8h, v19.h[0]\n" ".inst 0x4f52f22c // bfdot v12.4s, v17.8h, v18.h[0]\n" "ldr q17, [x10, #0x20]\n" ".inst 0x4f53f209 // bfdot v9.4s, v16.8h, v19.h[0]\n" ".inst 0x4f52f20d // bfdot v13.4s, v16.8h, v18.h[0]\n" "ldr q16, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f53f22a // bfdot v10.4s, v17.8h, v19.h[0]\n" ".inst 0x4f52f22e // bfdot v14.4s, v17.8h, v18.h[0]\n" - "add x10, x10, #0x40\n" ".inst 0x4f53f20b // bfdot v11.4s, v16.8h, v19.h[0]\n" ".inst 0x4f52f20f // bfdot v15.4s, v16.8h, v18.h[0]\n" "bge 56b\n" @@ -681,9 +683,9 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f40f209 // bfdot v9.4s, v16.8h, v0.h[0]\n" ".inst 0x4f41f20d // bfdot v13.4s, v16.8h, v1.h[0]\n" "ldr q16, [x10, #0x30]\n" + "add x10, x10, #0x40\n" ".inst 0x4f40f22a // bfdot v10.4s, v17.8h, v0.h[0]\n" ".inst 0x4f41f22e // bfdot v14.4s, v17.8h, v1.h[0]\n" - "add x10, x10, #0x40\n" ".inst 0x4f40f20b // bfdot v11.4s, v16.8h, v0.h[0]\n" ".inst 0x4f41f20f // bfdot v15.4s, v16.8h, v1.h[0]\n" "59:" // Height 2: Multiply loop: No odd multiplies @@ -692,13 +694,13 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "cmp x28, x20\n" "bne 50b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" "prfm pstl1keep, [x9, #0x0]\n" - "prfm pstl1keep, [x25, #0x0]\n" + "add x26, x9, x20, LSL #2\n" + "prfm pstl1keep, [x26, #0x0]\n" "tbz %x[flags], #1, 60f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v17.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v17.4s }, [x21]\n" "ld1r { v16.4s }, [x20]\n" "fmin v8.4s, v8.4s, v17.4s\n" "fmin v9.4s, v9.4s, v17.4s\n" @@ -722,63 +724,63 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "tbz x11, #3, 64f\n" "st1 { v8.4s }, [x9], #0x10\n" "st1 { v9.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v13.4s }, [x25], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v13.4s }, [x26], #0x10\n" "tbz x11, #2, 62f\n" "st1 { v10.4s }, [x9], #0x10\n" - "st1 { v14.4s }, [x25], #0x10\n" + "st1 { v14.4s }, [x26], #0x10\n" "tbz x11, #1, 61f\n" "str d11, [x9], #0x8\n" - "str d15, [x25], #0x8\n" + "str d15, [x26], #0x8\n" "tbz x11, #0, 68f\n" "st1 { v11.s }[2], [x9]\n" - "st1 { v15.s }[2], [x25]\n" + "st1 { v15.s }[2], [x26]\n" "b 68f\n" "61:" // Height 2: Partial direct writeback: partial_1_12 "tbz x11, #0, 68f\n" "str s11, [x9, #0x0]\n" - "str s15, [x25, #0x0]\n" + "str s15, [x26, #0x0]\n" "b 68f\n" "62:" // Height 2: Partial direct writeback: partial_2_8 "tbz x11, #1, 63f\n" "str d10, [x9], #0x8\n" - "str d14, [x25], #0x8\n" + "str d14, [x26], #0x8\n" "tbz x11, #0, 68f\n" "st1 { v10.s }[2], [x9]\n" - "st1 { v14.s }[2], [x25]\n" + "st1 { v14.s }[2], [x26]\n" "b 68f\n" "63:" // Height 2: Partial direct writeback: partial_1_8 "tbz x11, #0, 68f\n" "str s10, [x9, #0x0]\n" - "str s14, [x25, #0x0]\n" + "str s14, [x26, #0x0]\n" "b 68f\n" "64:" // Height 2: Partial direct writeback: partial_4_0 "tbz x11, #2, 66f\n" "st1 { v8.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" "tbz x11, #1, 65f\n" "str d9, [x9], #0x8\n" - "str d13, [x25], #0x8\n" + "str d13, [x26], #0x8\n" "tbz x11, #0, 68f\n" "st1 { v9.s }[2], [x9]\n" - "st1 { v13.s }[2], [x25]\n" + "st1 { v13.s }[2], [x26]\n" "b 68f\n" "65:" // Height 2: Partial direct writeback: partial_1_4 "tbz x11, #0, 68f\n" "str s9, [x9, #0x0]\n" - "str s13, [x25, #0x0]\n" + "str s13, [x26, #0x0]\n" "b 68f\n" "66:" // Height 2: Partial direct writeback: partial_2_0 "tbz x11, #1, 67f\n" "str d8, [x9], #0x8\n" - "str d12, [x25], #0x8\n" + "str d12, [x26], #0x8\n" "tbz x11, #0, 68f\n" "st1 { v8.s }[2], [x9]\n" - "st1 { v12.s }[2], [x25]\n" + "st1 { v12.s }[2], [x26]\n" "b 68f\n" "67:" // Height 2: Partial direct writeback: partial_1_0 "str s8, [x9, #0x0]\n" - "str s12, [x25, #0x0]\n" + "str s12, [x26, #0x0]\n" "68:" // Height 2: Partial direct writeback: Done "b 70f\n" "69:" // Height 2: Full writeback @@ -787,126 +789,126 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "str q10, [x9, #0x20]\n" "str q11, [x9, #0x30]\n" "add x9, x9, #0x40\n" - "str q12, [x25, #0x0]\n" - "str q13, [x25, #0x10]\n" - "str q14, [x25, #0x20]\n" - "str q15, [x25, #0x30]\n" + "str q12, [x26, #0x0]\n" + "str q13, [x26, #0x10]\n" + "str q14, [x26, #0x20]\n" + "str q15, [x26, #0x30]\n" "70:" // Height 2: Writeback done "subs x11, x11, #0x10\n" "bgt 37b\n" "b 212f\n" "71:" // Height 3 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "72:" // Height 3: Column loop "cbz x12, 73f\n" "ldr q8, [x12, #0x0]\n" "ldr q9, [x12, #0x10]\n" - "mov v12.16b, v8.16b\n" - "mov v13.16b, v9.16b\n" "ldr q10, [x12, #0x20]\n" "ldr q11, [x12, #0x30]\n" - "mov v14.16b, v10.16b\n" - "mov v15.16b, v11.16b\n" + "add x12, x12, #0x40\n" + "mov v12.16b, v8.16b\n" + "mov v13.16b, v9.16b\n" "mov v16.16b, v8.16b\n" "mov v17.16b, v9.16b\n" - "add x12, x12, #0x40\n" + "mov v14.16b, v10.16b\n" + "mov v15.16b, v11.16b\n" "mov v18.16b, v10.16b\n" "mov v19.16b, v11.16b\n" "b 84f\n" "73:" // Height 3: no bias "tbz %x[flags], #0, 83f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" "cmp x11, #0x10\n" - "add x24, x25, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" "bge 82f\n" "tbz x11, #3, 77f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" "ld1 { v9.4s }, [x9], #0x10\n" - "ld1 { v13.4s }, [x25], #0x10\n" - "ld1 { v17.4s }, [x24], #0x10\n" + "ld1 { v13.4s }, [x26], #0x10\n" + "ld1 { v17.4s }, [x25], #0x10\n" "tbz x11, #2, 75f\n" "ld1 { v10.4s }, [x9], #0x10\n" - "ld1 { v14.4s }, [x25], #0x10\n" - "ld1 { v18.4s }, [x24], #0x10\n" + "ld1 { v14.4s }, [x26], #0x10\n" + "ld1 { v18.4s }, [x25], #0x10\n" "tbz x11, #1, 74f\n" "ldr d11, [x9], #0x8\n" - "ldr d15, [x25], #0x8\n" + "ldr d15, [x26], #0x8\n" "mov x20, #0x38\n" - "ldr d19, [x24], #0x8\n" + "ldr d19, [x25], #0x8\n" "tbz x11, #0, 81f\n" "ld1 { v11.s }[2], [x9]\n" - "ld1 { v15.s }[2], [x25]\n" - "ld1 { v19.s }[2], [x24]\n" + "ld1 { v15.s }[2], [x26]\n" + "ld1 { v19.s }[2], [x25]\n" "b 81f\n" "74:" // Height 3: Partial accumulate: partial_1_12 "mov x20, #0x30\n" "tbz x11, #0, 81f\n" "ldr s11, [x9, #0x0]\n" - "ldr s15, [x25, #0x0]\n" - "ldr s19, [x24, #0x0]\n" + "ldr s15, [x26, #0x0]\n" + "ldr s19, [x25, #0x0]\n" "b 81f\n" "75:" // Height 3: Partial accumulate: partial_2_8 "tbz x11, #1, 76f\n" "ldr d10, [x9], #0x8\n" - "ldr d14, [x25], #0x8\n" + "ldr d14, [x26], #0x8\n" "mov x20, #0x28\n" - "ldr d18, [x24], #0x8\n" + "ldr d18, [x25], #0x8\n" "tbz x11, #0, 81f\n" "ld1 { v10.s }[2], [x9]\n" - "ld1 { v14.s }[2], [x25]\n" - "ld1 { v18.s }[2], [x24]\n" + "ld1 { v14.s }[2], [x26]\n" + "ld1 { v18.s }[2], [x25]\n" "b 81f\n" "76:" // Height 3: Partial accumulate: partial_1_8 "mov x20, #0x20\n" "tbz x11, #0, 81f\n" "ldr s10, [x9, #0x0]\n" - "ldr s14, [x25, #0x0]\n" - "ldr s18, [x24, #0x0]\n" + "ldr s14, [x26, #0x0]\n" + "ldr s18, [x25, #0x0]\n" "b 81f\n" "77:" // Height 3: Partial accumulate: partial_4_0 "tbz x11, #2, 79f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" "tbz x11, #1, 78f\n" "ldr d9, [x9], #0x8\n" - "ldr d13, [x25], #0x8\n" + "ldr d13, [x26], #0x8\n" "mov x20, #0x18\n" - "ldr d17, [x24], #0x8\n" + "ldr d17, [x25], #0x8\n" "tbz x11, #0, 81f\n" "ld1 { v9.s }[2], [x9]\n" - "ld1 { v13.s }[2], [x25]\n" - "ld1 { v17.s }[2], [x24]\n" + "ld1 { v13.s }[2], [x26]\n" + "ld1 { v17.s }[2], [x25]\n" "b 81f\n" "78:" // Height 3: Partial accumulate: partial_1_4 "mov x20, #0x10\n" "tbz x11, #0, 81f\n" "ldr s9, [x9, #0x0]\n" - "ldr s13, [x25, #0x0]\n" - "ldr s17, [x24, #0x0]\n" + "ldr s13, [x26, #0x0]\n" + "ldr s17, [x25, #0x0]\n" "b 81f\n" "79:" // Height 3: Partial accumulate: partial_2_0 "tbz x11, #1, 80f\n" "ldr d8, [x9], #0x8\n" - "ldr d12, [x25], #0x8\n" + "ldr d12, [x26], #0x8\n" "mov x20, #0x8\n" - "ldr d16, [x24], #0x8\n" + "ldr d16, [x25], #0x8\n" "tbz x11, #0, 81f\n" "ld1 { v8.s }[2], [x9]\n" - "ld1 { v12.s }[2], [x25]\n" - "ld1 { v16.s }[2], [x24]\n" + "ld1 { v12.s }[2], [x26]\n" + "ld1 { v16.s }[2], [x25]\n" "b 81f\n" "80:" // Height 3: Partial accumulate: partial_1_0 "ldr s8, [x9, #0x0]\n" - "ldr s12, [x25, #0x0]\n" + "ldr s12, [x26, #0x0]\n" "mov x20, #0x0\n" - "ldr s16, [x24, #0x0]\n" + "ldr s16, [x25, #0x0]\n" "81:" // Height 3: Partial accumulate: Done "sub x9, x9, x20\n" "b 84f\n" @@ -915,14 +917,14 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr q9, [x9, #0x10]\n" "ldr q10, [x9, #0x20]\n" "ldr q11, [x9, #0x30]\n" - "ldr q12, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "ldr q14, [x25, #0x20]\n" - "ldr q15, [x25, #0x30]\n" - "ldr q16, [x24, #0x0]\n" - "ldr q17, [x24, #0x10]\n" - "ldr q18, [x24, #0x20]\n" - "ldr q19, [x24, #0x30]\n" + "ldr q12, [x26, #0x0]\n" + "ldr q13, [x26, #0x10]\n" + "ldr q14, [x26, #0x20]\n" + "ldr q15, [x26, #0x30]\n" + "ldr q16, [x25, #0x0]\n" + "ldr q17, [x25, #0x10]\n" + "ldr q18, [x25, #0x20]\n" + "ldr q19, [x25, #0x30]\n" "b 84f\n" "83:" // Height 3: no accumulate "movi v8.16b, #0x0\n" @@ -941,8 +943,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "mov x28, #0x0\n" "85:" // 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, 86f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -982,18 +984,18 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" "ldr q20, [x10, #0x30]\n" "add x24, x24, #0x10\n" - ".inst 0x4f40f2aa // bfdot v10.4s, v21.8h, v0.h[0]\n" - ".inst 0x4f41f2ae // bfdot v14.4s, v21.8h, v1.h[0]\n" "cmp x27, #0x10\n" "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" + ".inst 0x4f40f2aa // bfdot v10.4s, v21.8h, v0.h[0]\n" + ".inst 0x4f41f2ae // bfdot v14.4s, v21.8h, v1.h[0]\n" ".inst 0x4f42f2b2 // bfdot v18.4s, v21.8h, v2.h[0]\n" "ldr q21, [x10, #0x40]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f40f28b // bfdot v11.4s, v20.8h, v0.h[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f41f28f // bfdot v15.4s, v20.8h, v1.h[0]\n" ".inst 0x4f42f293 // bfdot v19.4s, v20.8h, v2.h[0]\n" "ldr q20, [x10, #0x50]\n" - "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f60f2a8 // bfdot v8.4s, v21.8h, v0.h[1]\n" ".inst 0x4f61f2ac // bfdot v12.4s, v21.8h, v1.h[1]\n" ".inst 0x4f62f2b0 // bfdot v16.4s, v21.8h, v2.h[1]\n" @@ -1060,14 +1062,14 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" "ldr q20, [x10, #0x30]\n" "sub x27, x27, #0x8\n" - ".inst 0x4f40f2aa // bfdot v10.4s, v21.8h, v0.h[0]\n" - ".inst 0x4f41f2ae // bfdot v14.4s, v21.8h, v1.h[0]\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" + ".inst 0x4f40f2aa // bfdot v10.4s, v21.8h, v0.h[0]\n" + ".inst 0x4f41f2ae // bfdot v14.4s, v21.8h, v1.h[0]\n" ".inst 0x4f42f2b2 // bfdot v18.4s, v21.8h, v2.h[0]\n" "ldr q21, [x10, #0x40]\n" - ".inst 0x4f40f28b // bfdot v11.4s, v20.8h, v0.h[0]\n" "prfm pldl1keep, [x24, #0x80]\n" + ".inst 0x4f40f28b // bfdot v11.4s, v20.8h, v0.h[0]\n" ".inst 0x4f41f28f // bfdot v15.4s, v20.8h, v1.h[0]\n" ".inst 0x4f42f293 // bfdot v19.4s, v20.8h, v2.h[0]\n" "ldr q20, [x10, #0x50]\n" @@ -1126,12 +1128,12 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr s24, [x26], #0x4\n" "ldr s23, [x25], #0x4\n" "sub x27, x27, #0x2\n" - "cmp x27, #0x2\n" "ldr s22, [x24], #0x4\n" "ldr q21, [x10, #0x0]\n" + "cmp x27, #0x2\n" + "ldr q20, [x10, #0x10]\n" ".inst 0x4f58f2a8 // bfdot v8.4s, v21.8h, v24.h[0]\n" ".inst 0x4f57f2ac // bfdot v12.4s, v21.8h, v23.h[0]\n" - "ldr q20, [x10, #0x10]\n" ".inst 0x4f56f2b0 // bfdot v16.4s, v21.8h, v22.h[0]\n" "ldr q21, [x10, #0x20]\n" ".inst 0x4f58f289 // bfdot v9.4s, v20.8h, v24.h[0]\n" @@ -1175,15 +1177,15 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "cmp x28, x20\n" "bne 85b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x9, #0x0]\n" + "add x26, x9, x20, LSL #2\n" + "prfm pstl1keep, [x26, #0x0]\n" + "add x25, x26, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" - "prfm pstl1keep, [x24, #0x0]\n" "tbz %x[flags], #1, 95f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v21.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v21.4s }, [x21]\n" "ld1r { v20.4s }, [x20]\n" "fmin v8.4s, v8.4s, v21.4s\n" "fmin v9.4s, v9.4s, v21.4s\n" @@ -1215,79 +1217,79 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "tbz x11, #3, 99f\n" "st1 { v8.4s }, [x9], #0x10\n" "st1 { v9.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v13.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v17.4s }, [x24], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v13.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v17.4s }, [x25], #0x10\n" "tbz x11, #2, 97f\n" "st1 { v10.4s }, [x9], #0x10\n" - "st1 { v14.4s }, [x25], #0x10\n" - "st1 { v18.4s }, [x24], #0x10\n" + "st1 { v14.4s }, [x26], #0x10\n" + "st1 { v18.4s }, [x25], #0x10\n" "tbz x11, #1, 96f\n" "str d11, [x9], #0x8\n" - "str d15, [x25], #0x8\n" - "str d19, [x24], #0x8\n" + "str d15, [x26], #0x8\n" + "str d19, [x25], #0x8\n" "tbz x11, #0, 103f\n" "st1 { v11.s }[2], [x9]\n" - "st1 { v15.s }[2], [x25]\n" - "st1 { v19.s }[2], [x24]\n" + "st1 { v15.s }[2], [x26]\n" + "st1 { v19.s }[2], [x25]\n" "b 103f\n" "96:" // Height 3: Partial direct writeback: partial_1_12 "tbz x11, #0, 103f\n" "str s11, [x9, #0x0]\n" - "str s15, [x25, #0x0]\n" - "str s19, [x24, #0x0]\n" + "str s15, [x26, #0x0]\n" + "str s19, [x25, #0x0]\n" "b 103f\n" "97:" // Height 3: Partial direct writeback: partial_2_8 "tbz x11, #1, 98f\n" "str d10, [x9], #0x8\n" - "str d14, [x25], #0x8\n" - "str d18, [x24], #0x8\n" + "str d14, [x26], #0x8\n" + "str d18, [x25], #0x8\n" "tbz x11, #0, 103f\n" "st1 { v10.s }[2], [x9]\n" - "st1 { v14.s }[2], [x25]\n" - "st1 { v18.s }[2], [x24]\n" + "st1 { v14.s }[2], [x26]\n" + "st1 { v18.s }[2], [x25]\n" "b 103f\n" "98:" // Height 3: Partial direct writeback: partial_1_8 "tbz x11, #0, 103f\n" "str s10, [x9, #0x0]\n" - "str s14, [x25, #0x0]\n" - "str s18, [x24, #0x0]\n" + "str s14, [x26, #0x0]\n" + "str s18, [x25, #0x0]\n" "b 103f\n" "99:" // Height 3: Partial direct writeback: partial_4_0 "tbz x11, #2, 101f\n" "st1 { v8.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" "tbz x11, #1, 100f\n" "str d9, [x9], #0x8\n" - "str d13, [x25], #0x8\n" - "str d17, [x24], #0x8\n" + "str d13, [x26], #0x8\n" + "str d17, [x25], #0x8\n" "tbz x11, #0, 103f\n" "st1 { v9.s }[2], [x9]\n" - "st1 { v13.s }[2], [x25]\n" - "st1 { v17.s }[2], [x24]\n" + "st1 { v13.s }[2], [x26]\n" + "st1 { v17.s }[2], [x25]\n" "b 103f\n" "100:" // Height 3: Partial direct writeback: partial_1_4 "tbz x11, #0, 103f\n" "str s9, [x9, #0x0]\n" - "str s13, [x25, #0x0]\n" - "str s17, [x24, #0x0]\n" + "str s13, [x26, #0x0]\n" + "str s17, [x25, #0x0]\n" "b 103f\n" "101:" // Height 3: Partial direct writeback: partial_2_0 "tbz x11, #1, 102f\n" "str d8, [x9], #0x8\n" - "str d12, [x25], #0x8\n" - "str d16, [x24], #0x8\n" + "str d12, [x26], #0x8\n" + "str d16, [x25], #0x8\n" "tbz x11, #0, 103f\n" "st1 { v8.s }[2], [x9]\n" - "st1 { v12.s }[2], [x25]\n" - "st1 { v16.s }[2], [x24]\n" + "st1 { v12.s }[2], [x26]\n" + "st1 { v16.s }[2], [x25]\n" "b 103f\n" "102:" // Height 3: Partial direct writeback: partial_1_0 "str s8, [x9, #0x0]\n" - "str s12, [x25, #0x0]\n" - "str s16, [x24, #0x0]\n" + "str s12, [x26, #0x0]\n" + "str s16, [x25, #0x0]\n" "103:" // Height 3: Partial direct writeback: Done "b 105f\n" "104:" // Height 3: Full writeback @@ -1296,39 +1298,39 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "str q10, [x9, #0x20]\n" "str q11, [x9, #0x30]\n" "add x9, x9, #0x40\n" - "str q12, [x25, #0x0]\n" - "str q13, [x25, #0x10]\n" - "str q14, [x25, #0x20]\n" - "str q15, [x25, #0x30]\n" - "str q16, [x24, #0x0]\n" - "str q17, [x24, #0x10]\n" - "str q18, [x24, #0x20]\n" - "str q19, [x24, #0x30]\n" + "str q12, [x26, #0x0]\n" + "str q13, [x26, #0x10]\n" + "str q14, [x26, #0x20]\n" + "str q15, [x26, #0x30]\n" + "str q16, [x25, #0x0]\n" + "str q17, [x25, #0x10]\n" + "str q18, [x25, #0x20]\n" + "str q19, [x25, #0x30]\n" "105:" // Height 3: Writeback done "subs x11, x11, #0x10\n" "bgt 72b\n" "b 212f\n" "106:" // Height 4 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "107:" // Height 4: Column loop "cbz x12, 108f\n" "ldr q8, [x12, #0x0]\n" "ldr q9, [x12, #0x10]\n" - "mov v12.16b, v8.16b\n" - "mov v13.16b, v9.16b\n" "ldr q10, [x12, #0x20]\n" "ldr q11, [x12, #0x30]\n" - "mov v14.16b, v10.16b\n" - "mov v15.16b, v11.16b\n" + "add x12, x12, #0x40\n" + "mov v12.16b, v8.16b\n" + "mov v13.16b, v9.16b\n" "mov v16.16b, v8.16b\n" "mov v17.16b, v9.16b\n" - "add x12, x12, #0x40\n" + "mov v20.16b, v8.16b\n" + "mov v14.16b, v10.16b\n" + "mov v15.16b, v11.16b\n" "mov v18.16b, v10.16b\n" "mov v19.16b, v11.16b\n" - "mov v20.16b, v8.16b\n" "mov v21.16b, v9.16b\n" "mov v22.16b, v10.16b\n" "mov v23.16b, v11.16b\n" @@ -1336,111 +1338,111 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "108:" // Height 4: no bias "tbz %x[flags], #0, 118f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "cmp x11, #0x10\n" - "add x23, x24, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" + "add x24, x25, x20, LSL #2\n" "bge 117f\n" "tbz x11, #3, 112f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" "ld1 { v9.4s }, [x9], #0x10\n" - "ld1 { v13.4s }, [x25], #0x10\n" - "ld1 { v17.4s }, [x24], #0x10\n" - "ld1 { v21.4s }, [x23], #0x10\n" + "ld1 { v13.4s }, [x26], #0x10\n" + "ld1 { v17.4s }, [x25], #0x10\n" + "ld1 { v21.4s }, [x24], #0x10\n" "tbz x11, #2, 110f\n" "ld1 { v10.4s }, [x9], #0x10\n" - "ld1 { v14.4s }, [x25], #0x10\n" - "ld1 { v18.4s }, [x24], #0x10\n" - "ld1 { v22.4s }, [x23], #0x10\n" + "ld1 { v14.4s }, [x26], #0x10\n" + "ld1 { v18.4s }, [x25], #0x10\n" + "ld1 { v22.4s }, [x24], #0x10\n" "tbz x11, #1, 109f\n" "ldr d11, [x9], #0x8\n" - "ldr d15, [x25], #0x8\n" + "ldr d15, [x26], #0x8\n" "mov x20, #0x38\n" - "ldr d19, [x24], #0x8\n" - "ldr d23, [x23], #0x8\n" + "ldr d19, [x25], #0x8\n" + "ldr d23, [x24], #0x8\n" "tbz x11, #0, 116f\n" "ld1 { v11.s }[2], [x9]\n" - "ld1 { v15.s }[2], [x25]\n" - "ld1 { v19.s }[2], [x24]\n" - "ld1 { v23.s }[2], [x23]\n" + "ld1 { v15.s }[2], [x26]\n" + "ld1 { v19.s }[2], [x25]\n" + "ld1 { v23.s }[2], [x24]\n" "b 116f\n" "109:" // Height 4: Partial accumulate: partial_1_12 "mov x20, #0x30\n" "tbz x11, #0, 116f\n" "ldr s11, [x9, #0x0]\n" - "ldr s15, [x25, #0x0]\n" - "ldr s19, [x24, #0x0]\n" - "ldr s23, [x23, #0x0]\n" + "ldr s15, [x26, #0x0]\n" + "ldr s19, [x25, #0x0]\n" + "ldr s23, [x24, #0x0]\n" "b 116f\n" "110:" // Height 4: Partial accumulate: partial_2_8 "tbz x11, #1, 111f\n" "ldr d10, [x9], #0x8\n" - "ldr d14, [x25], #0x8\n" + "ldr d14, [x26], #0x8\n" "mov x20, #0x28\n" - "ldr d18, [x24], #0x8\n" - "ldr d22, [x23], #0x8\n" + "ldr d18, [x25], #0x8\n" + "ldr d22, [x24], #0x8\n" "tbz x11, #0, 116f\n" "ld1 { v10.s }[2], [x9]\n" - "ld1 { v14.s }[2], [x25]\n" - "ld1 { v18.s }[2], [x24]\n" - "ld1 { v22.s }[2], [x23]\n" + "ld1 { v14.s }[2], [x26]\n" + "ld1 { v18.s }[2], [x25]\n" + "ld1 { v22.s }[2], [x24]\n" "b 116f\n" "111:" // Height 4: Partial accumulate: partial_1_8 "mov x20, #0x20\n" "tbz x11, #0, 116f\n" "ldr s10, [x9, #0x0]\n" - "ldr s14, [x25, #0x0]\n" - "ldr s18, [x24, #0x0]\n" - "ldr s22, [x23, #0x0]\n" + "ldr s14, [x26, #0x0]\n" + "ldr s18, [x25, #0x0]\n" + "ldr s22, [x24, #0x0]\n" "b 116f\n" "112:" // Height 4: Partial accumulate: partial_4_0 "tbz x11, #2, 114f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" "tbz x11, #1, 113f\n" "ldr d9, [x9], #0x8\n" - "ldr d13, [x25], #0x8\n" + "ldr d13, [x26], #0x8\n" "mov x20, #0x18\n" - "ldr d17, [x24], #0x8\n" - "ldr d21, [x23], #0x8\n" + "ldr d17, [x25], #0x8\n" + "ldr d21, [x24], #0x8\n" "tbz x11, #0, 116f\n" "ld1 { v9.s }[2], [x9]\n" - "ld1 { v13.s }[2], [x25]\n" - "ld1 { v17.s }[2], [x24]\n" - "ld1 { v21.s }[2], [x23]\n" + "ld1 { v13.s }[2], [x26]\n" + "ld1 { v17.s }[2], [x25]\n" + "ld1 { v21.s }[2], [x24]\n" "b 116f\n" "113:" // Height 4: Partial accumulate: partial_1_4 "mov x20, #0x10\n" "tbz x11, #0, 116f\n" "ldr s9, [x9, #0x0]\n" - "ldr s13, [x25, #0x0]\n" - "ldr s17, [x24, #0x0]\n" - "ldr s21, [x23, #0x0]\n" + "ldr s13, [x26, #0x0]\n" + "ldr s17, [x25, #0x0]\n" + "ldr s21, [x24, #0x0]\n" "b 116f\n" "114:" // Height 4: Partial accumulate: partial_2_0 "tbz x11, #1, 115f\n" "ldr d8, [x9], #0x8\n" - "ldr d12, [x25], #0x8\n" + "ldr d12, [x26], #0x8\n" "mov x20, #0x8\n" - "ldr d16, [x24], #0x8\n" - "ldr d20, [x23], #0x8\n" + "ldr d16, [x25], #0x8\n" + "ldr d20, [x24], #0x8\n" "tbz x11, #0, 116f\n" "ld1 { v8.s }[2], [x9]\n" - "ld1 { v12.s }[2], [x25]\n" - "ld1 { v16.s }[2], [x24]\n" - "ld1 { v20.s }[2], [x23]\n" + "ld1 { v12.s }[2], [x26]\n" + "ld1 { v16.s }[2], [x25]\n" + "ld1 { v20.s }[2], [x24]\n" "b 116f\n" "115:" // Height 4: Partial accumulate: partial_1_0 "ldr s8, [x9, #0x0]\n" - "ldr s12, [x25, #0x0]\n" + "ldr s12, [x26, #0x0]\n" "mov x20, #0x0\n" - "ldr s16, [x24, #0x0]\n" - "ldr s20, [x23, #0x0]\n" + "ldr s16, [x25, #0x0]\n" + "ldr s20, [x24, #0x0]\n" "116:" // Height 4: Partial accumulate: Done "sub x9, x9, x20\n" "b 119f\n" @@ -1449,18 +1451,18 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr q9, [x9, #0x10]\n" "ldr q10, [x9, #0x20]\n" "ldr q11, [x9, #0x30]\n" - "ldr q12, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "ldr q14, [x25, #0x20]\n" - "ldr q15, [x25, #0x30]\n" - "ldr q16, [x24, #0x0]\n" - "ldr q17, [x24, #0x10]\n" - "ldr q18, [x24, #0x20]\n" - "ldr q19, [x24, #0x30]\n" - "ldr q20, [x23, #0x0]\n" - "ldr q21, [x23, #0x10]\n" - "ldr q22, [x23, #0x20]\n" - "ldr q23, [x23, #0x30]\n" + "ldr q12, [x26, #0x0]\n" + "ldr q13, [x26, #0x10]\n" + "ldr q14, [x26, #0x20]\n" + "ldr q15, [x26, #0x30]\n" + "ldr q16, [x25, #0x0]\n" + "ldr q17, [x25, #0x10]\n" + "ldr q18, [x25, #0x20]\n" + "ldr q19, [x25, #0x30]\n" + "ldr q20, [x24, #0x0]\n" + "ldr q21, [x24, #0x10]\n" + "ldr q22, [x24, #0x20]\n" + "ldr q23, [x24, #0x30]\n" "b 119f\n" "118:" // Height 4: no accumulate "movi v8.16b, #0x0\n" @@ -1483,8 +1485,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "mov x28, #0x0\n" "120:" // 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, 121f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -1624,14 +1626,14 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" "add x23, x23, #0x10\n" - "sub x27, x27, #0x8\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" "ldr q24, [x10, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f40f32a // bfdot v10.4s, v25.8h, v0.h[0]\n" ".inst 0x4f41f32e // bfdot v14.4s, v25.8h, v1.h[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "sub x27, x27, #0x8\n" "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f42f332 // bfdot v18.4s, v25.8h, v2.h[0]\n" ".inst 0x4f43f336 // bfdot v22.4s, v25.8h, v3.h[0]\n" @@ -1709,9 +1711,9 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr s29, [x26], #0x4\n" "ldr s28, [x25], #0x4\n" "sub x27, x27, #0x2\n" - "cmp x27, #0x2\n" "ldr s27, [x24], #0x4\n" "ldr s26, [x23], #0x4\n" + "cmp x27, #0x2\n" "ldr q25, [x10, #0x0]\n" "ldr q24, [x10, #0x10]\n" ".inst 0x4f5df328 // bfdot v8.4s, v25.8h, v29.h[0]\n" @@ -1768,17 +1770,17 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "cmp x28, x20\n" "bne 120b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x9, #0x0]\n" - "add x23, x24, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" + "prfm pstl1keep, [x26, #0x0]\n" + "add x25, x26, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" + "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x24, #0x0]\n" - "prfm pstl1keep, [x23, #0x0]\n" "tbz %x[flags], #1, 130f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v25.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v25.4s }, [x21]\n" "ld1r { v24.4s }, [x20]\n" "fmin v8.4s, v8.4s, v25.4s\n" "fmin v9.4s, v9.4s, v25.4s\n" @@ -1818,95 +1820,95 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "tbz x11, #3, 134f\n" "st1 { v8.4s }, [x9], #0x10\n" "st1 { v9.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v13.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v17.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" - "st1 { v21.4s }, [x23], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v13.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v17.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" + "st1 { v21.4s }, [x24], #0x10\n" "tbz x11, #2, 132f\n" "st1 { v10.4s }, [x9], #0x10\n" - "st1 { v14.4s }, [x25], #0x10\n" - "st1 { v18.4s }, [x24], #0x10\n" - "st1 { v22.4s }, [x23], #0x10\n" + "st1 { v14.4s }, [x26], #0x10\n" + "st1 { v18.4s }, [x25], #0x10\n" + "st1 { v22.4s }, [x24], #0x10\n" "tbz x11, #1, 131f\n" "str d11, [x9], #0x8\n" - "str d15, [x25], #0x8\n" - "str d19, [x24], #0x8\n" - "str d23, [x23], #0x8\n" + "str d15, [x26], #0x8\n" + "str d19, [x25], #0x8\n" + "str d23, [x24], #0x8\n" "tbz x11, #0, 138f\n" "st1 { v11.s }[2], [x9]\n" - "st1 { v15.s }[2], [x25]\n" - "st1 { v19.s }[2], [x24]\n" - "st1 { v23.s }[2], [x23]\n" + "st1 { v15.s }[2], [x26]\n" + "st1 { v19.s }[2], [x25]\n" + "st1 { v23.s }[2], [x24]\n" "b 138f\n" "131:" // Height 4: Partial direct writeback: partial_1_12 "tbz x11, #0, 138f\n" "str s11, [x9, #0x0]\n" - "str s15, [x25, #0x0]\n" - "str s19, [x24, #0x0]\n" - "str s23, [x23, #0x0]\n" + "str s15, [x26, #0x0]\n" + "str s19, [x25, #0x0]\n" + "str s23, [x24, #0x0]\n" "b 138f\n" "132:" // Height 4: Partial direct writeback: partial_2_8 "tbz x11, #1, 133f\n" "str d10, [x9], #0x8\n" - "str d14, [x25], #0x8\n" - "str d18, [x24], #0x8\n" - "str d22, [x23], #0x8\n" + "str d14, [x26], #0x8\n" + "str d18, [x25], #0x8\n" + "str d22, [x24], #0x8\n" "tbz x11, #0, 138f\n" "st1 { v10.s }[2], [x9]\n" - "st1 { v14.s }[2], [x25]\n" - "st1 { v18.s }[2], [x24]\n" - "st1 { v22.s }[2], [x23]\n" + "st1 { v14.s }[2], [x26]\n" + "st1 { v18.s }[2], [x25]\n" + "st1 { v22.s }[2], [x24]\n" "b 138f\n" "133:" // Height 4: Partial direct writeback: partial_1_8 "tbz x11, #0, 138f\n" "str s10, [x9, #0x0]\n" - "str s14, [x25, #0x0]\n" - "str s18, [x24, #0x0]\n" - "str s22, [x23, #0x0]\n" + "str s14, [x26, #0x0]\n" + "str s18, [x25, #0x0]\n" + "str s22, [x24, #0x0]\n" "b 138f\n" "134:" // Height 4: Partial direct writeback: partial_4_0 "tbz x11, #2, 136f\n" "st1 { v8.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" "tbz x11, #1, 135f\n" "str d9, [x9], #0x8\n" - "str d13, [x25], #0x8\n" - "str d17, [x24], #0x8\n" - "str d21, [x23], #0x8\n" + "str d13, [x26], #0x8\n" + "str d17, [x25], #0x8\n" + "str d21, [x24], #0x8\n" "tbz x11, #0, 138f\n" "st1 { v9.s }[2], [x9]\n" - "st1 { v13.s }[2], [x25]\n" - "st1 { v17.s }[2], [x24]\n" - "st1 { v21.s }[2], [x23]\n" + "st1 { v13.s }[2], [x26]\n" + "st1 { v17.s }[2], [x25]\n" + "st1 { v21.s }[2], [x24]\n" "b 138f\n" "135:" // Height 4: Partial direct writeback: partial_1_4 "tbz x11, #0, 138f\n" "str s9, [x9, #0x0]\n" - "str s13, [x25, #0x0]\n" - "str s17, [x24, #0x0]\n" - "str s21, [x23, #0x0]\n" + "str s13, [x26, #0x0]\n" + "str s17, [x25, #0x0]\n" + "str s21, [x24, #0x0]\n" "b 138f\n" "136:" // Height 4: Partial direct writeback: partial_2_0 "tbz x11, #1, 137f\n" "str d8, [x9], #0x8\n" - "str d12, [x25], #0x8\n" - "str d16, [x24], #0x8\n" - "str d20, [x23], #0x8\n" + "str d12, [x26], #0x8\n" + "str d16, [x25], #0x8\n" + "str d20, [x24], #0x8\n" "tbz x11, #0, 138f\n" "st1 { v8.s }[2], [x9]\n" - "st1 { v12.s }[2], [x25]\n" - "st1 { v16.s }[2], [x24]\n" - "st1 { v20.s }[2], [x23]\n" + "st1 { v12.s }[2], [x26]\n" + "st1 { v16.s }[2], [x25]\n" + "st1 { v20.s }[2], [x24]\n" "b 138f\n" "137:" // Height 4: Partial direct writeback: partial_1_0 "str s8, [x9, #0x0]\n" - "str s12, [x25, #0x0]\n" - "str s16, [x24, #0x0]\n" - "str s20, [x23, #0x0]\n" + "str s12, [x26, #0x0]\n" + "str s16, [x25, #0x0]\n" + "str s20, [x24, #0x0]\n" "138:" // Height 4: Partial direct writeback: Done "b 140f\n" "139:" // Height 4: Full writeback @@ -1915,43 +1917,43 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "str q10, [x9, #0x20]\n" "str q11, [x9, #0x30]\n" "add x9, x9, #0x40\n" - "str q12, [x25, #0x0]\n" - "str q13, [x25, #0x10]\n" - "str q14, [x25, #0x20]\n" - "str q15, [x25, #0x30]\n" - "str q16, [x24, #0x0]\n" - "str q17, [x24, #0x10]\n" - "str q18, [x24, #0x20]\n" - "str q19, [x24, #0x30]\n" - "str q20, [x23, #0x0]\n" - "str q21, [x23, #0x10]\n" - "str q22, [x23, #0x20]\n" - "str q23, [x23, #0x30]\n" + "str q12, [x26, #0x0]\n" + "str q13, [x26, #0x10]\n" + "str q14, [x26, #0x20]\n" + "str q15, [x26, #0x30]\n" + "str q16, [x25, #0x0]\n" + "str q17, [x25, #0x10]\n" + "str q18, [x25, #0x20]\n" + "str q19, [x25, #0x30]\n" + "str q20, [x24, #0x0]\n" + "str q21, [x24, #0x10]\n" + "str q22, [x24, #0x20]\n" + "str q23, [x24, #0x30]\n" "140:" // Height 4: Writeback done "subs x11, x11, #0x10\n" "bgt 107b\n" "b 212f\n" "141:" // Height 5 - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "142:" // Height 5: Column loop "cbz x12, 143f\n" "ldr q8, [x12, #0x0]\n" "ldr q9, [x12, #0x10]\n" - "mov v12.16b, v8.16b\n" - "mov v13.16b, v9.16b\n" "ldr q10, [x12, #0x20]\n" "ldr q11, [x12, #0x30]\n" - "mov v14.16b, v10.16b\n" - "mov v15.16b, v11.16b\n" + "add x12, x12, #0x40\n" + "mov v12.16b, v8.16b\n" + "mov v13.16b, v9.16b\n" "mov v16.16b, v8.16b\n" "mov v17.16b, v9.16b\n" - "add x12, x12, #0x40\n" + "mov v20.16b, v8.16b\n" + "mov v14.16b, v10.16b\n" + "mov v15.16b, v11.16b\n" "mov v18.16b, v10.16b\n" "mov v19.16b, v11.16b\n" - "mov v20.16b, v8.16b\n" "mov v21.16b, v9.16b\n" "mov v22.16b, v10.16b\n" "mov v23.16b, v11.16b\n" @@ -1963,128 +1965,128 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "143:" // Height 5: no bias "tbz %x[flags], #0, 153f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" + "cmp x11, #0x10\n" + "add x26, x9, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" "add x24, x25, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" - "cmp x11, #0x10\n" - "add x22, x23, x20, LSL #2\n" "bge 152f\n" "tbz x11, #3, 147f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" - "ld1 { v24.4s }, [x22], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" + "ld1 { v24.4s }, [x23], #0x10\n" "ld1 { v9.4s }, [x9], #0x10\n" - "ld1 { v13.4s }, [x25], #0x10\n" - "ld1 { v17.4s }, [x24], #0x10\n" - "ld1 { v21.4s }, [x23], #0x10\n" - "ld1 { v25.4s }, [x22], #0x10\n" + "ld1 { v13.4s }, [x26], #0x10\n" + "ld1 { v17.4s }, [x25], #0x10\n" + "ld1 { v21.4s }, [x24], #0x10\n" + "ld1 { v25.4s }, [x23], #0x10\n" "tbz x11, #2, 145f\n" "ld1 { v10.4s }, [x9], #0x10\n" - "ld1 { v14.4s }, [x25], #0x10\n" - "ld1 { v18.4s }, [x24], #0x10\n" - "ld1 { v22.4s }, [x23], #0x10\n" - "ld1 { v26.4s }, [x22], #0x10\n" + "ld1 { v14.4s }, [x26], #0x10\n" + "ld1 { v18.4s }, [x25], #0x10\n" + "ld1 { v22.4s }, [x24], #0x10\n" + "ld1 { v26.4s }, [x23], #0x10\n" "tbz x11, #1, 144f\n" "ldr d11, [x9], #0x8\n" - "ldr d15, [x25], #0x8\n" + "ldr d15, [x26], #0x8\n" "mov x20, #0x38\n" - "ldr d19, [x24], #0x8\n" - "ldr d23, [x23], #0x8\n" - "ldr d27, [x22], #0x8\n" + "ldr d19, [x25], #0x8\n" + "ldr d23, [x24], #0x8\n" + "ldr d27, [x23], #0x8\n" "tbz x11, #0, 151f\n" "ld1 { v11.s }[2], [x9]\n" - "ld1 { v15.s }[2], [x25]\n" - "ld1 { v19.s }[2], [x24]\n" - "ld1 { v23.s }[2], [x23]\n" - "ld1 { v27.s }[2], [x22]\n" + "ld1 { v15.s }[2], [x26]\n" + "ld1 { v19.s }[2], [x25]\n" + "ld1 { v23.s }[2], [x24]\n" + "ld1 { v27.s }[2], [x23]\n" "b 151f\n" "144:" // Height 5: Partial accumulate: partial_1_12 "mov x20, #0x30\n" "tbz x11, #0, 151f\n" "ldr s11, [x9, #0x0]\n" - "ldr s15, [x25, #0x0]\n" - "ldr s19, [x24, #0x0]\n" - "ldr s23, [x23, #0x0]\n" - "ldr s27, [x22, #0x0]\n" + "ldr s15, [x26, #0x0]\n" + "ldr s19, [x25, #0x0]\n" + "ldr s23, [x24, #0x0]\n" + "ldr s27, [x23, #0x0]\n" "b 151f\n" "145:" // Height 5: Partial accumulate: partial_2_8 "tbz x11, #1, 146f\n" "ldr d10, [x9], #0x8\n" - "ldr d14, [x25], #0x8\n" + "ldr d14, [x26], #0x8\n" "mov x20, #0x28\n" - "ldr d18, [x24], #0x8\n" - "ldr d22, [x23], #0x8\n" - "ldr d26, [x22], #0x8\n" + "ldr d18, [x25], #0x8\n" + "ldr d22, [x24], #0x8\n" + "ldr d26, [x23], #0x8\n" "tbz x11, #0, 151f\n" "ld1 { v10.s }[2], [x9]\n" - "ld1 { v14.s }[2], [x25]\n" - "ld1 { v18.s }[2], [x24]\n" - "ld1 { v22.s }[2], [x23]\n" - "ld1 { v26.s }[2], [x22]\n" + "ld1 { v14.s }[2], [x26]\n" + "ld1 { v18.s }[2], [x25]\n" + "ld1 { v22.s }[2], [x24]\n" + "ld1 { v26.s }[2], [x23]\n" "b 151f\n" "146:" // Height 5: Partial accumulate: partial_1_8 "mov x20, #0x20\n" "tbz x11, #0, 151f\n" "ldr s10, [x9, #0x0]\n" - "ldr s14, [x25, #0x0]\n" - "ldr s18, [x24, #0x0]\n" - "ldr s22, [x23, #0x0]\n" - "ldr s26, [x22, #0x0]\n" + "ldr s14, [x26, #0x0]\n" + "ldr s18, [x25, #0x0]\n" + "ldr s22, [x24, #0x0]\n" + "ldr s26, [x23, #0x0]\n" "b 151f\n" "147:" // Height 5: Partial accumulate: partial_4_0 "tbz x11, #2, 149f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" - "ld1 { v24.4s }, [x22], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" + "ld1 { v24.4s }, [x23], #0x10\n" "tbz x11, #1, 148f\n" "ldr d9, [x9], #0x8\n" - "ldr d13, [x25], #0x8\n" + "ldr d13, [x26], #0x8\n" "mov x20, #0x18\n" - "ldr d17, [x24], #0x8\n" - "ldr d21, [x23], #0x8\n" - "ldr d25, [x22], #0x8\n" + "ldr d17, [x25], #0x8\n" + "ldr d21, [x24], #0x8\n" + "ldr d25, [x23], #0x8\n" "tbz x11, #0, 151f\n" "ld1 { v9.s }[2], [x9]\n" - "ld1 { v13.s }[2], [x25]\n" - "ld1 { v17.s }[2], [x24]\n" - "ld1 { v21.s }[2], [x23]\n" - "ld1 { v25.s }[2], [x22]\n" + "ld1 { v13.s }[2], [x26]\n" + "ld1 { v17.s }[2], [x25]\n" + "ld1 { v21.s }[2], [x24]\n" + "ld1 { v25.s }[2], [x23]\n" "b 151f\n" "148:" // Height 5: Partial accumulate: partial_1_4 "mov x20, #0x10\n" "tbz x11, #0, 151f\n" "ldr s9, [x9, #0x0]\n" - "ldr s13, [x25, #0x0]\n" - "ldr s17, [x24, #0x0]\n" - "ldr s21, [x23, #0x0]\n" - "ldr s25, [x22, #0x0]\n" + "ldr s13, [x26, #0x0]\n" + "ldr s17, [x25, #0x0]\n" + "ldr s21, [x24, #0x0]\n" + "ldr s25, [x23, #0x0]\n" "b 151f\n" "149:" // Height 5: Partial accumulate: partial_2_0 "tbz x11, #1, 150f\n" "ldr d8, [x9], #0x8\n" - "ldr d12, [x25], #0x8\n" + "ldr d12, [x26], #0x8\n" "mov x20, #0x8\n" - "ldr d16, [x24], #0x8\n" - "ldr d20, [x23], #0x8\n" - "ldr d24, [x22], #0x8\n" + "ldr d16, [x25], #0x8\n" + "ldr d20, [x24], #0x8\n" + "ldr d24, [x23], #0x8\n" "tbz x11, #0, 151f\n" "ld1 { v8.s }[2], [x9]\n" - "ld1 { v12.s }[2], [x25]\n" - "ld1 { v16.s }[2], [x24]\n" - "ld1 { v20.s }[2], [x23]\n" - "ld1 { v24.s }[2], [x22]\n" + "ld1 { v12.s }[2], [x26]\n" + "ld1 { v16.s }[2], [x25]\n" + "ld1 { v20.s }[2], [x24]\n" + "ld1 { v24.s }[2], [x23]\n" "b 151f\n" "150:" // Height 5: Partial accumulate: partial_1_0 "ldr s8, [x9, #0x0]\n" - "ldr s12, [x25, #0x0]\n" + "ldr s12, [x26, #0x0]\n" "mov x20, #0x0\n" - "ldr s16, [x24, #0x0]\n" - "ldr s20, [x23, #0x0]\n" - "ldr s24, [x22, #0x0]\n" + "ldr s16, [x25, #0x0]\n" + "ldr s20, [x24, #0x0]\n" + "ldr s24, [x23, #0x0]\n" "151:" // Height 5: Partial accumulate: Done "sub x9, x9, x20\n" "b 154f\n" @@ -2093,22 +2095,22 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr q9, [x9, #0x10]\n" "ldr q10, [x9, #0x20]\n" "ldr q11, [x9, #0x30]\n" - "ldr q12, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "ldr q14, [x25, #0x20]\n" - "ldr q15, [x25, #0x30]\n" - "ldr q16, [x24, #0x0]\n" - "ldr q17, [x24, #0x10]\n" - "ldr q18, [x24, #0x20]\n" - "ldr q19, [x24, #0x30]\n" - "ldr q20, [x23, #0x0]\n" - "ldr q21, [x23, #0x10]\n" - "ldr q22, [x23, #0x20]\n" - "ldr q23, [x23, #0x30]\n" - "ldr q24, [x22, #0x0]\n" - "ldr q25, [x22, #0x10]\n" - "ldr q26, [x22, #0x20]\n" - "ldr q27, [x22, #0x30]\n" + "ldr q12, [x26, #0x0]\n" + "ldr q13, [x26, #0x10]\n" + "ldr q14, [x26, #0x20]\n" + "ldr q15, [x26, #0x30]\n" + "ldr q16, [x25, #0x0]\n" + "ldr q17, [x25, #0x10]\n" + "ldr q18, [x25, #0x20]\n" + "ldr q19, [x25, #0x30]\n" + "ldr q20, [x24, #0x0]\n" + "ldr q21, [x24, #0x10]\n" + "ldr q22, [x24, #0x20]\n" + "ldr q23, [x24, #0x30]\n" + "ldr q24, [x23, #0x0]\n" + "ldr q25, [x23, #0x10]\n" + "ldr q26, [x23, #0x20]\n" + "ldr q27, [x23, #0x30]\n" "b 154f\n" "153:" // Height 5: no accumulate "movi v8.16b, #0x0\n" @@ -2135,8 +2137,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "mov x28, #0x0\n" "155:" // 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, 156f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -2302,12 +2304,12 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "add x22, x22, #0x10\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "sub x27, x27, #0x8\n" "prfm pldl1keep, [x26, #0x80]\n" + "prfm pldl1keep, [x25, #0x80]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" "ldr q28, [x10, #0x30]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "sub x27, x27, #0x8\n" ".inst 0x4f40f3aa // bfdot v10.4s, v29.8h, v0.h[0]\n" ".inst 0x4f41f3ae // bfdot v14.4s, v29.8h, v1.h[0]\n" "prfm pldl1keep, [x24, #0x80]\n" @@ -2402,14 +2404,14 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr s2, [x26], #0x4\n" "ldr s1, [x25], #0x4\n" "sub x27, x27, #0x2\n" - "cmp x27, #0x2\n" "ldr s0, [x24], #0x4\n" "ldr s31, [x23], #0x4\n" + "cmp x27, #0x2\n" "ldr s30, [x22], #0x4\n" "ldr q29, [x10, #0x0]\n" + "ldr q28, [x10, #0x10]\n" ".inst 0x4f42f3a8 // bfdot v8.4s, v29.8h, v2.h[0]\n" ".inst 0x4f41f3ac // bfdot v12.4s, v29.8h, v1.h[0]\n" - "ldr q28, [x10, #0x10]\n" ".inst 0x4f40f3b0 // bfdot v16.4s, v29.8h, v0.h[0]\n" ".inst 0x4f5ff3b4 // bfdot v20.4s, v29.8h, v31.h[0]\n" ".inst 0x4f5ef3b8 // bfdot v24.4s, v29.8h, v30.h[0]\n" @@ -2471,19 +2473,19 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "cmp x28, x20\n" "bne 155b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x9, #0x0]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" + "prfm pstl1keep, [x26, #0x0]\n" + "add x25, x26, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" + "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x24, #0x0]\n" + "add x23, x24, x20, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" - "prfm pstl1keep, [x22, #0x0]\n" "tbz %x[flags], #1, 165f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v29.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v29.4s }, [x21]\n" "ld1r { v28.4s }, [x20]\n" "fmin v8.4s, v8.4s, v29.4s\n" "fmin v9.4s, v9.4s, v29.4s\n" @@ -2531,111 +2533,111 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "tbz x11, #3, 169f\n" "st1 { v8.4s }, [x9], #0x10\n" "st1 { v9.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v13.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v17.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" - "st1 { v21.4s }, [x23], #0x10\n" - "st1 { v24.4s }, [x22], #0x10\n" - "st1 { v25.4s }, [x22], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v13.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v17.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" + "st1 { v21.4s }, [x24], #0x10\n" + "st1 { v24.4s }, [x23], #0x10\n" + "st1 { v25.4s }, [x23], #0x10\n" "tbz x11, #2, 167f\n" "st1 { v10.4s }, [x9], #0x10\n" - "st1 { v14.4s }, [x25], #0x10\n" - "st1 { v18.4s }, [x24], #0x10\n" - "st1 { v22.4s }, [x23], #0x10\n" - "st1 { v26.4s }, [x22], #0x10\n" + "st1 { v14.4s }, [x26], #0x10\n" + "st1 { v18.4s }, [x25], #0x10\n" + "st1 { v22.4s }, [x24], #0x10\n" + "st1 { v26.4s }, [x23], #0x10\n" "tbz x11, #1, 166f\n" "str d11, [x9], #0x8\n" - "str d15, [x25], #0x8\n" - "str d19, [x24], #0x8\n" - "str d23, [x23], #0x8\n" - "str d27, [x22], #0x8\n" + "str d15, [x26], #0x8\n" + "str d19, [x25], #0x8\n" + "str d23, [x24], #0x8\n" + "str d27, [x23], #0x8\n" "tbz x11, #0, 173f\n" "st1 { v11.s }[2], [x9]\n" - "st1 { v15.s }[2], [x25]\n" - "st1 { v19.s }[2], [x24]\n" - "st1 { v23.s }[2], [x23]\n" - "st1 { v27.s }[2], [x22]\n" + "st1 { v15.s }[2], [x26]\n" + "st1 { v19.s }[2], [x25]\n" + "st1 { v23.s }[2], [x24]\n" + "st1 { v27.s }[2], [x23]\n" "b 173f\n" "166:" // Height 5: Partial direct writeback: partial_1_12 "tbz x11, #0, 173f\n" "str s11, [x9, #0x0]\n" - "str s15, [x25, #0x0]\n" - "str s19, [x24, #0x0]\n" - "str s23, [x23, #0x0]\n" - "str s27, [x22, #0x0]\n" + "str s15, [x26, #0x0]\n" + "str s19, [x25, #0x0]\n" + "str s23, [x24, #0x0]\n" + "str s27, [x23, #0x0]\n" "b 173f\n" "167:" // Height 5: Partial direct writeback: partial_2_8 "tbz x11, #1, 168f\n" "str d10, [x9], #0x8\n" - "str d14, [x25], #0x8\n" - "str d18, [x24], #0x8\n" - "str d22, [x23], #0x8\n" - "str d26, [x22], #0x8\n" + "str d14, [x26], #0x8\n" + "str d18, [x25], #0x8\n" + "str d22, [x24], #0x8\n" + "str d26, [x23], #0x8\n" "tbz x11, #0, 173f\n" "st1 { v10.s }[2], [x9]\n" - "st1 { v14.s }[2], [x25]\n" - "st1 { v18.s }[2], [x24]\n" - "st1 { v22.s }[2], [x23]\n" - "st1 { v26.s }[2], [x22]\n" + "st1 { v14.s }[2], [x26]\n" + "st1 { v18.s }[2], [x25]\n" + "st1 { v22.s }[2], [x24]\n" + "st1 { v26.s }[2], [x23]\n" "b 173f\n" "168:" // Height 5: Partial direct writeback: partial_1_8 "tbz x11, #0, 173f\n" "str s10, [x9, #0x0]\n" - "str s14, [x25, #0x0]\n" - "str s18, [x24, #0x0]\n" - "str s22, [x23, #0x0]\n" - "str s26, [x22, #0x0]\n" + "str s14, [x26, #0x0]\n" + "str s18, [x25, #0x0]\n" + "str s22, [x24, #0x0]\n" + "str s26, [x23, #0x0]\n" "b 173f\n" "169:" // Height 5: Partial direct writeback: partial_4_0 "tbz x11, #2, 171f\n" "st1 { v8.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" - "st1 { v24.4s }, [x22], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" + "st1 { v24.4s }, [x23], #0x10\n" "tbz x11, #1, 170f\n" "str d9, [x9], #0x8\n" - "str d13, [x25], #0x8\n" - "str d17, [x24], #0x8\n" - "str d21, [x23], #0x8\n" - "str d25, [x22], #0x8\n" + "str d13, [x26], #0x8\n" + "str d17, [x25], #0x8\n" + "str d21, [x24], #0x8\n" + "str d25, [x23], #0x8\n" "tbz x11, #0, 173f\n" "st1 { v9.s }[2], [x9]\n" - "st1 { v13.s }[2], [x25]\n" - "st1 { v17.s }[2], [x24]\n" - "st1 { v21.s }[2], [x23]\n" - "st1 { v25.s }[2], [x22]\n" + "st1 { v13.s }[2], [x26]\n" + "st1 { v17.s }[2], [x25]\n" + "st1 { v21.s }[2], [x24]\n" + "st1 { v25.s }[2], [x23]\n" "b 173f\n" "170:" // Height 5: Partial direct writeback: partial_1_4 "tbz x11, #0, 173f\n" "str s9, [x9, #0x0]\n" - "str s13, [x25, #0x0]\n" - "str s17, [x24, #0x0]\n" - "str s21, [x23, #0x0]\n" - "str s25, [x22, #0x0]\n" + "str s13, [x26, #0x0]\n" + "str s17, [x25, #0x0]\n" + "str s21, [x24, #0x0]\n" + "str s25, [x23, #0x0]\n" "b 173f\n" "171:" // Height 5: Partial direct writeback: partial_2_0 "tbz x11, #1, 172f\n" "str d8, [x9], #0x8\n" - "str d12, [x25], #0x8\n" - "str d16, [x24], #0x8\n" - "str d20, [x23], #0x8\n" - "str d24, [x22], #0x8\n" + "str d12, [x26], #0x8\n" + "str d16, [x25], #0x8\n" + "str d20, [x24], #0x8\n" + "str d24, [x23], #0x8\n" "tbz x11, #0, 173f\n" "st1 { v8.s }[2], [x9]\n" - "st1 { v12.s }[2], [x25]\n" - "st1 { v16.s }[2], [x24]\n" - "st1 { v20.s }[2], [x23]\n" - "st1 { v24.s }[2], [x22]\n" + "st1 { v12.s }[2], [x26]\n" + "st1 { v16.s }[2], [x25]\n" + "st1 { v20.s }[2], [x24]\n" + "st1 { v24.s }[2], [x23]\n" "b 173f\n" "172:" // Height 5: Partial direct writeback: partial_1_0 "str s8, [x9, #0x0]\n" - "str s12, [x25, #0x0]\n" - "str s16, [x24, #0x0]\n" - "str s20, [x23, #0x0]\n" - "str s24, [x22, #0x0]\n" + "str s12, [x26, #0x0]\n" + "str s16, [x25, #0x0]\n" + "str s20, [x24, #0x0]\n" + "str s24, [x23, #0x0]\n" "173:" // Height 5: Partial direct writeback: Done "b 175f\n" "174:" // Height 5: Full writeback @@ -2644,50 +2646,51 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "str q10, [x9, #0x20]\n" "str q11, [x9, #0x30]\n" "add x9, x9, #0x40\n" - "str q12, [x25, #0x0]\n" - "str q13, [x25, #0x10]\n" - "str q14, [x25, #0x20]\n" - "str q15, [x25, #0x30]\n" - "str q16, [x24, #0x0]\n" - "str q17, [x24, #0x10]\n" - "str q18, [x24, #0x20]\n" - "str q19, [x24, #0x30]\n" - "str q20, [x23, #0x0]\n" - "str q21, [x23, #0x10]\n" - "str q22, [x23, #0x20]\n" - "str q23, [x23, #0x30]\n" - "str q24, [x22, #0x0]\n" - "str q25, [x22, #0x10]\n" - "str q26, [x22, #0x20]\n" - "str q27, [x22, #0x30]\n" + "str q12, [x26, #0x0]\n" + "str q13, [x26, #0x10]\n" + "str q14, [x26, #0x20]\n" + "str q15, [x26, #0x30]\n" + "str q16, [x25, #0x0]\n" + "str q17, [x25, #0x10]\n" + "str q18, [x25, #0x20]\n" + "str q19, [x25, #0x30]\n" + "str q20, [x24, #0x0]\n" + "str q21, [x24, #0x10]\n" + "str q22, [x24, #0x20]\n" + "str q23, [x24, #0x30]\n" + "str q24, [x23, #0x0]\n" + "str q25, [x23, #0x10]\n" + "str q26, [x23, #0x20]\n" + "str q27, [x23, #0x30]\n" "175:" // Height 5: Writeback done "subs x11, x11, #0x10\n" "bgt 142b\n" "b 212f\n" "176:" // Height 6 "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n" + "ldr x9, [%x[args_ptr], %[offsetof_output_ptr]]\n" "mov x20, #0x18\n" - "mov x12, %x[bias]\n" + "ldr x12, [%x[args_ptr], %[offsetof_bias]]\n" "ldr x11, [%x[args_ptr], %[offsetof_N]]\n" "ldr x10, [%x[args_ptr], %[offsetof_B_ptr]]\n" - "mov x9, %x[output_ptr]\n" - "madd %x[output_ptr], x21, x20, %x[output_ptr]\n" + "madd x20, x21, x20, x9\n" + "str x20, [%x[args_ptr], %[offsetof_output_ptr]]\n" "177:" // Height 6: Column loop "cbz x12, 178f\n" "ldr q8, [x12, #0x0]\n" "ldr q9, [x12, #0x10]\n" - "mov v12.16b, v8.16b\n" - "mov v13.16b, v9.16b\n" "ldr q10, [x12, #0x20]\n" "ldr q11, [x12, #0x30]\n" - "mov v14.16b, v10.16b\n" - "mov v15.16b, v11.16b\n" + "add x12, x12, #0x40\n" + "mov v12.16b, v8.16b\n" + "mov v13.16b, v9.16b\n" "mov v16.16b, v8.16b\n" "mov v17.16b, v9.16b\n" - "add x12, x12, #0x40\n" + "mov v20.16b, v8.16b\n" + "mov v14.16b, v10.16b\n" + "mov v15.16b, v11.16b\n" "mov v18.16b, v10.16b\n" "mov v19.16b, v11.16b\n" - "mov v20.16b, v8.16b\n" "mov v21.16b, v9.16b\n" "mov v22.16b, v10.16b\n" "mov v23.16b, v11.16b\n" @@ -2703,145 +2706,145 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "178:" // Height 6: no bias "tbz %x[flags], #0, 188f\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" + "cmp x11, #0x10\n" + "add x26, x9, x20, LSL #2\n" + "add x25, x26, x20, LSL #2\n" "add x24, x25, x20, LSL #2\n" "add x23, x24, x20, LSL #2\n" "add x22, x23, x20, LSL #2\n" - "cmp x11, #0x10\n" - "add x21, x22, x20, LSL #2\n" "bge 187f\n" "tbz x11, #3, 182f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" - "ld1 { v24.4s }, [x22], #0x10\n" - "ld1 { v28.4s }, [x21], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" + "ld1 { v24.4s }, [x23], #0x10\n" + "ld1 { v28.4s }, [x22], #0x10\n" "ld1 { v9.4s }, [x9], #0x10\n" - "ld1 { v13.4s }, [x25], #0x10\n" - "ld1 { v17.4s }, [x24], #0x10\n" - "ld1 { v21.4s }, [x23], #0x10\n" - "ld1 { v25.4s }, [x22], #0x10\n" - "ld1 { v29.4s }, [x21], #0x10\n" + "ld1 { v13.4s }, [x26], #0x10\n" + "ld1 { v17.4s }, [x25], #0x10\n" + "ld1 { v21.4s }, [x24], #0x10\n" + "ld1 { v25.4s }, [x23], #0x10\n" + "ld1 { v29.4s }, [x22], #0x10\n" "tbz x11, #2, 180f\n" "ld1 { v10.4s }, [x9], #0x10\n" - "ld1 { v14.4s }, [x25], #0x10\n" - "ld1 { v18.4s }, [x24], #0x10\n" - "ld1 { v22.4s }, [x23], #0x10\n" - "ld1 { v26.4s }, [x22], #0x10\n" - "ld1 { v30.4s }, [x21], #0x10\n" + "ld1 { v14.4s }, [x26], #0x10\n" + "ld1 { v18.4s }, [x25], #0x10\n" + "ld1 { v22.4s }, [x24], #0x10\n" + "ld1 { v26.4s }, [x23], #0x10\n" + "ld1 { v30.4s }, [x22], #0x10\n" "tbz x11, #1, 179f\n" "ldr d11, [x9], #0x8\n" - "ldr d15, [x25], #0x8\n" + "ldr d15, [x26], #0x8\n" "mov x20, #0x38\n" - "ldr d19, [x24], #0x8\n" - "ldr d23, [x23], #0x8\n" - "ldr d27, [x22], #0x8\n" - "ldr d31, [x21], #0x8\n" + "ldr d19, [x25], #0x8\n" + "ldr d23, [x24], #0x8\n" + "ldr d27, [x23], #0x8\n" + "ldr d31, [x22], #0x8\n" "tbz x11, #0, 186f\n" "ld1 { v11.s }[2], [x9]\n" - "ld1 { v15.s }[2], [x25]\n" - "ld1 { v19.s }[2], [x24]\n" - "ld1 { v23.s }[2], [x23]\n" - "ld1 { v27.s }[2], [x22]\n" - "ld1 { v31.s }[2], [x21]\n" + "ld1 { v15.s }[2], [x26]\n" + "ld1 { v19.s }[2], [x25]\n" + "ld1 { v23.s }[2], [x24]\n" + "ld1 { v27.s }[2], [x23]\n" + "ld1 { v31.s }[2], [x22]\n" "b 186f\n" "179:" // Height 6: Partial accumulate: partial_1_12 "mov x20, #0x30\n" "tbz x11, #0, 186f\n" "ldr s11, [x9, #0x0]\n" - "ldr s15, [x25, #0x0]\n" - "ldr s19, [x24, #0x0]\n" - "ldr s23, [x23, #0x0]\n" - "ldr s27, [x22, #0x0]\n" - "ldr s31, [x21, #0x0]\n" + "ldr s15, [x26, #0x0]\n" + "ldr s19, [x25, #0x0]\n" + "ldr s23, [x24, #0x0]\n" + "ldr s27, [x23, #0x0]\n" + "ldr s31, [x22, #0x0]\n" "b 186f\n" "180:" // Height 6: Partial accumulate: partial_2_8 "tbz x11, #1, 181f\n" "ldr d10, [x9], #0x8\n" - "ldr d14, [x25], #0x8\n" + "ldr d14, [x26], #0x8\n" "mov x20, #0x28\n" - "ldr d18, [x24], #0x8\n" - "ldr d22, [x23], #0x8\n" - "ldr d26, [x22], #0x8\n" - "ldr d30, [x21], #0x8\n" + "ldr d18, [x25], #0x8\n" + "ldr d22, [x24], #0x8\n" + "ldr d26, [x23], #0x8\n" + "ldr d30, [x22], #0x8\n" "tbz x11, #0, 186f\n" "ld1 { v10.s }[2], [x9]\n" - "ld1 { v14.s }[2], [x25]\n" - "ld1 { v18.s }[2], [x24]\n" - "ld1 { v22.s }[2], [x23]\n" - "ld1 { v26.s }[2], [x22]\n" - "ld1 { v30.s }[2], [x21]\n" + "ld1 { v14.s }[2], [x26]\n" + "ld1 { v18.s }[2], [x25]\n" + "ld1 { v22.s }[2], [x24]\n" + "ld1 { v26.s }[2], [x23]\n" + "ld1 { v30.s }[2], [x22]\n" "b 186f\n" "181:" // Height 6: Partial accumulate: partial_1_8 "mov x20, #0x20\n" "tbz x11, #0, 186f\n" "ldr s10, [x9, #0x0]\n" - "ldr s14, [x25, #0x0]\n" - "ldr s18, [x24, #0x0]\n" - "ldr s22, [x23, #0x0]\n" - "ldr s26, [x22, #0x0]\n" - "ldr s30, [x21, #0x0]\n" + "ldr s14, [x26, #0x0]\n" + "ldr s18, [x25, #0x0]\n" + "ldr s22, [x24, #0x0]\n" + "ldr s26, [x23, #0x0]\n" + "ldr s30, [x22, #0x0]\n" "b 186f\n" "182:" // Height 6: Partial accumulate: partial_4_0 "tbz x11, #2, 184f\n" "ld1 { v8.4s }, [x9], #0x10\n" - "ld1 { v12.4s }, [x25], #0x10\n" - "ld1 { v16.4s }, [x24], #0x10\n" - "ld1 { v20.4s }, [x23], #0x10\n" - "ld1 { v24.4s }, [x22], #0x10\n" - "ld1 { v28.4s }, [x21], #0x10\n" + "ld1 { v12.4s }, [x26], #0x10\n" + "ld1 { v16.4s }, [x25], #0x10\n" + "ld1 { v20.4s }, [x24], #0x10\n" + "ld1 { v24.4s }, [x23], #0x10\n" + "ld1 { v28.4s }, [x22], #0x10\n" "tbz x11, #1, 183f\n" "ldr d9, [x9], #0x8\n" - "ldr d13, [x25], #0x8\n" + "ldr d13, [x26], #0x8\n" "mov x20, #0x18\n" - "ldr d17, [x24], #0x8\n" - "ldr d21, [x23], #0x8\n" - "ldr d25, [x22], #0x8\n" - "ldr d29, [x21], #0x8\n" + "ldr d17, [x25], #0x8\n" + "ldr d21, [x24], #0x8\n" + "ldr d25, [x23], #0x8\n" + "ldr d29, [x22], #0x8\n" "tbz x11, #0, 186f\n" "ld1 { v9.s }[2], [x9]\n" - "ld1 { v13.s }[2], [x25]\n" - "ld1 { v17.s }[2], [x24]\n" - "ld1 { v21.s }[2], [x23]\n" - "ld1 { v25.s }[2], [x22]\n" - "ld1 { v29.s }[2], [x21]\n" + "ld1 { v13.s }[2], [x26]\n" + "ld1 { v17.s }[2], [x25]\n" + "ld1 { v21.s }[2], [x24]\n" + "ld1 { v25.s }[2], [x23]\n" + "ld1 { v29.s }[2], [x22]\n" "b 186f\n" "183:" // Height 6: Partial accumulate: partial_1_4 "mov x20, #0x10\n" "tbz x11, #0, 186f\n" "ldr s9, [x9, #0x0]\n" - "ldr s13, [x25, #0x0]\n" - "ldr s17, [x24, #0x0]\n" - "ldr s21, [x23, #0x0]\n" - "ldr s25, [x22, #0x0]\n" - "ldr s29, [x21, #0x0]\n" + "ldr s13, [x26, #0x0]\n" + "ldr s17, [x25, #0x0]\n" + "ldr s21, [x24, #0x0]\n" + "ldr s25, [x23, #0x0]\n" + "ldr s29, [x22, #0x0]\n" "b 186f\n" "184:" // Height 6: Partial accumulate: partial_2_0 "tbz x11, #1, 185f\n" "ldr d8, [x9], #0x8\n" - "ldr d12, [x25], #0x8\n" + "ldr d12, [x26], #0x8\n" "mov x20, #0x8\n" - "ldr d16, [x24], #0x8\n" - "ldr d20, [x23], #0x8\n" - "ldr d24, [x22], #0x8\n" - "ldr d28, [x21], #0x8\n" + "ldr d16, [x25], #0x8\n" + "ldr d20, [x24], #0x8\n" + "ldr d24, [x23], #0x8\n" + "ldr d28, [x22], #0x8\n" "tbz x11, #0, 186f\n" "ld1 { v8.s }[2], [x9]\n" - "ld1 { v12.s }[2], [x25]\n" - "ld1 { v16.s }[2], [x24]\n" - "ld1 { v20.s }[2], [x23]\n" - "ld1 { v24.s }[2], [x22]\n" - "ld1 { v28.s }[2], [x21]\n" + "ld1 { v12.s }[2], [x26]\n" + "ld1 { v16.s }[2], [x25]\n" + "ld1 { v20.s }[2], [x24]\n" + "ld1 { v24.s }[2], [x23]\n" + "ld1 { v28.s }[2], [x22]\n" "b 186f\n" "185:" // Height 6: Partial accumulate: partial_1_0 "ldr s8, [x9, #0x0]\n" - "ldr s12, [x25, #0x0]\n" + "ldr s12, [x26, #0x0]\n" "mov x20, #0x0\n" - "ldr s16, [x24, #0x0]\n" - "ldr s20, [x23, #0x0]\n" - "ldr s24, [x22, #0x0]\n" - "ldr s28, [x21, #0x0]\n" + "ldr s16, [x25, #0x0]\n" + "ldr s20, [x24, #0x0]\n" + "ldr s24, [x23, #0x0]\n" + "ldr s28, [x22, #0x0]\n" "186:" // Height 6: Partial accumulate: Done "sub x9, x9, x20\n" "b 189f\n" @@ -2850,26 +2853,26 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr q9, [x9, #0x10]\n" "ldr q10, [x9, #0x20]\n" "ldr q11, [x9, #0x30]\n" - "ldr q12, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "ldr q14, [x25, #0x20]\n" - "ldr q15, [x25, #0x30]\n" - "ldr q16, [x24, #0x0]\n" - "ldr q17, [x24, #0x10]\n" - "ldr q18, [x24, #0x20]\n" - "ldr q19, [x24, #0x30]\n" - "ldr q20, [x23, #0x0]\n" - "ldr q21, [x23, #0x10]\n" - "ldr q22, [x23, #0x20]\n" - "ldr q23, [x23, #0x30]\n" - "ldr q24, [x22, #0x0]\n" - "ldr q25, [x22, #0x10]\n" - "ldr q26, [x22, #0x20]\n" - "ldr q27, [x22, #0x30]\n" - "ldr q28, [x21, #0x0]\n" - "ldr q29, [x21, #0x10]\n" - "ldr q30, [x21, #0x20]\n" - "ldr q31, [x21, #0x30]\n" + "ldr q12, [x26, #0x0]\n" + "ldr q13, [x26, #0x10]\n" + "ldr q14, [x26, #0x20]\n" + "ldr q15, [x26, #0x30]\n" + "ldr q16, [x25, #0x0]\n" + "ldr q17, [x25, #0x10]\n" + "ldr q18, [x25, #0x20]\n" + "ldr q19, [x25, #0x30]\n" + "ldr q20, [x24, #0x0]\n" + "ldr q21, [x24, #0x10]\n" + "ldr q22, [x24, #0x20]\n" + "ldr q23, [x24, #0x30]\n" + "ldr q24, [x23, #0x0]\n" + "ldr q25, [x23, #0x10]\n" + "ldr q26, [x23, #0x20]\n" + "ldr q27, [x23, #0x30]\n" + "ldr q28, [x22, #0x0]\n" + "ldr q29, [x22, #0x10]\n" + "ldr q30, [x22, #0x20]\n" + "ldr q31, [x22, #0x30]\n" "b 189f\n" "188:" // Height 6: no accumulate "movi v8.16b, #0x0\n" @@ -2900,8 +2903,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "mov x28, #0x0\n" "190:" // 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, 191f\n" "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" "add x20, x20, x21, LSL #3\n" @@ -3091,18 +3094,18 @@ void a64_hybrid_bf16fp32_dot_6x16 ( ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" "add x21, x21, #0x10\n" - "sub x27, x27, #0x8\n" + "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" ".inst 0x4f45f0fd // bfdot v29.4s, v7.8h, v5.h[0]\n" "ldr q7, [x10, #0x30]\n" - "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" + "sub x27, x27, #0x8\n" "prfm pldl1keep, [x22, #0x80]\n" ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" @@ -3208,9 +3211,9 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr s7, [x26], #0x4\n" "ldr s6, [x25], #0x4\n" "sub x27, x27, #0x2\n" - "cmp x27, #0x2\n" "ldr s5, [x24], #0x4\n" "ldr s4, [x23], #0x4\n" + "cmp x27, #0x2\n" "ldr s3, [x22], #0x4\n" "ldr s2, [x21], #0x4\n" "ldr q1, [x10, #0x0]\n" @@ -3287,21 +3290,21 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "cmp x28, x20\n" "bne 190b\n" "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n" - "add x25, x9, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x9, #0x0]\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x26, x9, x20, LSL #2\n" + "prfm pstl1keep, [x26, #0x0]\n" + "add x25, x26, x20, LSL #2\n" "prfm pstl1keep, [x25, #0x0]\n" + "add x24, x25, x20, LSL #2\n" "prfm pstl1keep, [x24, #0x0]\n" - "add x21, x22, x20, LSL #2\n" + "add x23, x24, x20, LSL #2\n" + "add x22, x23, x20, LSL #2\n" "prfm pstl1keep, [x23, #0x0]\n" "prfm pstl1keep, [x22, #0x0]\n" - "prfm pstl1keep, [x21, #0x0]\n" "tbz %x[flags], #1, 200f\n" - "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "add x21, %x[args_ptr], %[offset_max]\n" "add x20, %x[args_ptr], %[offset_min]\n" + "ld1r { v1.4s }, [x21]\n" "ld1r { v0.4s }, [x20]\n" "fmin v8.4s, v8.4s, v1.4s\n" "fmin v9.4s, v9.4s, v1.4s\n" @@ -3357,127 +3360,127 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "tbz x11, #3, 204f\n" "st1 { v8.4s }, [x9], #0x10\n" "st1 { v9.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v13.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v17.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" - "st1 { v21.4s }, [x23], #0x10\n" - "st1 { v24.4s }, [x22], #0x10\n" - "st1 { v25.4s }, [x22], #0x10\n" - "st1 { v28.4s }, [x21], #0x10\n" - "st1 { v29.4s }, [x21], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v13.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v17.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" + "st1 { v21.4s }, [x24], #0x10\n" + "st1 { v24.4s }, [x23], #0x10\n" + "st1 { v25.4s }, [x23], #0x10\n" + "st1 { v28.4s }, [x22], #0x10\n" + "st1 { v29.4s }, [x22], #0x10\n" "tbz x11, #2, 202f\n" "st1 { v10.4s }, [x9], #0x10\n" - "st1 { v14.4s }, [x25], #0x10\n" - "st1 { v18.4s }, [x24], #0x10\n" - "st1 { v22.4s }, [x23], #0x10\n" - "st1 { v26.4s }, [x22], #0x10\n" - "st1 { v30.4s }, [x21], #0x10\n" + "st1 { v14.4s }, [x26], #0x10\n" + "st1 { v18.4s }, [x25], #0x10\n" + "st1 { v22.4s }, [x24], #0x10\n" + "st1 { v26.4s }, [x23], #0x10\n" + "st1 { v30.4s }, [x22], #0x10\n" "tbz x11, #1, 201f\n" "str d11, [x9], #0x8\n" - "str d15, [x25], #0x8\n" - "str d19, [x24], #0x8\n" - "str d23, [x23], #0x8\n" - "str d27, [x22], #0x8\n" - "str d31, [x21], #0x8\n" + "str d15, [x26], #0x8\n" + "str d19, [x25], #0x8\n" + "str d23, [x24], #0x8\n" + "str d27, [x23], #0x8\n" + "str d31, [x22], #0x8\n" "tbz x11, #0, 208f\n" "st1 { v11.s }[2], [x9]\n" - "st1 { v15.s }[2], [x25]\n" - "st1 { v19.s }[2], [x24]\n" - "st1 { v23.s }[2], [x23]\n" - "st1 { v27.s }[2], [x22]\n" - "st1 { v31.s }[2], [x21]\n" + "st1 { v15.s }[2], [x26]\n" + "st1 { v19.s }[2], [x25]\n" + "st1 { v23.s }[2], [x24]\n" + "st1 { v27.s }[2], [x23]\n" + "st1 { v31.s }[2], [x22]\n" "b 208f\n" "201:" // Height 6: Partial direct writeback: partial_1_12 "tbz x11, #0, 208f\n" "str s11, [x9, #0x0]\n" - "str s15, [x25, #0x0]\n" - "str s19, [x24, #0x0]\n" - "str s23, [x23, #0x0]\n" - "str s27, [x22, #0x0]\n" - "str s31, [x21, #0x0]\n" + "str s15, [x26, #0x0]\n" + "str s19, [x25, #0x0]\n" + "str s23, [x24, #0x0]\n" + "str s27, [x23, #0x0]\n" + "str s31, [x22, #0x0]\n" "b 208f\n" "202:" // Height 6: Partial direct writeback: partial_2_8 "tbz x11, #1, 203f\n" "str d10, [x9], #0x8\n" - "str d14, [x25], #0x8\n" - "str d18, [x24], #0x8\n" - "str d22, [x23], #0x8\n" - "str d26, [x22], #0x8\n" - "str d30, [x21], #0x8\n" + "str d14, [x26], #0x8\n" + "str d18, [x25], #0x8\n" + "str d22, [x24], #0x8\n" + "str d26, [x23], #0x8\n" + "str d30, [x22], #0x8\n" "tbz x11, #0, 208f\n" "st1 { v10.s }[2], [x9]\n" - "st1 { v14.s }[2], [x25]\n" - "st1 { v18.s }[2], [x24]\n" - "st1 { v22.s }[2], [x23]\n" - "st1 { v26.s }[2], [x22]\n" - "st1 { v30.s }[2], [x21]\n" + "st1 { v14.s }[2], [x26]\n" + "st1 { v18.s }[2], [x25]\n" + "st1 { v22.s }[2], [x24]\n" + "st1 { v26.s }[2], [x23]\n" + "st1 { v30.s }[2], [x22]\n" "b 208f\n" "203:" // Height 6: Partial direct writeback: partial_1_8 "tbz x11, #0, 208f\n" "str s10, [x9, #0x0]\n" - "str s14, [x25, #0x0]\n" - "str s18, [x24, #0x0]\n" - "str s22, [x23, #0x0]\n" - "str s26, [x22, #0x0]\n" - "str s30, [x21, #0x0]\n" + "str s14, [x26, #0x0]\n" + "str s18, [x25, #0x0]\n" + "str s22, [x24, #0x0]\n" + "str s26, [x23, #0x0]\n" + "str s30, [x22, #0x0]\n" "b 208f\n" "204:" // Height 6: Partial direct writeback: partial_4_0 "tbz x11, #2, 206f\n" "st1 { v8.4s }, [x9], #0x10\n" - "st1 { v12.4s }, [x25], #0x10\n" - "st1 { v16.4s }, [x24], #0x10\n" - "st1 { v20.4s }, [x23], #0x10\n" - "st1 { v24.4s }, [x22], #0x10\n" - "st1 { v28.4s }, [x21], #0x10\n" + "st1 { v12.4s }, [x26], #0x10\n" + "st1 { v16.4s }, [x25], #0x10\n" + "st1 { v20.4s }, [x24], #0x10\n" + "st1 { v24.4s }, [x23], #0x10\n" + "st1 { v28.4s }, [x22], #0x10\n" "tbz x11, #1, 205f\n" "str d9, [x9], #0x8\n" - "str d13, [x25], #0x8\n" - "str d17, [x24], #0x8\n" - "str d21, [x23], #0x8\n" - "str d25, [x22], #0x8\n" - "str d29, [x21], #0x8\n" + "str d13, [x26], #0x8\n" + "str d17, [x25], #0x8\n" + "str d21, [x24], #0x8\n" + "str d25, [x23], #0x8\n" + "str d29, [x22], #0x8\n" "tbz x11, #0, 208f\n" "st1 { v9.s }[2], [x9]\n" - "st1 { v13.s }[2], [x25]\n" - "st1 { v17.s }[2], [x24]\n" - "st1 { v21.s }[2], [x23]\n" - "st1 { v25.s }[2], [x22]\n" - "st1 { v29.s }[2], [x21]\n" + "st1 { v13.s }[2], [x26]\n" + "st1 { v17.s }[2], [x25]\n" + "st1 { v21.s }[2], [x24]\n" + "st1 { v25.s }[2], [x23]\n" + "st1 { v29.s }[2], [x22]\n" "b 208f\n" "205:" // Height 6: Partial direct writeback: partial_1_4 "tbz x11, #0, 208f\n" "str s9, [x9, #0x0]\n" - "str s13, [x25, #0x0]\n" - "str s17, [x24, #0x0]\n" - "str s21, [x23, #0x0]\n" - "str s25, [x22, #0x0]\n" - "str s29, [x21, #0x0]\n" + "str s13, [x26, #0x0]\n" + "str s17, [x25, #0x0]\n" + "str s21, [x24, #0x0]\n" + "str s25, [x23, #0x0]\n" + "str s29, [x22, #0x0]\n" "b 208f\n" "206:" // Height 6: Partial direct writeback: partial_2_0 "tbz x11, #1, 207f\n" "str d8, [x9], #0x8\n" - "str d12, [x25], #0x8\n" - "str d16, [x24], #0x8\n" - "str d20, [x23], #0x8\n" - "str d24, [x22], #0x8\n" - "str d28, [x21], #0x8\n" + "str d12, [x26], #0x8\n" + "str d16, [x25], #0x8\n" + "str d20, [x24], #0x8\n" + "str d24, [x23], #0x8\n" + "str d28, [x22], #0x8\n" "tbz x11, #0, 208f\n" "st1 { v8.s }[2], [x9]\n" - "st1 { v12.s }[2], [x25]\n" - "st1 { v16.s }[2], [x24]\n" - "st1 { v20.s }[2], [x23]\n" - "st1 { v24.s }[2], [x22]\n" - "st1 { v28.s }[2], [x21]\n" + "st1 { v12.s }[2], [x26]\n" + "st1 { v16.s }[2], [x25]\n" + "st1 { v20.s }[2], [x24]\n" + "st1 { v24.s }[2], [x23]\n" + "st1 { v28.s }[2], [x22]\n" "b 208f\n" "207:" // Height 6: Partial direct writeback: partial_1_0 "str s8, [x9, #0x0]\n" - "str s12, [x25, #0x0]\n" - "str s16, [x24, #0x0]\n" - "str s20, [x23, #0x0]\n" - "str s24, [x22, #0x0]\n" - "str s28, [x21, #0x0]\n" + "str s12, [x26, #0x0]\n" + "str s16, [x25, #0x0]\n" + "str s20, [x24, #0x0]\n" + "str s24, [x23, #0x0]\n" + "str s28, [x22, #0x0]\n" "208:" // Height 6: Partial direct writeback: Done "b 210f\n" "209:" // Height 6: Full writeback @@ -3486,26 +3489,26 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "str q10, [x9, #0x20]\n" "str q11, [x9, #0x30]\n" "add x9, x9, #0x40\n" - "str q12, [x25, #0x0]\n" - "str q13, [x25, #0x10]\n" - "str q14, [x25, #0x20]\n" - "str q15, [x25, #0x30]\n" - "str q16, [x24, #0x0]\n" - "str q17, [x24, #0x10]\n" - "str q18, [x24, #0x20]\n" - "str q19, [x24, #0x30]\n" - "str q20, [x23, #0x0]\n" - "str q21, [x23, #0x10]\n" - "str q22, [x23, #0x20]\n" - "str q23, [x23, #0x30]\n" - "str q24, [x22, #0x0]\n" - "str q25, [x22, #0x10]\n" - "str q26, [x22, #0x20]\n" - "str q27, [x22, #0x30]\n" - "str q28, [x21, #0x0]\n" - "str q29, [x21, #0x10]\n" - "str q30, [x21, #0x20]\n" - "str q31, [x21, #0x30]\n" + "str q12, [x26, #0x0]\n" + "str q13, [x26, #0x10]\n" + "str q14, [x26, #0x20]\n" + "str q15, [x26, #0x30]\n" + "str q16, [x25, #0x0]\n" + "str q17, [x25, #0x10]\n" + "str q18, [x25, #0x20]\n" + "str q19, [x25, #0x30]\n" + "str q20, [x24, #0x0]\n" + "str q21, [x24, #0x10]\n" + "str q22, [x24, #0x20]\n" + "str q23, [x24, #0x30]\n" + "str q24, [x23, #0x0]\n" + "str q25, [x23, #0x10]\n" + "str q26, [x23, #0x20]\n" + "str q27, [x23, #0x30]\n" + "str q28, [x22, #0x0]\n" + "str q29, [x22, #0x10]\n" + "str q30, [x22, #0x20]\n" + "str q31, [x22, #0x30]\n" "210:" // Height 6: Writeback done "subs x11, x11, #0x10\n" "bgt 177b\n" @@ -3521,8 +3524,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "212:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) - : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) + : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr) + : [args_ptr] "r" (&ka), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_bias] "I" (offsetof(KernelArgs, bias)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_output_ptr] "I" (offsetof(KernelArgs, output_ptr)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } |