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 | 1962 |
1 files changed, 980 insertions, 982 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 02d2434356..fc323ea4fc 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 @@ -93,7 +93,6 @@ void a64_hybrid_bf16fp32_dot_6x16 ( break; } __asm__ __volatile__( - "1:" // Row loop "cmp %x[M], #0x6\n" "bge 176f\n" @@ -190,11 +189,11 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "15:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 16f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" "cbnz x28, 17f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -211,37 +210,37 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "blt 19f\n" "18:" // Height 1: Multiply loop: Main loop head ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - "ldr q7, [x10, #0xf0]\n" + "ldr q16, [x10, #0x30]\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" + "ldr q16, [x10, #0x50]\n" + ".inst 0x4f60f228 // bfdot v8.4s, v17.8h, v0.h[1]\n" + "ldr q17, [x10, #0x60]\n" + ".inst 0x4f60f209 // bfdot v9.4s, v16.8h, v0.h[1]\n" + "ldr q16, [x10, #0x70]\n" + ".inst 0x4f60f22a // bfdot v10.4s, v17.8h, v0.h[1]\n" + "ldr q17, [x10, #0x80]\n" + ".inst 0x4f60f20b // bfdot v11.4s, v16.8h, v0.h[1]\n" + "ldr q16, [x10, #0x90]\n" + ".inst 0x4f40fa28 // bfdot v8.4s, v17.8h, v0.h[2]\n" + "ldr q17, [x10, #0xa0]\n" + ".inst 0x4f40fa09 // bfdot v9.4s, v16.8h, v0.h[2]\n" + "ldr q16, [x10, #0xb0]\n" + ".inst 0x4f40fa2a // bfdot v10.4s, v17.8h, v0.h[2]\n" + "ldr q17, [x10, #0xc0]\n" + ".inst 0x4f40fa0b // bfdot v11.4s, v16.8h, v0.h[2]\n" + "ldr q16, [x10, #0xd0]\n" + ".inst 0x4f60fa28 // bfdot v8.4s, v17.8h, v0.h[3]\n" + "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" - ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" + ".inst 0x4f60fa2a // bfdot v10.4s, v17.8h, v0.h[3]\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" @@ -251,37 +250,37 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "bge 18b\n" "19:" // Height 1: Multiply loop: Single iteration only ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - "ldr q7, [x10, #0xf0]\n" + "ldr q16, [x10, #0x30]\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" + "ldr q16, [x10, #0x50]\n" + ".inst 0x4f60f228 // bfdot v8.4s, v17.8h, v0.h[1]\n" + "ldr q17, [x10, #0x60]\n" + ".inst 0x4f60f209 // bfdot v9.4s, v16.8h, v0.h[1]\n" + "ldr q16, [x10, #0x70]\n" + ".inst 0x4f60f22a // bfdot v10.4s, v17.8h, v0.h[1]\n" + "ldr q17, [x10, #0x80]\n" + ".inst 0x4f60f20b // bfdot v11.4s, v16.8h, v0.h[1]\n" + "ldr q16, [x10, #0x90]\n" + ".inst 0x4f40fa28 // bfdot v8.4s, v17.8h, v0.h[2]\n" + "ldr q17, [x10, #0xa0]\n" + ".inst 0x4f40fa09 // bfdot v9.4s, v16.8h, v0.h[2]\n" + "ldr q16, [x10, #0xb0]\n" + ".inst 0x4f40fa2a // bfdot v10.4s, v17.8h, v0.h[2]\n" + "ldr q17, [x10, #0xc0]\n" + ".inst 0x4f40fa0b // bfdot v11.4s, v16.8h, v0.h[2]\n" + "ldr q16, [x10, #0xd0]\n" + ".inst 0x4f60fa28 // bfdot v8.4s, v17.8h, v0.h[3]\n" + "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" - ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\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 @@ -289,31 +288,31 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "cmp x27, #0x2\n" "blt 22f\n" "21:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr q6, [x10, #0x0]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" + "ldr s18, [x26], #0x4\n" + "ldr q16, [x10, #0x0]\n" + ".inst 0x4f52f208 // bfdot v8.4s, v16.8h, v18.h[0]\n" "sub x27, x27, #0x2\n" - "ldr q7, [x10, #0x10]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" + "ldr q16, [x10, #0x10]\n" + "ldr q17, [x10, #0x20]\n" + ".inst 0x4f52f209 // bfdot v9.4s, v16.8h, v18.h[0]\n" "cmp x27, #0x2\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" + "ldr q16, [x10, #0x30]\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" "ldr h0, [x26, #0x0]\n" "23:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - "ldr q6, [x10, #0x20]\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" + "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" + "ldr q16, [x10, #0x30]\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" @@ -323,17 +322,17 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "prfm pstl1keep, [x9, #0x0]\n" "tbz %x[flags], #1, 25f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" + "ld1r { v16.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v17.4s\n" + "fmin v9.4s, v9.4s, v17.4s\n" + "fmin v10.4s, v10.4s, v17.4s\n" + "fmin v11.4s, v11.4s, v17.4s\n" + "fmax v8.4s, v8.4s, v16.4s\n" + "fmax v9.4s, v9.4s, v16.4s\n" + "fmax v10.4s, v10.4s, v16.4s\n" + "fmax v11.4s, v11.4s, v16.4s\n" "25:" // Height 1: No activation "cmp x11, #0x10\n" "bge 34f\n" @@ -511,12 +510,12 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "50:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 51f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" "cbnz x28, 52f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -524,7 +523,7 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "b 52f\n" "51:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" "52:" // Height 2: input setup done "cmp x27, #0x8\n" "blt 55f\n" @@ -537,156 +536,156 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "53:" // Height 2: Multiply loop: Main loop head ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "sub x27, x27, #0x8\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q16, [x10, #0x30]\n" "add x26, x26, #0x10\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - "ldr q6, [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" + "ldr q17, [x10, #0x40]\n" "add x25, x25, #0x10\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - "ldr q7, [x10, #0x50]\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 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" - "ldr q6, [x10, #0x60]\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 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" - "ldr q7, [x10, #0x70]\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 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" - "ldr q7, [x10, #0xf0]\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" + ".inst 0x4f60f20b // bfdot v11.4s, v16.8h, v0.h[1]\n" + ".inst 0x4f61f20f // bfdot v15.4s, v16.8h, v1.h[1]\n" + "ldr q16, [x10, #0x90]\n" + ".inst 0x4f40fa28 // bfdot v8.4s, v17.8h, v0.h[2]\n" + ".inst 0x4f41fa2c // bfdot v12.4s, v17.8h, v1.h[2]\n" + "ldr q17, [x10, #0xa0]\n" + ".inst 0x4f40fa09 // bfdot v9.4s, v16.8h, v0.h[2]\n" + ".inst 0x4f41fa0d // bfdot v13.4s, v16.8h, v1.h[2]\n" + "ldr q16, [x10, #0xb0]\n" + ".inst 0x4f40fa2a // bfdot v10.4s, v17.8h, v0.h[2]\n" + ".inst 0x4f41fa2e // bfdot v14.4s, v17.8h, v1.h[2]\n" + "ldr q17, [x10, #0xc0]\n" + ".inst 0x4f40fa0b // bfdot v11.4s, v16.8h, v0.h[2]\n" + ".inst 0x4f41fa0f // bfdot v15.4s, v16.8h, v1.h[2]\n" + "ldr q16, [x10, #0xd0]\n" + ".inst 0x4f60fa28 // bfdot v8.4s, v17.8h, v0.h[3]\n" + ".inst 0x4f61fa2c // bfdot v12.4s, v17.8h, v1.h[3]\n" + "ldr q17, [x10, #0xe0]\n" + ".inst 0x4f60fa09 // bfdot v9.4s, v16.8h, v0.h[3]\n" + ".inst 0x4f61fa0d // bfdot v13.4s, v16.8h, v1.h[3]\n" + "ldr q16, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" + ".inst 0x4f60fa2a // bfdot v10.4s, v17.8h, v0.h[3]\n" + ".inst 0x4f61fa2e // bfdot v14.4s, v17.8h, v1.h[3]\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" + ".inst 0x4f60fa0b // bfdot v11.4s, v16.8h, v0.h[3]\n" "ldr q0, [x26, #0x0]\n" - ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" + ".inst 0x4f61fa0f // bfdot v15.4s, v16.8h, v1.h[3]\n" "ldr q1, [x25, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 53b\n" "54:" // Height 2: Multiply loop: Single iteration only ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "add x26, x26, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q16, [x10, #0x30]\n" "add x25, x25, #0x10\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - "ldr q6, [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" + "ldr q17, [x10, #0x40]\n" "sub x27, x27, #0x8\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - "ldr q7, [x10, #0x50]\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 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" - "ldr q6, [x10, #0x60]\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 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" - "ldr q7, [x10, #0xf0]\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" + ".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" + ".inst 0x4f60f20b // bfdot v11.4s, v16.8h, v0.h[1]\n" + ".inst 0x4f61f20f // bfdot v15.4s, v16.8h, v1.h[1]\n" + "ldr q16, [x10, #0x90]\n" + ".inst 0x4f40fa28 // bfdot v8.4s, v17.8h, v0.h[2]\n" + ".inst 0x4f41fa2c // bfdot v12.4s, v17.8h, v1.h[2]\n" + "ldr q17, [x10, #0xa0]\n" + ".inst 0x4f40fa09 // bfdot v9.4s, v16.8h, v0.h[2]\n" + ".inst 0x4f41fa0d // bfdot v13.4s, v16.8h, v1.h[2]\n" + "ldr q16, [x10, #0xb0]\n" + ".inst 0x4f40fa2a // bfdot v10.4s, v17.8h, v0.h[2]\n" + ".inst 0x4f41fa2e // bfdot v14.4s, v17.8h, v1.h[2]\n" + "ldr q17, [x10, #0xc0]\n" + ".inst 0x4f40fa0b // bfdot v11.4s, v16.8h, v0.h[2]\n" + ".inst 0x4f41fa0f // bfdot v15.4s, v16.8h, v1.h[2]\n" + "ldr q16, [x10, #0xd0]\n" + ".inst 0x4f60fa28 // bfdot v8.4s, v17.8h, v0.h[3]\n" + ".inst 0x4f61fa2c // bfdot v12.4s, v17.8h, v1.h[3]\n" + "ldr q17, [x10, #0xe0]\n" + ".inst 0x4f60fa09 // bfdot v9.4s, v16.8h, v0.h[3]\n" + ".inst 0x4f61fa0d // bfdot v13.4s, v16.8h, v1.h[3]\n" + "ldr q16, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" + ".inst 0x4f60fa2a // bfdot v10.4s, v17.8h, v0.h[3]\n" + ".inst 0x4f61fa2e // bfdot v14.4s, v17.8h, v1.h[3]\n" + ".inst 0x4f60fa0b // bfdot v11.4s, v16.8h, v0.h[3]\n" + ".inst 0x4f61fa0f // bfdot v15.4s, v16.8h, v1.h[3]\n" "55:" // Height 2: Multiply loop: Main loop skip "cbz x27, 59f\n" "cmp x27, #0x2\n" "blt 57f\n" "56:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr s1, [x25], #0x4\n" + "ldr s19, [x26], #0x4\n" + "ldr s18, [x25], #0x4\n" "sub x27, x27, #0x2\n" "cmp x27, #0x2\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" + "ldr q17, [x10, #0x0]\n" + "ldr q16, [x10, #0x10]\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" + ".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 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\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" "57:" // Height 2: Multiply loop: Skip odd blocks "cbz x27, 59f\n" "ldr h0, [x26, #0x0]\n" "ldr h1, [x25, #0x0]\n" "58:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" + "ldr q17, [x10, #0x0]\n" + "ldr q16, [x10, #0x10]\n" + ".inst 0x4f40f228 // bfdot v8.4s, v17.8h, v0.h[0]\n" + ".inst 0x4f41f22c // bfdot v12.4s, v17.8h, v1.h[0]\n" + "ldr q17, [x10, #0x20]\n" + ".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" + ".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 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\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 "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -698,25 +697,25 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "prfm pstl1keep, [x25, #0x0]\n" "tbz %x[flags], #1, 60f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" + "ld1r { v16.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v17.4s\n" + "fmin v9.4s, v9.4s, v17.4s\n" + "fmin v10.4s, v10.4s, v17.4s\n" + "fmin v11.4s, v11.4s, v17.4s\n" + "fmin v12.4s, v12.4s, v17.4s\n" + "fmin v13.4s, v13.4s, v17.4s\n" + "fmin v14.4s, v14.4s, v17.4s\n" + "fmin v15.4s, v15.4s, v17.4s\n" + "fmax v8.4s, v8.4s, v16.4s\n" + "fmax v9.4s, v9.4s, v16.4s\n" + "fmax v10.4s, v10.4s, v16.4s\n" + "fmax v11.4s, v11.4s, v16.4s\n" + "fmax v12.4s, v12.4s, v16.4s\n" + "fmax v13.4s, v13.4s, v16.4s\n" + "fmax v14.4s, v14.4s, v16.4s\n" + "fmax v15.4s, v15.4s, v16.4s\n" "60:" // Height 2: No activation "cmp x11, #0x10\n" "bge 69f\n" @@ -943,13 +942,13 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "85:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 86f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" "cbnz x28, 87f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -958,8 +957,8 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "b 87f\n" "86:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" "87:" // Height 3: input setup done "cmp x27, #0x8\n" "blt 90f\n" @@ -976,75 +975,75 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "sub x27, x27, #0x8\n" "add x26, x26, #0x10\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q21, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" "add x25, x25, #0x10\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q20, [x10, #0x30]\n" "add x24, x24, #0x10\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\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" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.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, [x25, #0x80]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - "ldr q7, [x10, #0x50]\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 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" - ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" - "ldr q7, [x10, #0xf0]\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" + "ldr q21, [x10, #0x60]\n" + ".inst 0x4f60f289 // bfdot v9.4s, v20.8h, v0.h[1]\n" + ".inst 0x4f61f28d // bfdot v13.4s, v20.8h, v1.h[1]\n" + ".inst 0x4f62f291 // bfdot v17.4s, v20.8h, v2.h[1]\n" + "ldr q20, [x10, #0x70]\n" + ".inst 0x4f60f2aa // bfdot v10.4s, v21.8h, v0.h[1]\n" + ".inst 0x4f61f2ae // bfdot v14.4s, v21.8h, v1.h[1]\n" + ".inst 0x4f62f2b2 // bfdot v18.4s, v21.8h, v2.h[1]\n" + "ldr q21, [x10, #0x80]\n" + ".inst 0x4f60f28b // bfdot v11.4s, v20.8h, v0.h[1]\n" + ".inst 0x4f61f28f // bfdot v15.4s, v20.8h, v1.h[1]\n" + ".inst 0x4f62f293 // bfdot v19.4s, v20.8h, v2.h[1]\n" + "ldr q20, [x10, #0x90]\n" + ".inst 0x4f40faa8 // bfdot v8.4s, v21.8h, v0.h[2]\n" + ".inst 0x4f41faac // bfdot v12.4s, v21.8h, v1.h[2]\n" + ".inst 0x4f42fab0 // bfdot v16.4s, v21.8h, v2.h[2]\n" + "ldr q21, [x10, #0xa0]\n" + ".inst 0x4f40fa89 // bfdot v9.4s, v20.8h, v0.h[2]\n" + ".inst 0x4f41fa8d // bfdot v13.4s, v20.8h, v1.h[2]\n" + ".inst 0x4f42fa91 // bfdot v17.4s, v20.8h, v2.h[2]\n" + "ldr q20, [x10, #0xb0]\n" + ".inst 0x4f40faaa // bfdot v10.4s, v21.8h, v0.h[2]\n" + ".inst 0x4f41faae // bfdot v14.4s, v21.8h, v1.h[2]\n" + ".inst 0x4f42fab2 // bfdot v18.4s, v21.8h, v2.h[2]\n" + "ldr q21, [x10, #0xc0]\n" + ".inst 0x4f40fa8b // bfdot v11.4s, v20.8h, v0.h[2]\n" + ".inst 0x4f41fa8f // bfdot v15.4s, v20.8h, v1.h[2]\n" + ".inst 0x4f42fa93 // bfdot v19.4s, v20.8h, v2.h[2]\n" + "ldr q20, [x10, #0xd0]\n" + ".inst 0x4f60faa8 // bfdot v8.4s, v21.8h, v0.h[3]\n" + ".inst 0x4f61faac // bfdot v12.4s, v21.8h, v1.h[3]\n" + ".inst 0x4f62fab0 // bfdot v16.4s, v21.8h, v2.h[3]\n" + "ldr q21, [x10, #0xe0]\n" + ".inst 0x4f60fa89 // bfdot v9.4s, v20.8h, v0.h[3]\n" + ".inst 0x4f61fa8d // bfdot v13.4s, v20.8h, v1.h[3]\n" + ".inst 0x4f62fa91 // bfdot v17.4s, v20.8h, v2.h[3]\n" + "ldr q20, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" + ".inst 0x4f60faaa // bfdot v10.4s, v21.8h, v0.h[3]\n" + ".inst 0x4f61faae // bfdot v14.4s, v21.8h, v1.h[3]\n" + ".inst 0x4f62fab2 // bfdot v18.4s, v21.8h, v2.h[3]\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" + ".inst 0x4f60fa8b // bfdot v11.4s, v20.8h, v0.h[3]\n" "ldr q0, [x26, #0x0]\n" - ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" + ".inst 0x4f61fa8f // bfdot v15.4s, v20.8h, v1.h[3]\n" "ldr q1, [x25, #0x0]\n" - ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" + ".inst 0x4f62fa93 // bfdot v19.4s, v20.8h, v2.h[3]\n" "ldr q2, [x24, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 88b\n" @@ -1054,98 +1053,98 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q21, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" "add x24, x24, #0x10\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q20, [x10, #0x30]\n" "sub x27, x27, #0x8\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\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 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.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 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" - ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" - "ldr q7, [x10, #0xf0]\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" + ".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" + "ldr q21, [x10, #0x60]\n" + ".inst 0x4f60f289 // bfdot v9.4s, v20.8h, v0.h[1]\n" + ".inst 0x4f61f28d // bfdot v13.4s, v20.8h, v1.h[1]\n" + ".inst 0x4f62f291 // bfdot v17.4s, v20.8h, v2.h[1]\n" + "ldr q20, [x10, #0x70]\n" + ".inst 0x4f60f2aa // bfdot v10.4s, v21.8h, v0.h[1]\n" + ".inst 0x4f61f2ae // bfdot v14.4s, v21.8h, v1.h[1]\n" + ".inst 0x4f62f2b2 // bfdot v18.4s, v21.8h, v2.h[1]\n" + "ldr q21, [x10, #0x80]\n" + ".inst 0x4f60f28b // bfdot v11.4s, v20.8h, v0.h[1]\n" + ".inst 0x4f61f28f // bfdot v15.4s, v20.8h, v1.h[1]\n" + ".inst 0x4f62f293 // bfdot v19.4s, v20.8h, v2.h[1]\n" + "ldr q20, [x10, #0x90]\n" + ".inst 0x4f40faa8 // bfdot v8.4s, v21.8h, v0.h[2]\n" + ".inst 0x4f41faac // bfdot v12.4s, v21.8h, v1.h[2]\n" + ".inst 0x4f42fab0 // bfdot v16.4s, v21.8h, v2.h[2]\n" + "ldr q21, [x10, #0xa0]\n" + ".inst 0x4f40fa89 // bfdot v9.4s, v20.8h, v0.h[2]\n" + ".inst 0x4f41fa8d // bfdot v13.4s, v20.8h, v1.h[2]\n" + ".inst 0x4f42fa91 // bfdot v17.4s, v20.8h, v2.h[2]\n" + "ldr q20, [x10, #0xb0]\n" + ".inst 0x4f40faaa // bfdot v10.4s, v21.8h, v0.h[2]\n" + ".inst 0x4f41faae // bfdot v14.4s, v21.8h, v1.h[2]\n" + ".inst 0x4f42fab2 // bfdot v18.4s, v21.8h, v2.h[2]\n" + "ldr q21, [x10, #0xc0]\n" + ".inst 0x4f40fa8b // bfdot v11.4s, v20.8h, v0.h[2]\n" + ".inst 0x4f41fa8f // bfdot v15.4s, v20.8h, v1.h[2]\n" + ".inst 0x4f42fa93 // bfdot v19.4s, v20.8h, v2.h[2]\n" + "ldr q20, [x10, #0xd0]\n" + ".inst 0x4f60faa8 // bfdot v8.4s, v21.8h, v0.h[3]\n" + ".inst 0x4f61faac // bfdot v12.4s, v21.8h, v1.h[3]\n" + ".inst 0x4f62fab0 // bfdot v16.4s, v21.8h, v2.h[3]\n" + "ldr q21, [x10, #0xe0]\n" + ".inst 0x4f60fa89 // bfdot v9.4s, v20.8h, v0.h[3]\n" + ".inst 0x4f61fa8d // bfdot v13.4s, v20.8h, v1.h[3]\n" + ".inst 0x4f62fa91 // bfdot v17.4s, v20.8h, v2.h[3]\n" + "ldr q20, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" - ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" - ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" + ".inst 0x4f60faaa // bfdot v10.4s, v21.8h, v0.h[3]\n" + ".inst 0x4f61faae // bfdot v14.4s, v21.8h, v1.h[3]\n" + ".inst 0x4f62fab2 // bfdot v18.4s, v21.8h, v2.h[3]\n" + ".inst 0x4f60fa8b // bfdot v11.4s, v20.8h, v0.h[3]\n" + ".inst 0x4f61fa8f // bfdot v15.4s, v20.8h, v1.h[3]\n" + ".inst 0x4f62fa93 // bfdot v19.4s, v20.8h, v2.h[3]\n" "90:" // Height 3: Multiply loop: Main loop skip "cbz x27, 94f\n" "cmp x27, #0x2\n" "blt 92f\n" "91:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr s1, [x25], #0x4\n" + "ldr s24, [x26], #0x4\n" + "ldr s23, [x25], #0x4\n" "sub x27, x27, #0x2\n" "cmp x27, #0x2\n" - "ldr s2, [x24], #0x4\n" - "ldr q6, [x10, #0x0]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr s22, [x24], #0x4\n" + "ldr q21, [x10, #0x0]\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" + ".inst 0x4f57f28d // bfdot v13.4s, v20.8h, v23.h[0]\n" + ".inst 0x4f56f291 // bfdot v17.4s, v20.8h, v22.h[0]\n" + "ldr q20, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" + ".inst 0x4f58f2aa // bfdot v10.4s, v21.8h, v24.h[0]\n" + ".inst 0x4f57f2ae // bfdot v14.4s, v21.8h, v23.h[0]\n" + ".inst 0x4f56f2b2 // bfdot v18.4s, v21.8h, v22.h[0]\n" + ".inst 0x4f58f28b // bfdot v11.4s, v20.8h, v24.h[0]\n" + ".inst 0x4f57f28f // bfdot v15.4s, v20.8h, v23.h[0]\n" + ".inst 0x4f56f293 // bfdot v19.4s, v20.8h, v22.h[0]\n" "bge 91b\n" "92:" // Height 3: Multiply loop: Skip odd blocks "cbz x27, 94f\n" @@ -1153,23 +1152,23 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr h1, [x25, #0x0]\n" "ldr h2, [x24, #0x0]\n" "93:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q21, [x10, #0x0]\n" + "ldr q20, [x10, #0x10]\n" + ".inst 0x4f40f2a8 // bfdot v8.4s, v21.8h, v0.h[0]\n" + ".inst 0x4f41f2ac // bfdot v12.4s, v21.8h, v1.h[0]\n" + ".inst 0x4f42f2b0 // bfdot v16.4s, v21.8h, v2.h[0]\n" + "ldr q21, [x10, #0x20]\n" + ".inst 0x4f40f289 // bfdot v9.4s, v20.8h, v0.h[0]\n" + ".inst 0x4f41f28d // bfdot v13.4s, v20.8h, v1.h[0]\n" + ".inst 0x4f42f291 // bfdot v17.4s, v20.8h, v2.h[0]\n" + "ldr q20, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\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" + ".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" "94:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1183,33 +1182,33 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "prfm pstl1keep, [x24, #0x0]\n" "tbz %x[flags], #1, 95f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v21.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmin v16.4s, v16.4s, v1.4s\n" - "fmin v17.4s, v17.4s, v1.4s\n" - "fmin v18.4s, v18.4s, v1.4s\n" - "fmin v19.4s, v19.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" - "fmax v16.4s, v16.4s, v0.4s\n" - "fmax v17.4s, v17.4s, v0.4s\n" - "fmax v18.4s, v18.4s, v0.4s\n" - "fmax v19.4s, v19.4s, v0.4s\n" + "ld1r { v20.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v21.4s\n" + "fmin v9.4s, v9.4s, v21.4s\n" + "fmin v10.4s, v10.4s, v21.4s\n" + "fmin v11.4s, v11.4s, v21.4s\n" + "fmin v12.4s, v12.4s, v21.4s\n" + "fmin v13.4s, v13.4s, v21.4s\n" + "fmin v14.4s, v14.4s, v21.4s\n" + "fmin v15.4s, v15.4s, v21.4s\n" + "fmin v16.4s, v16.4s, v21.4s\n" + "fmin v17.4s, v17.4s, v21.4s\n" + "fmin v18.4s, v18.4s, v21.4s\n" + "fmin v19.4s, v19.4s, v21.4s\n" + "fmax v8.4s, v8.4s, v20.4s\n" + "fmax v9.4s, v9.4s, v20.4s\n" + "fmax v10.4s, v10.4s, v20.4s\n" + "fmax v11.4s, v11.4s, v20.4s\n" + "fmax v12.4s, v12.4s, v20.4s\n" + "fmax v13.4s, v13.4s, v20.4s\n" + "fmax v14.4s, v14.4s, v20.4s\n" + "fmax v15.4s, v15.4s, v20.4s\n" + "fmax v16.4s, v16.4s, v20.4s\n" + "fmax v17.4s, v17.4s, v20.4s\n" + "fmax v18.4s, v18.4s, v20.4s\n" + "fmax v19.4s, v19.4s, v20.4s\n" "95:" // Height 3: No activation "cmp x11, #0x10\n" "bge 104f\n" @@ -1485,14 +1484,14 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "120:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 121f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" "cbnz x28, 122f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -1502,9 +1501,9 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "b 122f\n" "121:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" "122:" // Height 4: input setup done "cmp x27, #0x8\n" "blt 125f\n" @@ -1523,7 +1522,7 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "add x26, x26, #0x10\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q25, [x10, #0x20]\n" "add x25, x25, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" @@ -1531,85 +1530,85 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "add x23, x23, #0x10\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q24, [x10, #0x30]\n" "cmp x27, #0x10\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\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, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" - "ldr q6, [x10, #0x40]\n" + ".inst 0x4f42f332 // bfdot v18.4s, v25.8h, v2.h[0]\n" + ".inst 0x4f43f336 // bfdot v22.4s, v25.8h, v3.h[0]\n" + "ldr q25, [x10, #0x40]\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" + ".inst 0x4f40f30b // bfdot v11.4s, v24.8h, v0.h[0]\n" + ".inst 0x4f41f30f // bfdot v15.4s, v24.8h, v1.h[0]\n" "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" - ".inst 0x4f63f0d4 // bfdot v20.4s, v6.8h, v3.h[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" - ".inst 0x4f63f0f5 // bfdot v21.4s, v7.8h, v3.h[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" - ".inst 0x4f63f0d6 // bfdot v22.4s, v6.8h, v3.h[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" - ".inst 0x4f63f0f7 // bfdot v23.4s, v7.8h, v3.h[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" - ".inst 0x4f43f8d4 // bfdot v20.4s, v6.8h, v3.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" - ".inst 0x4f43f8f5 // bfdot v21.4s, v7.8h, v3.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" - ".inst 0x4f43f8d6 // bfdot v22.4s, v6.8h, v3.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" - ".inst 0x4f43f8f7 // bfdot v23.4s, v7.8h, v3.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" - ".inst 0x4f63f8d4 // bfdot v20.4s, v6.8h, v3.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" - ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" - ".inst 0x4f63f8f5 // bfdot v21.4s, v7.8h, v3.h[3]\n" - "ldr q7, [x10, #0xf0]\n" + ".inst 0x4f42f313 // bfdot v19.4s, v24.8h, v2.h[0]\n" + ".inst 0x4f43f317 // bfdot v23.4s, v24.8h, v3.h[0]\n" + "ldr q24, [x10, #0x50]\n" + ".inst 0x4f60f328 // bfdot v8.4s, v25.8h, v0.h[1]\n" + ".inst 0x4f61f32c // bfdot v12.4s, v25.8h, v1.h[1]\n" + ".inst 0x4f62f330 // bfdot v16.4s, v25.8h, v2.h[1]\n" + ".inst 0x4f63f334 // bfdot v20.4s, v25.8h, v3.h[1]\n" + "ldr q25, [x10, #0x60]\n" + ".inst 0x4f60f309 // bfdot v9.4s, v24.8h, v0.h[1]\n" + ".inst 0x4f61f30d // bfdot v13.4s, v24.8h, v1.h[1]\n" + ".inst 0x4f62f311 // bfdot v17.4s, v24.8h, v2.h[1]\n" + ".inst 0x4f63f315 // bfdot v21.4s, v24.8h, v3.h[1]\n" + "ldr q24, [x10, #0x70]\n" + ".inst 0x4f60f32a // bfdot v10.4s, v25.8h, v0.h[1]\n" + ".inst 0x4f61f32e // bfdot v14.4s, v25.8h, v1.h[1]\n" + ".inst 0x4f62f332 // bfdot v18.4s, v25.8h, v2.h[1]\n" + ".inst 0x4f63f336 // bfdot v22.4s, v25.8h, v3.h[1]\n" + "ldr q25, [x10, #0x80]\n" + ".inst 0x4f60f30b // bfdot v11.4s, v24.8h, v0.h[1]\n" + ".inst 0x4f61f30f // bfdot v15.4s, v24.8h, v1.h[1]\n" + ".inst 0x4f62f313 // bfdot v19.4s, v24.8h, v2.h[1]\n" + ".inst 0x4f63f317 // bfdot v23.4s, v24.8h, v3.h[1]\n" + "ldr q24, [x10, #0x90]\n" + ".inst 0x4f40fb28 // bfdot v8.4s, v25.8h, v0.h[2]\n" + ".inst 0x4f41fb2c // bfdot v12.4s, v25.8h, v1.h[2]\n" + ".inst 0x4f42fb30 // bfdot v16.4s, v25.8h, v2.h[2]\n" + ".inst 0x4f43fb34 // bfdot v20.4s, v25.8h, v3.h[2]\n" + "ldr q25, [x10, #0xa0]\n" + ".inst 0x4f40fb09 // bfdot v9.4s, v24.8h, v0.h[2]\n" + ".inst 0x4f41fb0d // bfdot v13.4s, v24.8h, v1.h[2]\n" + ".inst 0x4f42fb11 // bfdot v17.4s, v24.8h, v2.h[2]\n" + ".inst 0x4f43fb15 // bfdot v21.4s, v24.8h, v3.h[2]\n" + "ldr q24, [x10, #0xb0]\n" + ".inst 0x4f40fb2a // bfdot v10.4s, v25.8h, v0.h[2]\n" + ".inst 0x4f41fb2e // bfdot v14.4s, v25.8h, v1.h[2]\n" + ".inst 0x4f42fb32 // bfdot v18.4s, v25.8h, v2.h[2]\n" + ".inst 0x4f43fb36 // bfdot v22.4s, v25.8h, v3.h[2]\n" + "ldr q25, [x10, #0xc0]\n" + ".inst 0x4f40fb0b // bfdot v11.4s, v24.8h, v0.h[2]\n" + ".inst 0x4f41fb0f // bfdot v15.4s, v24.8h, v1.h[2]\n" + ".inst 0x4f42fb13 // bfdot v19.4s, v24.8h, v2.h[2]\n" + ".inst 0x4f43fb17 // bfdot v23.4s, v24.8h, v3.h[2]\n" + "ldr q24, [x10, #0xd0]\n" + ".inst 0x4f60fb28 // bfdot v8.4s, v25.8h, v0.h[3]\n" + ".inst 0x4f61fb2c // bfdot v12.4s, v25.8h, v1.h[3]\n" + ".inst 0x4f62fb30 // bfdot v16.4s, v25.8h, v2.h[3]\n" + ".inst 0x4f63fb34 // bfdot v20.4s, v25.8h, v3.h[3]\n" + "ldr q25, [x10, #0xe0]\n" + ".inst 0x4f60fb09 // bfdot v9.4s, v24.8h, v0.h[3]\n" + ".inst 0x4f61fb0d // bfdot v13.4s, v24.8h, v1.h[3]\n" + ".inst 0x4f62fb11 // bfdot v17.4s, v24.8h, v2.h[3]\n" + ".inst 0x4f63fb15 // bfdot v21.4s, v24.8h, v3.h[3]\n" + "ldr q24, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" - ".inst 0x4f63f8d6 // bfdot v22.4s, v6.8h, v3.h[3]\n" + ".inst 0x4f60fb2a // bfdot v10.4s, v25.8h, v0.h[3]\n" + ".inst 0x4f61fb2e // bfdot v14.4s, v25.8h, v1.h[3]\n" + ".inst 0x4f62fb32 // bfdot v18.4s, v25.8h, v2.h[3]\n" + ".inst 0x4f63fb36 // bfdot v22.4s, v25.8h, v3.h[3]\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" + ".inst 0x4f60fb0b // bfdot v11.4s, v24.8h, v0.h[3]\n" "ldr q0, [x26, #0x0]\n" - ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" + ".inst 0x4f61fb0f // bfdot v15.4s, v24.8h, v1.h[3]\n" "ldr q1, [x25, #0x0]\n" - ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" + ".inst 0x4f62fb13 // bfdot v19.4s, v24.8h, v2.h[3]\n" "ldr q2, [x24, #0x0]\n" - ".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n" + ".inst 0x4f63fb17 // bfdot v23.4s, v24.8h, v3.h[3]\n" "ldr q3, [x23, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 123b\n" @@ -1620,7 +1619,7 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "add x25, x25, #0x10\n" ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q25, [x10, #0x20]\n" "add x24, x24, #0x10\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" @@ -1628,112 +1627,112 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "sub x27, x27, #0x8\n" ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q24, [x10, #0x30]\n" "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\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" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" - "ldr q6, [x10, #0x40]\n" + ".inst 0x4f42f332 // bfdot v18.4s, v25.8h, v2.h[0]\n" + ".inst 0x4f43f336 // bfdot v22.4s, v25.8h, v3.h[0]\n" + "ldr q25, [x10, #0x40]\n" "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" - ".inst 0x4f63f0d4 // bfdot v20.4s, v6.8h, v3.h[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" - ".inst 0x4f63f0f5 // bfdot v21.4s, v7.8h, v3.h[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" - ".inst 0x4f63f0d6 // bfdot v22.4s, v6.8h, v3.h[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" - ".inst 0x4f63f0f7 // bfdot v23.4s, v7.8h, v3.h[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" - ".inst 0x4f43f8d4 // bfdot v20.4s, v6.8h, v3.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" - ".inst 0x4f43f8f5 // bfdot v21.4s, v7.8h, v3.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" - ".inst 0x4f43f8d6 // bfdot v22.4s, v6.8h, v3.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" - ".inst 0x4f43f8f7 // bfdot v23.4s, v7.8h, v3.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" - ".inst 0x4f63f8d4 // bfdot v20.4s, v6.8h, v3.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" - ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" - ".inst 0x4f63f8f5 // bfdot v21.4s, v7.8h, v3.h[3]\n" - "ldr q7, [x10, #0xf0]\n" + ".inst 0x4f40f30b // bfdot v11.4s, v24.8h, v0.h[0]\n" + ".inst 0x4f41f30f // bfdot v15.4s, v24.8h, v1.h[0]\n" + ".inst 0x4f42f313 // bfdot v19.4s, v24.8h, v2.h[0]\n" + ".inst 0x4f43f317 // bfdot v23.4s, v24.8h, v3.h[0]\n" + "ldr q24, [x10, #0x50]\n" + ".inst 0x4f60f328 // bfdot v8.4s, v25.8h, v0.h[1]\n" + ".inst 0x4f61f32c // bfdot v12.4s, v25.8h, v1.h[1]\n" + ".inst 0x4f62f330 // bfdot v16.4s, v25.8h, v2.h[1]\n" + ".inst 0x4f63f334 // bfdot v20.4s, v25.8h, v3.h[1]\n" + "ldr q25, [x10, #0x60]\n" + ".inst 0x4f60f309 // bfdot v9.4s, v24.8h, v0.h[1]\n" + ".inst 0x4f61f30d // bfdot v13.4s, v24.8h, v1.h[1]\n" + ".inst 0x4f62f311 // bfdot v17.4s, v24.8h, v2.h[1]\n" + ".inst 0x4f63f315 // bfdot v21.4s, v24.8h, v3.h[1]\n" + "ldr q24, [x10, #0x70]\n" + ".inst 0x4f60f32a // bfdot v10.4s, v25.8h, v0.h[1]\n" + ".inst 0x4f61f32e // bfdot v14.4s, v25.8h, v1.h[1]\n" + ".inst 0x4f62f332 // bfdot v18.4s, v25.8h, v2.h[1]\n" + ".inst 0x4f63f336 // bfdot v22.4s, v25.8h, v3.h[1]\n" + "ldr q25, [x10, #0x80]\n" + ".inst 0x4f60f30b // bfdot v11.4s, v24.8h, v0.h[1]\n" + ".inst 0x4f61f30f // bfdot v15.4s, v24.8h, v1.h[1]\n" + ".inst 0x4f62f313 // bfdot v19.4s, v24.8h, v2.h[1]\n" + ".inst 0x4f63f317 // bfdot v23.4s, v24.8h, v3.h[1]\n" + "ldr q24, [x10, #0x90]\n" + ".inst 0x4f40fb28 // bfdot v8.4s, v25.8h, v0.h[2]\n" + ".inst 0x4f41fb2c // bfdot v12.4s, v25.8h, v1.h[2]\n" + ".inst 0x4f42fb30 // bfdot v16.4s, v25.8h, v2.h[2]\n" + ".inst 0x4f43fb34 // bfdot v20.4s, v25.8h, v3.h[2]\n" + "ldr q25, [x10, #0xa0]\n" + ".inst 0x4f40fb09 // bfdot v9.4s, v24.8h, v0.h[2]\n" + ".inst 0x4f41fb0d // bfdot v13.4s, v24.8h, v1.h[2]\n" + ".inst 0x4f42fb11 // bfdot v17.4s, v24.8h, v2.h[2]\n" + ".inst 0x4f43fb15 // bfdot v21.4s, v24.8h, v3.h[2]\n" + "ldr q24, [x10, #0xb0]\n" + ".inst 0x4f40fb2a // bfdot v10.4s, v25.8h, v0.h[2]\n" + ".inst 0x4f41fb2e // bfdot v14.4s, v25.8h, v1.h[2]\n" + ".inst 0x4f42fb32 // bfdot v18.4s, v25.8h, v2.h[2]\n" + ".inst 0x4f43fb36 // bfdot v22.4s, v25.8h, v3.h[2]\n" + "ldr q25, [x10, #0xc0]\n" + ".inst 0x4f40fb0b // bfdot v11.4s, v24.8h, v0.h[2]\n" + ".inst 0x4f41fb0f // bfdot v15.4s, v24.8h, v1.h[2]\n" + ".inst 0x4f42fb13 // bfdot v19.4s, v24.8h, v2.h[2]\n" + ".inst 0x4f43fb17 // bfdot v23.4s, v24.8h, v3.h[2]\n" + "ldr q24, [x10, #0xd0]\n" + ".inst 0x4f60fb28 // bfdot v8.4s, v25.8h, v0.h[3]\n" + ".inst 0x4f61fb2c // bfdot v12.4s, v25.8h, v1.h[3]\n" + ".inst 0x4f62fb30 // bfdot v16.4s, v25.8h, v2.h[3]\n" + ".inst 0x4f63fb34 // bfdot v20.4s, v25.8h, v3.h[3]\n" + "ldr q25, [x10, #0xe0]\n" + ".inst 0x4f60fb09 // bfdot v9.4s, v24.8h, v0.h[3]\n" + ".inst 0x4f61fb0d // bfdot v13.4s, v24.8h, v1.h[3]\n" + ".inst 0x4f62fb11 // bfdot v17.4s, v24.8h, v2.h[3]\n" + ".inst 0x4f63fb15 // bfdot v21.4s, v24.8h, v3.h[3]\n" + "ldr q24, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" - ".inst 0x4f63f8d6 // bfdot v22.4s, v6.8h, v3.h[3]\n" - ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" - ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" - ".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n" + ".inst 0x4f60fb2a // bfdot v10.4s, v25.8h, v0.h[3]\n" + ".inst 0x4f61fb2e // bfdot v14.4s, v25.8h, v1.h[3]\n" + ".inst 0x4f62fb32 // bfdot v18.4s, v25.8h, v2.h[3]\n" + ".inst 0x4f63fb36 // bfdot v22.4s, v25.8h, v3.h[3]\n" + ".inst 0x4f60fb0b // bfdot v11.4s, v24.8h, v0.h[3]\n" + ".inst 0x4f61fb0f // bfdot v15.4s, v24.8h, v1.h[3]\n" + ".inst 0x4f62fb13 // bfdot v19.4s, v24.8h, v2.h[3]\n" + ".inst 0x4f63fb17 // bfdot v23.4s, v24.8h, v3.h[3]\n" "125:" // Height 4: Multiply loop: Main loop skip "cbz x27, 129f\n" "cmp x27, #0x2\n" "blt 127f\n" "126:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr s1, [x25], #0x4\n" + "ldr s29, [x26], #0x4\n" + "ldr s28, [x25], #0x4\n" "sub x27, x27, #0x2\n" "cmp x27, #0x2\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr s27, [x24], #0x4\n" + "ldr s26, [x23], #0x4\n" + "ldr q25, [x10, #0x0]\n" + "ldr q24, [x10, #0x10]\n" + ".inst 0x4f5df328 // bfdot v8.4s, v25.8h, v29.h[0]\n" + ".inst 0x4f5cf32c // bfdot v12.4s, v25.8h, v28.h[0]\n" + ".inst 0x4f5bf330 // bfdot v16.4s, v25.8h, v27.h[0]\n" + ".inst 0x4f5af334 // bfdot v20.4s, v25.8h, v26.h[0]\n" + "ldr q25, [x10, #0x20]\n" + ".inst 0x4f5df309 // bfdot v9.4s, v24.8h, v29.h[0]\n" + ".inst 0x4f5cf30d // bfdot v13.4s, v24.8h, v28.h[0]\n" + ".inst 0x4f5bf311 // bfdot v17.4s, v24.8h, v27.h[0]\n" + ".inst 0x4f5af315 // bfdot v21.4s, v24.8h, v26.h[0]\n" + "ldr q24, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" + ".inst 0x4f5df32a // bfdot v10.4s, v25.8h, v29.h[0]\n" + ".inst 0x4f5cf32e // bfdot v14.4s, v25.8h, v28.h[0]\n" + ".inst 0x4f5bf332 // bfdot v18.4s, v25.8h, v27.h[0]\n" + ".inst 0x4f5af336 // bfdot v22.4s, v25.8h, v26.h[0]\n" + ".inst 0x4f5df30b // bfdot v11.4s, v24.8h, v29.h[0]\n" + ".inst 0x4f5cf30f // bfdot v15.4s, v24.8h, v28.h[0]\n" + ".inst 0x4f5bf313 // bfdot v19.4s, v24.8h, v27.h[0]\n" + ".inst 0x4f5af317 // bfdot v23.4s, v24.8h, v26.h[0]\n" "bge 126b\n" "127:" // Height 4: Multiply loop: Skip odd blocks "cbz x27, 129f\n" @@ -1742,27 +1741,27 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr h2, [x24, #0x0]\n" "ldr h3, [x23, #0x0]\n" "128:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q25, [x10, #0x0]\n" + "ldr q24, [x10, #0x10]\n" + ".inst 0x4f40f328 // bfdot v8.4s, v25.8h, v0.h[0]\n" + ".inst 0x4f41f32c // bfdot v12.4s, v25.8h, v1.h[0]\n" + ".inst 0x4f42f330 // bfdot v16.4s, v25.8h, v2.h[0]\n" + ".inst 0x4f43f334 // bfdot v20.4s, v25.8h, v3.h[0]\n" + "ldr q25, [x10, #0x20]\n" + ".inst 0x4f40f309 // bfdot v9.4s, v24.8h, v0.h[0]\n" + ".inst 0x4f41f30d // bfdot v13.4s, v24.8h, v1.h[0]\n" + ".inst 0x4f42f311 // bfdot v17.4s, v24.8h, v2.h[0]\n" + ".inst 0x4f43f315 // bfdot v21.4s, v24.8h, v3.h[0]\n" + "ldr q24, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" + ".inst 0x4f40f32a // bfdot v10.4s, v25.8h, v0.h[0]\n" + ".inst 0x4f41f32e // bfdot v14.4s, v25.8h, v1.h[0]\n" + ".inst 0x4f42f332 // bfdot v18.4s, v25.8h, v2.h[0]\n" + ".inst 0x4f43f336 // bfdot v22.4s, v25.8h, v3.h[0]\n" + ".inst 0x4f40f30b // bfdot v11.4s, v24.8h, v0.h[0]\n" + ".inst 0x4f41f30f // bfdot v15.4s, v24.8h, v1.h[0]\n" + ".inst 0x4f42f313 // bfdot v19.4s, v24.8h, v2.h[0]\n" + ".inst 0x4f43f317 // bfdot v23.4s, v24.8h, v3.h[0]\n" "129:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -1778,41 +1777,41 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "prfm pstl1keep, [x23, #0x0]\n" "tbz %x[flags], #1, 130f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v25.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmin v16.4s, v16.4s, v1.4s\n" - "fmin v17.4s, v17.4s, v1.4s\n" - "fmin v18.4s, v18.4s, v1.4s\n" - "fmin v19.4s, v19.4s, v1.4s\n" - "fmin v20.4s, v20.4s, v1.4s\n" - "fmin v21.4s, v21.4s, v1.4s\n" - "fmin v22.4s, v22.4s, v1.4s\n" - "fmin v23.4s, v23.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" - "fmax v16.4s, v16.4s, v0.4s\n" - "fmax v17.4s, v17.4s, v0.4s\n" - "fmax v18.4s, v18.4s, v0.4s\n" - "fmax v19.4s, v19.4s, v0.4s\n" - "fmax v20.4s, v20.4s, v0.4s\n" - "fmax v21.4s, v21.4s, v0.4s\n" - "fmax v22.4s, v22.4s, v0.4s\n" - "fmax v23.4s, v23.4s, v0.4s\n" + "ld1r { v24.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v25.4s\n" + "fmin v9.4s, v9.4s, v25.4s\n" + "fmin v10.4s, v10.4s, v25.4s\n" + "fmin v11.4s, v11.4s, v25.4s\n" + "fmin v12.4s, v12.4s, v25.4s\n" + "fmin v13.4s, v13.4s, v25.4s\n" + "fmin v14.4s, v14.4s, v25.4s\n" + "fmin v15.4s, v15.4s, v25.4s\n" + "fmin v16.4s, v16.4s, v25.4s\n" + "fmin v17.4s, v17.4s, v25.4s\n" + "fmin v18.4s, v18.4s, v25.4s\n" + "fmin v19.4s, v19.4s, v25.4s\n" + "fmin v20.4s, v20.4s, v25.4s\n" + "fmin v21.4s, v21.4s, v25.4s\n" + "fmin v22.4s, v22.4s, v25.4s\n" + "fmin v23.4s, v23.4s, v25.4s\n" + "fmax v8.4s, v8.4s, v24.4s\n" + "fmax v9.4s, v9.4s, v24.4s\n" + "fmax v10.4s, v10.4s, v24.4s\n" + "fmax v11.4s, v11.4s, v24.4s\n" + "fmax v12.4s, v12.4s, v24.4s\n" + "fmax v13.4s, v13.4s, v24.4s\n" + "fmax v14.4s, v14.4s, v24.4s\n" + "fmax v15.4s, v15.4s, v24.4s\n" + "fmax v16.4s, v16.4s, v24.4s\n" + "fmax v17.4s, v17.4s, v24.4s\n" + "fmax v18.4s, v18.4s, v24.4s\n" + "fmax v19.4s, v19.4s, v24.4s\n" + "fmax v20.4s, v20.4s, v24.4s\n" + "fmax v21.4s, v21.4s, v24.4s\n" + "fmax v22.4s, v22.4s, v24.4s\n" + "fmax v23.4s, v23.4s, v24.4s\n" "130:" // Height 4: No activation "cmp x11, #0x10\n" "bge 139f\n" @@ -2137,15 +2136,15 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "155:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 156f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" "cbnz x28, 157f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -2156,10 +2155,10 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "b 157f\n" "156:" // Height 5: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" - "add x22, x23, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" + "add x22, x23, x21, LSL #1\n" "157:" // Height 5: input setup done "cmp x27, #0x8\n" "blt 160f\n" @@ -2182,7 +2181,7 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "add x25, x25, #0x10\n" "add x24, x24, #0x10\n" ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q29, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" "add x23, x23, #0x10\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" @@ -2191,100 +2190,100 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "cmp x27, #0x10\n" ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q28, [x10, #0x30]\n" "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\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, [x25, #0x80]\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" + ".inst 0x4f42f3b2 // bfdot v18.4s, v29.8h, v2.h[0]\n" + ".inst 0x4f43f3b6 // bfdot v22.4s, v29.8h, v3.h[0]\n" "prfm pldl1keep, [x23, #0x80]\n" "prfm pldl1keep, [x22, #0x80]\n" - ".inst 0x4f44f0da // bfdot v26.4s, v6.8h, v4.h[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" - ".inst 0x4f63f0d4 // bfdot v20.4s, v6.8h, v3.h[1]\n" - ".inst 0x4f64f0d8 // bfdot v24.4s, v6.8h, v4.h[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" - ".inst 0x4f63f0f5 // bfdot v21.4s, v7.8h, v3.h[1]\n" - ".inst 0x4f64f0f9 // bfdot v25.4s, v7.8h, v4.h[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" - ".inst 0x4f63f0d6 // bfdot v22.4s, v6.8h, v3.h[1]\n" - ".inst 0x4f64f0da // bfdot v26.4s, v6.8h, v4.h[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" - ".inst 0x4f63f0f7 // bfdot v23.4s, v7.8h, v3.h[1]\n" - ".inst 0x4f64f0fb // bfdot v27.4s, v7.8h, v4.h[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" - ".inst 0x4f43f8d4 // bfdot v20.4s, v6.8h, v3.h[2]\n" - ".inst 0x4f44f8d8 // bfdot v24.4s, v6.8h, v4.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" - ".inst 0x4f43f8f5 // bfdot v21.4s, v7.8h, v3.h[2]\n" - ".inst 0x4f44f8f9 // bfdot v25.4s, v7.8h, v4.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" - ".inst 0x4f43f8d6 // bfdot v22.4s, v6.8h, v3.h[2]\n" - ".inst 0x4f44f8da // bfdot v26.4s, v6.8h, v4.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" - ".inst 0x4f43f8f7 // bfdot v23.4s, v7.8h, v3.h[2]\n" - ".inst 0x4f44f8fb // bfdot v27.4s, v7.8h, v4.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" - ".inst 0x4f63f8d4 // bfdot v20.4s, v6.8h, v3.h[3]\n" - ".inst 0x4f64f8d8 // bfdot v24.4s, v6.8h, v4.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" - ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" - ".inst 0x4f63f8f5 // bfdot v21.4s, v7.8h, v3.h[3]\n" - ".inst 0x4f64f8f9 // bfdot v25.4s, v7.8h, v4.h[3]\n" - "ldr q7, [x10, #0xf0]\n" + ".inst 0x4f44f3ba // bfdot v26.4s, v29.8h, v4.h[0]\n" + "ldr q29, [x10, #0x40]\n" + ".inst 0x4f40f38b // bfdot v11.4s, v28.8h, v0.h[0]\n" + ".inst 0x4f41f38f // bfdot v15.4s, v28.8h, v1.h[0]\n" + ".inst 0x4f42f393 // bfdot v19.4s, v28.8h, v2.h[0]\n" + ".inst 0x4f43f397 // bfdot v23.4s, v28.8h, v3.h[0]\n" + ".inst 0x4f44f39b // bfdot v27.4s, v28.8h, v4.h[0]\n" + "ldr q28, [x10, #0x50]\n" + ".inst 0x4f60f3a8 // bfdot v8.4s, v29.8h, v0.h[1]\n" + ".inst 0x4f61f3ac // bfdot v12.4s, v29.8h, v1.h[1]\n" + ".inst 0x4f62f3b0 // bfdot v16.4s, v29.8h, v2.h[1]\n" + ".inst 0x4f63f3b4 // bfdot v20.4s, v29.8h, v3.h[1]\n" + ".inst 0x4f64f3b8 // bfdot v24.4s, v29.8h, v4.h[1]\n" + "ldr q29, [x10, #0x60]\n" + ".inst 0x4f60f389 // bfdot v9.4s, v28.8h, v0.h[1]\n" + ".inst 0x4f61f38d // bfdot v13.4s, v28.8h, v1.h[1]\n" + ".inst 0x4f62f391 // bfdot v17.4s, v28.8h, v2.h[1]\n" + ".inst 0x4f63f395 // bfdot v21.4s, v28.8h, v3.h[1]\n" + ".inst 0x4f64f399 // bfdot v25.4s, v28.8h, v4.h[1]\n" + "ldr q28, [x10, #0x70]\n" + ".inst 0x4f60f3aa // bfdot v10.4s, v29.8h, v0.h[1]\n" + ".inst 0x4f61f3ae // bfdot v14.4s, v29.8h, v1.h[1]\n" + ".inst 0x4f62f3b2 // bfdot v18.4s, v29.8h, v2.h[1]\n" + ".inst 0x4f63f3b6 // bfdot v22.4s, v29.8h, v3.h[1]\n" + ".inst 0x4f64f3ba // bfdot v26.4s, v29.8h, v4.h[1]\n" + "ldr q29, [x10, #0x80]\n" + ".inst 0x4f60f38b // bfdot v11.4s, v28.8h, v0.h[1]\n" + ".inst 0x4f61f38f // bfdot v15.4s, v28.8h, v1.h[1]\n" + ".inst 0x4f62f393 // bfdot v19.4s, v28.8h, v2.h[1]\n" + ".inst 0x4f63f397 // bfdot v23.4s, v28.8h, v3.h[1]\n" + ".inst 0x4f64f39b // bfdot v27.4s, v28.8h, v4.h[1]\n" + "ldr q28, [x10, #0x90]\n" + ".inst 0x4f40fba8 // bfdot v8.4s, v29.8h, v0.h[2]\n" + ".inst 0x4f41fbac // bfdot v12.4s, v29.8h, v1.h[2]\n" + ".inst 0x4f42fbb0 // bfdot v16.4s, v29.8h, v2.h[2]\n" + ".inst 0x4f43fbb4 // bfdot v20.4s, v29.8h, v3.h[2]\n" + ".inst 0x4f44fbb8 // bfdot v24.4s, v29.8h, v4.h[2]\n" + "ldr q29, [x10, #0xa0]\n" + ".inst 0x4f40fb89 // bfdot v9.4s, v28.8h, v0.h[2]\n" + ".inst 0x4f41fb8d // bfdot v13.4s, v28.8h, v1.h[2]\n" + ".inst 0x4f42fb91 // bfdot v17.4s, v28.8h, v2.h[2]\n" + ".inst 0x4f43fb95 // bfdot v21.4s, v28.8h, v3.h[2]\n" + ".inst 0x4f44fb99 // bfdot v25.4s, v28.8h, v4.h[2]\n" + "ldr q28, [x10, #0xb0]\n" + ".inst 0x4f40fbaa // bfdot v10.4s, v29.8h, v0.h[2]\n" + ".inst 0x4f41fbae // bfdot v14.4s, v29.8h, v1.h[2]\n" + ".inst 0x4f42fbb2 // bfdot v18.4s, v29.8h, v2.h[2]\n" + ".inst 0x4f43fbb6 // bfdot v22.4s, v29.8h, v3.h[2]\n" + ".inst 0x4f44fbba // bfdot v26.4s, v29.8h, v4.h[2]\n" + "ldr q29, [x10, #0xc0]\n" + ".inst 0x4f40fb8b // bfdot v11.4s, v28.8h, v0.h[2]\n" + ".inst 0x4f41fb8f // bfdot v15.4s, v28.8h, v1.h[2]\n" + ".inst 0x4f42fb93 // bfdot v19.4s, v28.8h, v2.h[2]\n" + ".inst 0x4f43fb97 // bfdot v23.4s, v28.8h, v3.h[2]\n" + ".inst 0x4f44fb9b // bfdot v27.4s, v28.8h, v4.h[2]\n" + "ldr q28, [x10, #0xd0]\n" + ".inst 0x4f60fba8 // bfdot v8.4s, v29.8h, v0.h[3]\n" + ".inst 0x4f61fbac // bfdot v12.4s, v29.8h, v1.h[3]\n" + ".inst 0x4f62fbb0 // bfdot v16.4s, v29.8h, v2.h[3]\n" + ".inst 0x4f63fbb4 // bfdot v20.4s, v29.8h, v3.h[3]\n" + ".inst 0x4f64fbb8 // bfdot v24.4s, v29.8h, v4.h[3]\n" + "ldr q29, [x10, #0xe0]\n" + ".inst 0x4f60fb89 // bfdot v9.4s, v28.8h, v0.h[3]\n" + ".inst 0x4f61fb8d // bfdot v13.4s, v28.8h, v1.h[3]\n" + ".inst 0x4f62fb91 // bfdot v17.4s, v28.8h, v2.h[3]\n" + ".inst 0x4f63fb95 // bfdot v21.4s, v28.8h, v3.h[3]\n" + ".inst 0x4f64fb99 // bfdot v25.4s, v28.8h, v4.h[3]\n" + "ldr q28, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" - ".inst 0x4f63f8d6 // bfdot v22.4s, v6.8h, v3.h[3]\n" - ".inst 0x4f64f8da // bfdot v26.4s, v6.8h, v4.h[3]\n" + ".inst 0x4f60fbaa // bfdot v10.4s, v29.8h, v0.h[3]\n" + ".inst 0x4f61fbae // bfdot v14.4s, v29.8h, v1.h[3]\n" + ".inst 0x4f62fbb2 // bfdot v18.4s, v29.8h, v2.h[3]\n" + ".inst 0x4f63fbb6 // bfdot v22.4s, v29.8h, v3.h[3]\n" + ".inst 0x4f64fbba // bfdot v26.4s, v29.8h, v4.h[3]\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" + ".inst 0x4f60fb8b // bfdot v11.4s, v28.8h, v0.h[3]\n" "ldr q0, [x26, #0x0]\n" - ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" + ".inst 0x4f61fb8f // bfdot v15.4s, v28.8h, v1.h[3]\n" "ldr q1, [x25, #0x0]\n" - ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" + ".inst 0x4f62fb93 // bfdot v19.4s, v28.8h, v2.h[3]\n" "ldr q2, [x24, #0x0]\n" - ".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n" + ".inst 0x4f63fb97 // bfdot v23.4s, v28.8h, v3.h[3]\n" "ldr q3, [x23, #0x0]\n" - ".inst 0x4f64f8fb // bfdot v27.4s, v7.8h, v4.h[3]\n" + ".inst 0x4f64fb9b // bfdot v27.4s, v28.8h, v4.h[3]\n" "ldr q4, [x22, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 158b\n" @@ -2298,7 +2297,7 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q29, [x10, #0x20]\n" ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" "add x22, x22, #0x10\n" ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" @@ -2307,131 +2306,131 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "prfm pldl1keep, [x26, #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 q7, [x10, #0x30]\n" + "ldr q28, [x10, #0x30]\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\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" "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" + ".inst 0x4f42f3b2 // bfdot v18.4s, v29.8h, v2.h[0]\n" + ".inst 0x4f43f3b6 // bfdot v22.4s, v29.8h, v3.h[0]\n" "prfm pldl1keep, [x22, #0x80]\n" - ".inst 0x4f44f0da // bfdot v26.4s, v6.8h, v4.h[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x4f60f0c8 // bfdot v8.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0cc // bfdot v12.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d0 // bfdot v16.4s, v6.8h, v2.h[1]\n" - ".inst 0x4f63f0d4 // bfdot v20.4s, v6.8h, v3.h[1]\n" - ".inst 0x4f64f0d8 // bfdot v24.4s, v6.8h, v4.h[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x4f60f0e9 // bfdot v9.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ed // bfdot v13.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f1 // bfdot v17.4s, v7.8h, v2.h[1]\n" - ".inst 0x4f63f0f5 // bfdot v21.4s, v7.8h, v3.h[1]\n" - ".inst 0x4f64f0f9 // bfdot v25.4s, v7.8h, v4.h[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x4f60f0ca // bfdot v10.4s, v6.8h, v0.h[1]\n" - ".inst 0x4f61f0ce // bfdot v14.4s, v6.8h, v1.h[1]\n" - ".inst 0x4f62f0d2 // bfdot v18.4s, v6.8h, v2.h[1]\n" - ".inst 0x4f63f0d6 // bfdot v22.4s, v6.8h, v3.h[1]\n" - ".inst 0x4f64f0da // bfdot v26.4s, v6.8h, v4.h[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x4f60f0eb // bfdot v11.4s, v7.8h, v0.h[1]\n" - ".inst 0x4f61f0ef // bfdot v15.4s, v7.8h, v1.h[1]\n" - ".inst 0x4f62f0f3 // bfdot v19.4s, v7.8h, v2.h[1]\n" - ".inst 0x4f63f0f7 // bfdot v23.4s, v7.8h, v3.h[1]\n" - ".inst 0x4f64f0fb // bfdot v27.4s, v7.8h, v4.h[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x4f40f8c8 // bfdot v8.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8cc // bfdot v12.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d0 // bfdot v16.4s, v6.8h, v2.h[2]\n" - ".inst 0x4f43f8d4 // bfdot v20.4s, v6.8h, v3.h[2]\n" - ".inst 0x4f44f8d8 // bfdot v24.4s, v6.8h, v4.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x4f40f8e9 // bfdot v9.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ed // bfdot v13.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f1 // bfdot v17.4s, v7.8h, v2.h[2]\n" - ".inst 0x4f43f8f5 // bfdot v21.4s, v7.8h, v3.h[2]\n" - ".inst 0x4f44f8f9 // bfdot v25.4s, v7.8h, v4.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x4f40f8ca // bfdot v10.4s, v6.8h, v0.h[2]\n" - ".inst 0x4f41f8ce // bfdot v14.4s, v6.8h, v1.h[2]\n" - ".inst 0x4f42f8d2 // bfdot v18.4s, v6.8h, v2.h[2]\n" - ".inst 0x4f43f8d6 // bfdot v22.4s, v6.8h, v3.h[2]\n" - ".inst 0x4f44f8da // bfdot v26.4s, v6.8h, v4.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x4f40f8eb // bfdot v11.4s, v7.8h, v0.h[2]\n" - ".inst 0x4f41f8ef // bfdot v15.4s, v7.8h, v1.h[2]\n" - ".inst 0x4f42f8f3 // bfdot v19.4s, v7.8h, v2.h[2]\n" - ".inst 0x4f43f8f7 // bfdot v23.4s, v7.8h, v3.h[2]\n" - ".inst 0x4f44f8fb // bfdot v27.4s, v7.8h, v4.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x4f60f8c8 // bfdot v8.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8cc // bfdot v12.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d0 // bfdot v16.4s, v6.8h, v2.h[3]\n" - ".inst 0x4f63f8d4 // bfdot v20.4s, v6.8h, v3.h[3]\n" - ".inst 0x4f64f8d8 // bfdot v24.4s, v6.8h, v4.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x4f60f8e9 // bfdot v9.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ed // bfdot v13.4s, v7.8h, v1.h[3]\n" - ".inst 0x4f62f8f1 // bfdot v17.4s, v7.8h, v2.h[3]\n" - ".inst 0x4f63f8f5 // bfdot v21.4s, v7.8h, v3.h[3]\n" - ".inst 0x4f64f8f9 // bfdot v25.4s, v7.8h, v4.h[3]\n" - "ldr q7, [x10, #0xf0]\n" + ".inst 0x4f44f3ba // bfdot v26.4s, v29.8h, v4.h[0]\n" + "ldr q29, [x10, #0x40]\n" + ".inst 0x4f40f38b // bfdot v11.4s, v28.8h, v0.h[0]\n" + ".inst 0x4f41f38f // bfdot v15.4s, v28.8h, v1.h[0]\n" + ".inst 0x4f42f393 // bfdot v19.4s, v28.8h, v2.h[0]\n" + ".inst 0x4f43f397 // bfdot v23.4s, v28.8h, v3.h[0]\n" + ".inst 0x4f44f39b // bfdot v27.4s, v28.8h, v4.h[0]\n" + "ldr q28, [x10, #0x50]\n" + ".inst 0x4f60f3a8 // bfdot v8.4s, v29.8h, v0.h[1]\n" + ".inst 0x4f61f3ac // bfdot v12.4s, v29.8h, v1.h[1]\n" + ".inst 0x4f62f3b0 // bfdot v16.4s, v29.8h, v2.h[1]\n" + ".inst 0x4f63f3b4 // bfdot v20.4s, v29.8h, v3.h[1]\n" + ".inst 0x4f64f3b8 // bfdot v24.4s, v29.8h, v4.h[1]\n" + "ldr q29, [x10, #0x60]\n" + ".inst 0x4f60f389 // bfdot v9.4s, v28.8h, v0.h[1]\n" + ".inst 0x4f61f38d // bfdot v13.4s, v28.8h, v1.h[1]\n" + ".inst 0x4f62f391 // bfdot v17.4s, v28.8h, v2.h[1]\n" + ".inst 0x4f63f395 // bfdot v21.4s, v28.8h, v3.h[1]\n" + ".inst 0x4f64f399 // bfdot v25.4s, v28.8h, v4.h[1]\n" + "ldr q28, [x10, #0x70]\n" + ".inst 0x4f60f3aa // bfdot v10.4s, v29.8h, v0.h[1]\n" + ".inst 0x4f61f3ae // bfdot v14.4s, v29.8h, v1.h[1]\n" + ".inst 0x4f62f3b2 // bfdot v18.4s, v29.8h, v2.h[1]\n" + ".inst 0x4f63f3b6 // bfdot v22.4s, v29.8h, v3.h[1]\n" + ".inst 0x4f64f3ba // bfdot v26.4s, v29.8h, v4.h[1]\n" + "ldr q29, [x10, #0x80]\n" + ".inst 0x4f60f38b // bfdot v11.4s, v28.8h, v0.h[1]\n" + ".inst 0x4f61f38f // bfdot v15.4s, v28.8h, v1.h[1]\n" + ".inst 0x4f62f393 // bfdot v19.4s, v28.8h, v2.h[1]\n" + ".inst 0x4f63f397 // bfdot v23.4s, v28.8h, v3.h[1]\n" + ".inst 0x4f64f39b // bfdot v27.4s, v28.8h, v4.h[1]\n" + "ldr q28, [x10, #0x90]\n" + ".inst 0x4f40fba8 // bfdot v8.4s, v29.8h, v0.h[2]\n" + ".inst 0x4f41fbac // bfdot v12.4s, v29.8h, v1.h[2]\n" + ".inst 0x4f42fbb0 // bfdot v16.4s, v29.8h, v2.h[2]\n" + ".inst 0x4f43fbb4 // bfdot v20.4s, v29.8h, v3.h[2]\n" + ".inst 0x4f44fbb8 // bfdot v24.4s, v29.8h, v4.h[2]\n" + "ldr q29, [x10, #0xa0]\n" + ".inst 0x4f40fb89 // bfdot v9.4s, v28.8h, v0.h[2]\n" + ".inst 0x4f41fb8d // bfdot v13.4s, v28.8h, v1.h[2]\n" + ".inst 0x4f42fb91 // bfdot v17.4s, v28.8h, v2.h[2]\n" + ".inst 0x4f43fb95 // bfdot v21.4s, v28.8h, v3.h[2]\n" + ".inst 0x4f44fb99 // bfdot v25.4s, v28.8h, v4.h[2]\n" + "ldr q28, [x10, #0xb0]\n" + ".inst 0x4f40fbaa // bfdot v10.4s, v29.8h, v0.h[2]\n" + ".inst 0x4f41fbae // bfdot v14.4s, v29.8h, v1.h[2]\n" + ".inst 0x4f42fbb2 // bfdot v18.4s, v29.8h, v2.h[2]\n" + ".inst 0x4f43fbb6 // bfdot v22.4s, v29.8h, v3.h[2]\n" + ".inst 0x4f44fbba // bfdot v26.4s, v29.8h, v4.h[2]\n" + "ldr q29, [x10, #0xc0]\n" + ".inst 0x4f40fb8b // bfdot v11.4s, v28.8h, v0.h[2]\n" + ".inst 0x4f41fb8f // bfdot v15.4s, v28.8h, v1.h[2]\n" + ".inst 0x4f42fb93 // bfdot v19.4s, v28.8h, v2.h[2]\n" + ".inst 0x4f43fb97 // bfdot v23.4s, v28.8h, v3.h[2]\n" + ".inst 0x4f44fb9b // bfdot v27.4s, v28.8h, v4.h[2]\n" + "ldr q28, [x10, #0xd0]\n" + ".inst 0x4f60fba8 // bfdot v8.4s, v29.8h, v0.h[3]\n" + ".inst 0x4f61fbac // bfdot v12.4s, v29.8h, v1.h[3]\n" + ".inst 0x4f62fbb0 // bfdot v16.4s, v29.8h, v2.h[3]\n" + ".inst 0x4f63fbb4 // bfdot v20.4s, v29.8h, v3.h[3]\n" + ".inst 0x4f64fbb8 // bfdot v24.4s, v29.8h, v4.h[3]\n" + "ldr q29, [x10, #0xe0]\n" + ".inst 0x4f60fb89 // bfdot v9.4s, v28.8h, v0.h[3]\n" + ".inst 0x4f61fb8d // bfdot v13.4s, v28.8h, v1.h[3]\n" + ".inst 0x4f62fb91 // bfdot v17.4s, v28.8h, v2.h[3]\n" + ".inst 0x4f63fb95 // bfdot v21.4s, v28.8h, v3.h[3]\n" + ".inst 0x4f64fb99 // bfdot v25.4s, v28.8h, v4.h[3]\n" + "ldr q28, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x4f60f8ca // bfdot v10.4s, v6.8h, v0.h[3]\n" - ".inst 0x4f61f8ce // bfdot v14.4s, v6.8h, v1.h[3]\n" - ".inst 0x4f62f8d2 // bfdot v18.4s, v6.8h, v2.h[3]\n" - ".inst 0x4f63f8d6 // bfdot v22.4s, v6.8h, v3.h[3]\n" - ".inst 0x4f64f8da // bfdot v26.4s, v6.8h, v4.h[3]\n" - ".inst 0x4f60f8eb // bfdot v11.4s, v7.8h, v0.h[3]\n" - ".inst 0x4f61f8ef // bfdot v15.4s, v7.8h, v1.h[3]\n" - ".inst 0x4f62f8f3 // bfdot v19.4s, v7.8h, v2.h[3]\n" - ".inst 0x4f63f8f7 // bfdot v23.4s, v7.8h, v3.h[3]\n" - ".inst 0x4f64f8fb // bfdot v27.4s, v7.8h, v4.h[3]\n" + ".inst 0x4f60fbaa // bfdot v10.4s, v29.8h, v0.h[3]\n" + ".inst 0x4f61fbae // bfdot v14.4s, v29.8h, v1.h[3]\n" + ".inst 0x4f62fbb2 // bfdot v18.4s, v29.8h, v2.h[3]\n" + ".inst 0x4f63fbb6 // bfdot v22.4s, v29.8h, v3.h[3]\n" + ".inst 0x4f64fbba // bfdot v26.4s, v29.8h, v4.h[3]\n" + ".inst 0x4f60fb8b // bfdot v11.4s, v28.8h, v0.h[3]\n" + ".inst 0x4f61fb8f // bfdot v15.4s, v28.8h, v1.h[3]\n" + ".inst 0x4f62fb93 // bfdot v19.4s, v28.8h, v2.h[3]\n" + ".inst 0x4f63fb97 // bfdot v23.4s, v28.8h, v3.h[3]\n" + ".inst 0x4f64fb9b // bfdot v27.4s, v28.8h, v4.h[3]\n" "160:" // Height 5: Multiply loop: Main loop skip "cbz x27, 164f\n" "cmp x27, #0x2\n" "blt 162f\n" "161:" // Height 5: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" + "ldr s2, [x26], #0x4\n" "ldr s1, [x25], #0x4\n" "sub x27, x27, #0x2\n" "cmp x27, #0x2\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr s4, [x22], #0x4\n" - "ldr q6, [x10, #0x0]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr s0, [x24], #0x4\n" + "ldr s31, [x23], #0x4\n" + "ldr s30, [x22], #0x4\n" + "ldr q29, [x10, #0x0]\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" + "ldr q29, [x10, #0x20]\n" + ".inst 0x4f42f389 // bfdot v9.4s, v28.8h, v2.h[0]\n" + ".inst 0x4f41f38d // bfdot v13.4s, v28.8h, v1.h[0]\n" + ".inst 0x4f40f391 // bfdot v17.4s, v28.8h, v0.h[0]\n" + ".inst 0x4f5ff395 // bfdot v21.4s, v28.8h, v31.h[0]\n" + ".inst 0x4f5ef399 // bfdot v25.4s, v28.8h, v30.h[0]\n" + "ldr q28, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" - ".inst 0x4f44f0da // bfdot v26.4s, v6.8h, v4.h[0]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" + ".inst 0x4f42f3aa // bfdot v10.4s, v29.8h, v2.h[0]\n" + ".inst 0x4f41f3ae // bfdot v14.4s, v29.8h, v1.h[0]\n" + ".inst 0x4f40f3b2 // bfdot v18.4s, v29.8h, v0.h[0]\n" + ".inst 0x4f5ff3b6 // bfdot v22.4s, v29.8h, v31.h[0]\n" + ".inst 0x4f5ef3ba // bfdot v26.4s, v29.8h, v30.h[0]\n" + ".inst 0x4f42f38b // bfdot v11.4s, v28.8h, v2.h[0]\n" + ".inst 0x4f41f38f // bfdot v15.4s, v28.8h, v1.h[0]\n" + ".inst 0x4f40f393 // bfdot v19.4s, v28.8h, v0.h[0]\n" + ".inst 0x4f5ff397 // bfdot v23.4s, v28.8h, v31.h[0]\n" + ".inst 0x4f5ef39b // bfdot v27.4s, v28.8h, v30.h[0]\n" "bge 161b\n" "162:" // Height 5: Multiply loop: Skip odd blocks "cbz x27, 164f\n" @@ -2441,31 +2440,31 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr h3, [x23, #0x0]\n" "ldr h4, [x22, #0x0]\n" "163:" // Height 5: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\n" - ".inst 0x4f44f0f9 // bfdot v25.4s, v7.8h, v4.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q29, [x10, #0x0]\n" + "ldr q28, [x10, #0x10]\n" + ".inst 0x4f40f3a8 // bfdot v8.4s, v29.8h, v0.h[0]\n" + ".inst 0x4f41f3ac // bfdot v12.4s, v29.8h, v1.h[0]\n" + ".inst 0x4f42f3b0 // bfdot v16.4s, v29.8h, v2.h[0]\n" + ".inst 0x4f43f3b4 // bfdot v20.4s, v29.8h, v3.h[0]\n" + ".inst 0x4f44f3b8 // bfdot v24.4s, v29.8h, v4.h[0]\n" + "ldr q29, [x10, #0x20]\n" + ".inst 0x4f40f389 // bfdot v9.4s, v28.8h, v0.h[0]\n" + ".inst 0x4f41f38d // bfdot v13.4s, v28.8h, v1.h[0]\n" + ".inst 0x4f42f391 // bfdot v17.4s, v28.8h, v2.h[0]\n" + ".inst 0x4f43f395 // bfdot v21.4s, v28.8h, v3.h[0]\n" + ".inst 0x4f44f399 // bfdot v25.4s, v28.8h, v4.h[0]\n" + "ldr q28, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" - ".inst 0x4f44f0da // bfdot v26.4s, v6.8h, v4.h[0]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" + ".inst 0x4f40f3aa // bfdot v10.4s, v29.8h, v0.h[0]\n" + ".inst 0x4f41f3ae // bfdot v14.4s, v29.8h, v1.h[0]\n" + ".inst 0x4f42f3b2 // bfdot v18.4s, v29.8h, v2.h[0]\n" + ".inst 0x4f43f3b6 // bfdot v22.4s, v29.8h, v3.h[0]\n" + ".inst 0x4f44f3ba // bfdot v26.4s, v29.8h, v4.h[0]\n" + ".inst 0x4f40f38b // bfdot v11.4s, v28.8h, v0.h[0]\n" + ".inst 0x4f41f38f // bfdot v15.4s, v28.8h, v1.h[0]\n" + ".inst 0x4f42f393 // bfdot v19.4s, v28.8h, v2.h[0]\n" + ".inst 0x4f43f397 // bfdot v23.4s, v28.8h, v3.h[0]\n" + ".inst 0x4f44f39b // bfdot v27.4s, v28.8h, v4.h[0]\n" "164:" // Height 5: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -2483,49 +2482,49 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "prfm pstl1keep, [x22, #0x0]\n" "tbz %x[flags], #1, 165f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v29.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmin v16.4s, v16.4s, v1.4s\n" - "fmin v17.4s, v17.4s, v1.4s\n" - "fmin v18.4s, v18.4s, v1.4s\n" - "fmin v19.4s, v19.4s, v1.4s\n" - "fmin v20.4s, v20.4s, v1.4s\n" - "fmin v21.4s, v21.4s, v1.4s\n" - "fmin v22.4s, v22.4s, v1.4s\n" - "fmin v23.4s, v23.4s, v1.4s\n" - "fmin v24.4s, v24.4s, v1.4s\n" - "fmin v25.4s, v25.4s, v1.4s\n" - "fmin v26.4s, v26.4s, v1.4s\n" - "fmin v27.4s, v27.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" - "fmax v16.4s, v16.4s, v0.4s\n" - "fmax v17.4s, v17.4s, v0.4s\n" - "fmax v18.4s, v18.4s, v0.4s\n" - "fmax v19.4s, v19.4s, v0.4s\n" - "fmax v20.4s, v20.4s, v0.4s\n" - "fmax v21.4s, v21.4s, v0.4s\n" - "fmax v22.4s, v22.4s, v0.4s\n" - "fmax v23.4s, v23.4s, v0.4s\n" - "fmax v24.4s, v24.4s, v0.4s\n" - "fmax v25.4s, v25.4s, v0.4s\n" - "fmax v26.4s, v26.4s, v0.4s\n" - "fmax v27.4s, v27.4s, v0.4s\n" + "ld1r { v28.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v29.4s\n" + "fmin v9.4s, v9.4s, v29.4s\n" + "fmin v10.4s, v10.4s, v29.4s\n" + "fmin v11.4s, v11.4s, v29.4s\n" + "fmin v12.4s, v12.4s, v29.4s\n" + "fmin v13.4s, v13.4s, v29.4s\n" + "fmin v14.4s, v14.4s, v29.4s\n" + "fmin v15.4s, v15.4s, v29.4s\n" + "fmin v16.4s, v16.4s, v29.4s\n" + "fmin v17.4s, v17.4s, v29.4s\n" + "fmin v18.4s, v18.4s, v29.4s\n" + "fmin v19.4s, v19.4s, v29.4s\n" + "fmin v20.4s, v20.4s, v29.4s\n" + "fmin v21.4s, v21.4s, v29.4s\n" + "fmin v22.4s, v22.4s, v29.4s\n" + "fmin v23.4s, v23.4s, v29.4s\n" + "fmin v24.4s, v24.4s, v29.4s\n" + "fmin v25.4s, v25.4s, v29.4s\n" + "fmin v26.4s, v26.4s, v29.4s\n" + "fmin v27.4s, v27.4s, v29.4s\n" + "fmax v8.4s, v8.4s, v28.4s\n" + "fmax v9.4s, v9.4s, v28.4s\n" + "fmax v10.4s, v10.4s, v28.4s\n" + "fmax v11.4s, v11.4s, v28.4s\n" + "fmax v12.4s, v12.4s, v28.4s\n" + "fmax v13.4s, v13.4s, v28.4s\n" + "fmax v14.4s, v14.4s, v28.4s\n" + "fmax v15.4s, v15.4s, v28.4s\n" + "fmax v16.4s, v16.4s, v28.4s\n" + "fmax v17.4s, v17.4s, v28.4s\n" + "fmax v18.4s, v18.4s, v28.4s\n" + "fmax v19.4s, v19.4s, v28.4s\n" + "fmax v20.4s, v20.4s, v28.4s\n" + "fmax v21.4s, v21.4s, v28.4s\n" + "fmax v22.4s, v22.4s, v28.4s\n" + "fmax v23.4s, v23.4s, v28.4s\n" + "fmax v24.4s, v24.4s, v28.4s\n" + "fmax v25.4s, v25.4s, v28.4s\n" + "fmax v26.4s, v26.4s, v28.4s\n" + "fmax v27.4s, v27.4s, v28.4s\n" "165:" // Height 5: No activation "cmp x11, #0x10\n" "bge 174f\n" @@ -2902,16 +2901,16 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "190:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 191f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" - "ldr x21, [x21, #0x28]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" + "ldr x21, [x20, #0x28]\n" "cbnz x28, 192f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -2923,11 +2922,11 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "b 192f\n" "191:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" - "add x22, x23, x20, LSL #1\n" - "add x21, x22, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" + "add x22, x23, x21, LSL #1\n" + "add x21, x22, x21, LSL #1\n" "192:" // Height 6: input setup done "cmp x27, #0x8\n" "blt 195f\n" @@ -3206,43 +3205,43 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "cmp x27, #0x2\n" "blt 197f\n" "196:" // Height 6: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr s1, [x25], #0x4\n" + "ldr s7, [x26], #0x4\n" + "ldr s6, [x25], #0x4\n" "sub x27, x27, #0x2\n" "cmp x27, #0x2\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr s4, [x22], #0x4\n" - "ldr s5, [x21], #0x4\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - ".inst 0x4f45f0dc // bfdot v28.4s, v6.8h, v5.h[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\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" + "ldr s5, [x24], #0x4\n" + "ldr s4, [x23], #0x4\n" + "ldr s3, [x22], #0x4\n" + "ldr s2, [x21], #0x4\n" + "ldr q1, [x10, #0x0]\n" + "ldr q0, [x10, #0x10]\n" + ".inst 0x4f47f028 // bfdot v8.4s, v1.8h, v7.h[0]\n" + ".inst 0x4f46f02c // bfdot v12.4s, v1.8h, v6.h[0]\n" + ".inst 0x4f45f030 // bfdot v16.4s, v1.8h, v5.h[0]\n" + ".inst 0x4f44f034 // bfdot v20.4s, v1.8h, v4.h[0]\n" + ".inst 0x4f43f038 // bfdot v24.4s, v1.8h, v3.h[0]\n" + ".inst 0x4f42f03c // bfdot v28.4s, v1.8h, v2.h[0]\n" + "ldr q1, [x10, #0x20]\n" + ".inst 0x4f47f009 // bfdot v9.4s, v0.8h, v7.h[0]\n" + ".inst 0x4f46f00d // bfdot v13.4s, v0.8h, v6.h[0]\n" + ".inst 0x4f45f011 // bfdot v17.4s, v0.8h, v5.h[0]\n" + ".inst 0x4f44f015 // bfdot v21.4s, v0.8h, v4.h[0]\n" + ".inst 0x4f43f019 // bfdot v25.4s, v0.8h, v3.h[0]\n" + ".inst 0x4f42f01d // bfdot v29.4s, v0.8h, v2.h[0]\n" + "ldr q0, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" - ".inst 0x4f44f0da // bfdot v26.4s, v6.8h, v4.h[0]\n" - ".inst 0x4f45f0de // bfdot v30.4s, v6.8h, v5.h[0]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" - ".inst 0x4f45f0ff // bfdot v31.4s, v7.8h, v5.h[0]\n" + ".inst 0x4f47f02a // bfdot v10.4s, v1.8h, v7.h[0]\n" + ".inst 0x4f46f02e // bfdot v14.4s, v1.8h, v6.h[0]\n" + ".inst 0x4f45f032 // bfdot v18.4s, v1.8h, v5.h[0]\n" + ".inst 0x4f44f036 // bfdot v22.4s, v1.8h, v4.h[0]\n" + ".inst 0x4f43f03a // bfdot v26.4s, v1.8h, v3.h[0]\n" + ".inst 0x4f42f03e // bfdot v30.4s, v1.8h, v2.h[0]\n" + ".inst 0x4f47f00b // bfdot v11.4s, v0.8h, v7.h[0]\n" + ".inst 0x4f46f00f // bfdot v15.4s, v0.8h, v6.h[0]\n" + ".inst 0x4f45f013 // bfdot v19.4s, v0.8h, v5.h[0]\n" + ".inst 0x4f44f017 // bfdot v23.4s, v0.8h, v4.h[0]\n" + ".inst 0x4f43f01b // bfdot v27.4s, v0.8h, v3.h[0]\n" + ".inst 0x4f42f01f // bfdot v31.4s, v0.8h, v2.h[0]\n" "bge 196b\n" "197:" // Height 6: Multiply loop: Skip odd blocks "cbz x27, 199f\n" @@ -3253,35 +3252,35 @@ void a64_hybrid_bf16fp32_dot_6x16 ( "ldr h4, [x22, #0x0]\n" "ldr h5, [x21, #0x0]\n" "198:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x4f40f0c8 // bfdot v8.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0cc // bfdot v12.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d0 // bfdot v16.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d4 // bfdot v20.4s, v6.8h, v3.h[0]\n" - ".inst 0x4f44f0d8 // bfdot v24.4s, v6.8h, v4.h[0]\n" - ".inst 0x4f45f0dc // bfdot v28.4s, v6.8h, v5.h[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x4f40f0e9 // bfdot v9.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ed // bfdot v13.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f1 // bfdot v17.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f5 // bfdot v21.4s, v7.8h, v3.h[0]\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" + "ldr q7, [x10, #0x0]\n" + "ldr q6, [x10, #0x10]\n" + ".inst 0x4f40f0e8 // bfdot v8.4s, v7.8h, v0.h[0]\n" + ".inst 0x4f41f0ec // bfdot v12.4s, v7.8h, v1.h[0]\n" + ".inst 0x4f42f0f0 // bfdot v16.4s, v7.8h, v2.h[0]\n" + ".inst 0x4f43f0f4 // bfdot v20.4s, v7.8h, v3.h[0]\n" + ".inst 0x4f44f0f8 // bfdot v24.4s, v7.8h, v4.h[0]\n" + ".inst 0x4f45f0fc // bfdot v28.4s, v7.8h, v5.h[0]\n" + "ldr q7, [x10, #0x20]\n" + ".inst 0x4f40f0c9 // bfdot v9.4s, v6.8h, v0.h[0]\n" + ".inst 0x4f41f0cd // bfdot v13.4s, v6.8h, v1.h[0]\n" + ".inst 0x4f42f0d1 // bfdot v17.4s, v6.8h, v2.h[0]\n" + ".inst 0x4f43f0d5 // bfdot v21.4s, v6.8h, v3.h[0]\n" + ".inst 0x4f44f0d9 // bfdot v25.4s, v6.8h, v4.h[0]\n" + ".inst 0x4f45f0dd // bfdot v29.4s, v6.8h, v5.h[0]\n" + "ldr q6, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x4f40f0ca // bfdot v10.4s, v6.8h, v0.h[0]\n" - ".inst 0x4f41f0ce // bfdot v14.4s, v6.8h, v1.h[0]\n" - ".inst 0x4f42f0d2 // bfdot v18.4s, v6.8h, v2.h[0]\n" - ".inst 0x4f43f0d6 // bfdot v22.4s, v6.8h, v3.h[0]\n" - ".inst 0x4f44f0da // bfdot v26.4s, v6.8h, v4.h[0]\n" - ".inst 0x4f45f0de // bfdot v30.4s, v6.8h, v5.h[0]\n" - ".inst 0x4f40f0eb // bfdot v11.4s, v7.8h, v0.h[0]\n" - ".inst 0x4f41f0ef // bfdot v15.4s, v7.8h, v1.h[0]\n" - ".inst 0x4f42f0f3 // bfdot v19.4s, v7.8h, v2.h[0]\n" - ".inst 0x4f43f0f7 // bfdot v23.4s, v7.8h, v3.h[0]\n" - ".inst 0x4f44f0fb // bfdot v27.4s, v7.8h, v4.h[0]\n" - ".inst 0x4f45f0ff // bfdot v31.4s, v7.8h, v5.h[0]\n" + ".inst 0x4f40f0ea // bfdot v10.4s, v7.8h, v0.h[0]\n" + ".inst 0x4f41f0ee // bfdot v14.4s, v7.8h, v1.h[0]\n" + ".inst 0x4f42f0f2 // bfdot v18.4s, v7.8h, v2.h[0]\n" + ".inst 0x4f43f0f6 // bfdot v22.4s, v7.8h, v3.h[0]\n" + ".inst 0x4f44f0fa // bfdot v26.4s, v7.8h, v4.h[0]\n" + ".inst 0x4f45f0fe // bfdot v30.4s, v7.8h, v5.h[0]\n" + ".inst 0x4f40f0cb // bfdot v11.4s, v6.8h, v0.h[0]\n" + ".inst 0x4f41f0cf // bfdot v15.4s, v6.8h, v1.h[0]\n" + ".inst 0x4f42f0d3 // bfdot v19.4s, v6.8h, v2.h[0]\n" + ".inst 0x4f43f0d7 // bfdot v23.4s, v6.8h, v3.h[0]\n" + ".inst 0x4f44f0db // bfdot v27.4s, v6.8h, v4.h[0]\n" + ".inst 0x4f45f0df // bfdot v31.4s, v6.8h, v5.h[0]\n" "199:" // Height 6: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x28, x28, #0x1\n" @@ -3522,7 +3521,6 @@ 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)) : "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" |