From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 12 Apr 2023 17:43:17 +0100 Subject: Update CPU kernel implementations and guard directives Resolves COMPMID-6023 Change-Id: I868975d14c4f98af6716726feda22405a6a4c891 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686 Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp | 2542 ++++++++++---------- 1 file changed, 1270 insertions(+), 1272 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp index 8046b2ebb0..289d38c3b6 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_s8s32_dot_6x16/a55.cpp @@ -77,7 +77,6 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( ka.N = N; ka.B_ptr = B_ptr; __asm__ __volatile__( - "1:" // Row loop "cmp %x[M], #0x6\n" "bge 171f\n" @@ -165,11 +164,11 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "14:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w14, [x20, x15, 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, 15f\n" - "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x13, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x13, [x20, #0x0]\n" "cbnz x15, 16f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x13, x13, x20\n" @@ -186,129 +185,129 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "blt 18f\n" "17:" // Height 1: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr d6, [x16, #0x20]\n" - "ldr x12, [x16, #0x28]\n" + "ldr d17, [x16, #0x20]\n" + "ldr x20, [x16, #0x28]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr d7, [x16, #0x30]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0x38]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "ldr d6, [x16, #0x40]\n" - "ldr x12, [x16, #0x48]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "ldr d7, [x16, #0x50]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0x58]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - "ldr d6, [x16, #0x60]\n" - "ldr x12, [x16, #0x68]\n" - ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "ldr d7, [x16, #0x70]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0x78]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - "ldr d6, [x16, #0x80]\n" - "ldr x12, [x16, #0x88]\n" - ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "ldr d7, [x16, #0x90]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0x98]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - "ldr d6, [x16, #0xa0]\n" - "ldr x12, [x16, #0xa8]\n" - ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "ldr d7, [x16, #0xb0]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0xb8]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - "ldr d6, [x16, #0xc0]\n" - "ldr x12, [x16, #0xc8]\n" - ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "ldr d7, [x16, #0xd0]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0xd8]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - "ldr d6, [x16, #0xe0]\n" - "ldr x12, [x16, #0xe8]\n" - ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "ldr d7, [x16, #0xf0]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0xf8]\n" - "mov v7.d[1], x11\n" + "ldr d16, [x16, #0x30]\n" + "mov v17.d[1], x20\n" + "ldr x20, [x16, #0x38]\n" + "mov v16.d[1], x20\n" + ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n" + "ldr d17, [x16, #0x40]\n" + "ldr x20, [x16, #0x48]\n" + ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n" + "ldr d16, [x16, #0x50]\n" + "mov v17.d[1], x20\n" + "ldr x20, [x16, #0x58]\n" + "mov v16.d[1], x20\n" + ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n" + "ldr d17, [x16, #0x60]\n" + "ldr x20, [x16, #0x68]\n" + ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n" + "ldr d16, [x16, #0x70]\n" + "mov v17.d[1], x20\n" + "ldr x20, [x16, #0x78]\n" + "mov v16.d[1], x20\n" + ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n" + "ldr d17, [x16, #0x80]\n" + "ldr x20, [x16, #0x88]\n" + ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n" + "ldr d16, [x16, #0x90]\n" + "mov v17.d[1], x20\n" + "ldr x20, [x16, #0x98]\n" + "mov v16.d[1], x20\n" + ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n" + "ldr d17, [x16, #0xa0]\n" + "ldr x20, [x16, #0xa8]\n" + ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n" + "ldr d16, [x16, #0xb0]\n" + "mov v17.d[1], x20\n" + "ldr x20, [x16, #0xb8]\n" + "mov v16.d[1], x20\n" + ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n" + "ldr d17, [x16, #0xc0]\n" + "ldr x20, [x16, #0xc8]\n" + ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n" + "ldr d16, [x16, #0xd0]\n" + "mov v17.d[1], x20\n" + "ldr x20, [x16, #0xd8]\n" + "mov v16.d[1], x20\n" + ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n" + "ldr d17, [x16, #0xe0]\n" + "ldr x20, [x16, #0xe8]\n" + ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n" + "ldr d16, [x16, #0xf0]\n" + "mov v17.d[1], x20\n" + "ldr x20, [x16, #0xf8]\n" + "mov v16.d[1], x20\n" "add x13, x13, #0x10\n" "add x16, x16, #0x100\n" - ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n" "ldr d6, [x16, #0x0]\n" - "ldr x12, [x16, #0x8]\n" - ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr x20, [x16, #0x8]\n" + ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n" "ldr d0, [x13, #0x0]\n" "sub x14, x14, #0x10\n" "ldr d7, [x16, #0x10]\n" "cmp x14, #0x20\n" - "ldr x10, [x13, #0x8]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0x18]\n" - "mov v0.d[1], x10\n" - "mov v7.d[1], x11\n" + "ldr x21, [x13, #0x8]\n" + "mov v6.d[1], x20\n" + "ldr x20, [x16, #0x18]\n" + "mov v0.d[1], x21\n" + "mov v7.d[1], x20\n" "prfm pldl1keep, [x13, #0x80]\n" "bge 17b\n" "18:" // Height 1: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x16, #0x20]\n" + "ldr q17, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x16, #0x40]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x50]\n" - ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x16, #0x60]\n" - ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x16, #0x70]\n" - ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x16, #0x80]\n" - ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x16, #0x90]\n" - ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x16, #0xa0]\n" - ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x16, #0xb0]\n" - ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x16, #0xc0]\n" - ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x16, #0xd0]\n" - ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - "ldr q6, [x16, #0xe0]\n" - ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "ldr q7, [x16, #0xf0]\n" + "ldr q16, [x16, #0x30]\n" + ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n" + "ldr q17, [x16, #0x40]\n" + ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n" + "ldr q16, [x16, #0x50]\n" + ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n" + "ldr q17, [x16, #0x60]\n" + ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n" + "ldr q16, [x16, #0x70]\n" + ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n" + "ldr q17, [x16, #0x80]\n" + ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n" + "ldr q16, [x16, #0x90]\n" + ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n" + "ldr q17, [x16, #0xa0]\n" + ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n" + "ldr q16, [x16, #0xb0]\n" + ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n" + "ldr q17, [x16, #0xc0]\n" + ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n" + "ldr q16, [x16, #0xd0]\n" + ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n" + "ldr q17, [x16, #0xe0]\n" + ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n" + "ldr q16, [x16, #0xf0]\n" "add x13, x13, #0x10\n" "sub x14, x14, #0x10\n" - ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n" "prfm pldl1keep, [x13, #0x80]\n" - ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n" "add x16, x16, #0x100\n" "19:" // Height 1: Multiply loop: Main loop skip "cbz x14, 24f\n" "cmp x14, #0x4\n" "blt 21f\n" "20:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x13], #0x4\n" + "ldr s18, [x13], #0x4\n" "sub x14, x14, #0x4\n" - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q6, [x16, #0x20]\n" + "ldr q16, [x16, #0x0]\n" + ".inst 0x4f92e208 // sdot v8.4s, v16.16b, v18.4b[0]\n" + "ldr q16, [x16, #0x10]\n" + ".inst 0x4f92e209 // sdot v9.4s, v16.16b, v18.4b[0]\n" + "ldr q17, [x16, #0x20]\n" "cmp x14, #0x4\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "ldr q16, [x16, #0x30]\n" + ".inst 0x4f92e22a // sdot v10.4s, v17.16b, v18.4b[0]\n" + ".inst 0x4f92e20b // sdot v11.4s, v16.16b, v18.4b[0]\n" "add x16, x16, #0x40\n" "bge 20b\n" "21:" // Height 1: Multiply loop: Skip odd blocks @@ -321,14 +320,14 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "22:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 "ldr b0, [x13, #0x0]\n" "23:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q6, [x16, #0x20]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + "ldr q16, [x16, #0x0]\n" + ".inst 0x4f80e208 // sdot v8.4s, v16.16b, v0.4b[0]\n" + "ldr q16, [x16, #0x10]\n" + ".inst 0x4f80e209 // sdot v9.4s, v16.16b, v0.4b[0]\n" + "ldr q16, [x16, #0x20]\n" + ".inst 0x4f80e20a // sdot v10.4s, v16.16b, v0.4b[0]\n" + "ldr q16, [x16, #0x30]\n" + ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n" "add x16, x16, #0x40\n" "24:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -499,226 +498,226 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "48:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w14, [x20, x15, 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, 49f\n" - "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x13, [x21, #0x0]\n" - "ldr x9, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x13, [x20, #0x0]\n" + "ldr x12, [x20, #0x8]\n" "cbnz x15, 50f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x13, x13, x20\n" - "add x9, x9, x20\n" + "add x12, x12, x20\n" "b 50f\n" "49:" // Height 2: setup direct input "mov x13, %x[input_ptr]\n" - "add x9, x13, x20\n" + "add x12, x13, x21\n" "50:" // Height 2: input setup done "cmp x14, #0x10\n" "blt 53f\n" "ldr q0, [x13, #0x0]\n" "cmp x14, #0x20\n" - "ldr q1, [x9, #0x0]\n" + "ldr q1, [x12, #0x0]\n" "ldr q6, [x16, #0x0]\n" "ldr q7, [x16, #0x10]\n" "blt 52f\n" "51:" // Height 2: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr x12, [x16, #0x28]\n" + "ldr x21, [x16, #0x28]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr d6, [x16, #0x20]\n" + "ldr d17, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "ldr x11, [x16, #0x38]\n" + "ldr x20, [x16, #0x38]\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr d7, [x16, #0x30]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "ldr d6, [x16, #0x40]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "ldr x12, [x16, #0x48]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr d7, [x16, #0x50]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0x58]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - "ldr x12, [x16, #0x68]\n" - ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "ldr d6, [x16, #0x60]\n" - ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "ldr x11, [x16, #0x78]\n" - ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "ldr d7, [x16, #0x70]\n" - "mov v6.d[1], x12\n" - ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "ldr d6, [x16, #0x80]\n" - ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "ldr x12, [x16, #0x88]\n" - ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "ldr d7, [x16, #0x90]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0x98]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - "ldr x12, [x16, #0xa8]\n" - ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "ldr d6, [x16, #0xa0]\n" - ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "ldr x11, [x16, #0xb8]\n" - ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "ldr d7, [x16, #0xb0]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "ldr d6, [x16, #0xc0]\n" - ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "ldr x12, [x16, #0xc8]\n" - ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "ldr d7, [x16, #0xd0]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0xd8]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - "ldr x12, [x16, #0xe8]\n" - ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "ldr d6, [x16, #0xe0]\n" - ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "ldr x11, [x16, #0xf8]\n" - ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - "ldr d7, [x16, #0xf0]\n" - "mov v6.d[1], x12\n" + "ldr d16, [x16, #0x30]\n" + "mov v17.d[1], x21\n" + ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n" + "mov v16.d[1], x20\n" + ".inst 0x4f81e22e // sdot v14.4s, v17.16b, v1.4b[0]\n" + "ldr d17, [x16, #0x40]\n" + ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n" + "ldr x20, [x16, #0x48]\n" + ".inst 0x4f81e20f // sdot v15.4s, v16.16b, v1.4b[0]\n" + "ldr d16, [x16, #0x50]\n" + "mov v17.d[1], x20\n" + "ldr x20, [x16, #0x58]\n" + "mov v16.d[1], x20\n" + ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n" + "ldr x21, [x16, #0x68]\n" + ".inst 0x4fa1e22c // sdot v12.4s, v17.16b, v1.4b[1]\n" + "ldr d17, [x16, #0x60]\n" + ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n" + "ldr x20, [x16, #0x78]\n" + ".inst 0x4fa1e20d // sdot v13.4s, v16.16b, v1.4b[1]\n" + "ldr d16, [x16, #0x70]\n" + "mov v17.d[1], x21\n" + ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n" + "mov v16.d[1], x20\n" + ".inst 0x4fa1e22e // sdot v14.4s, v17.16b, v1.4b[1]\n" + "ldr d17, [x16, #0x80]\n" + ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n" + "ldr x20, [x16, #0x88]\n" + ".inst 0x4fa1e20f // sdot v15.4s, v16.16b, v1.4b[1]\n" + "ldr d16, [x16, #0x90]\n" + "mov v17.d[1], x20\n" + "ldr x20, [x16, #0x98]\n" + "mov v16.d[1], x20\n" + ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n" + "ldr x21, [x16, #0xa8]\n" + ".inst 0x4f81ea2c // sdot v12.4s, v17.16b, v1.4b[2]\n" + "ldr d17, [x16, #0xa0]\n" + ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n" + "ldr x20, [x16, #0xb8]\n" + ".inst 0x4f81ea0d // sdot v13.4s, v16.16b, v1.4b[2]\n" + "ldr d16, [x16, #0xb0]\n" + "mov v17.d[1], x21\n" + ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n" + "mov v16.d[1], x20\n" + ".inst 0x4f81ea2e // sdot v14.4s, v17.16b, v1.4b[2]\n" + "ldr d17, [x16, #0xc0]\n" + ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n" + "ldr x20, [x16, #0xc8]\n" + ".inst 0x4f81ea0f // sdot v15.4s, v16.16b, v1.4b[2]\n" + "ldr d16, [x16, #0xd0]\n" + "mov v17.d[1], x20\n" + "ldr x20, [x16, #0xd8]\n" + "mov v16.d[1], x20\n" + ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n" + "ldr x21, [x16, #0xe8]\n" + ".inst 0x4fa1ea2c // sdot v12.4s, v17.16b, v1.4b[3]\n" + "ldr d17, [x16, #0xe0]\n" + ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n" + "ldr x20, [x16, #0xf8]\n" + ".inst 0x4fa1ea0d // sdot v13.4s, v16.16b, v1.4b[3]\n" + "ldr d16, [x16, #0xf0]\n" + "mov v17.d[1], x21\n" "add x13, x13, #0x10\n" - "mov v7.d[1], x11\n" - "add x9, x9, #0x10\n" + "mov v16.d[1], x20\n" + "add x12, x12, #0x10\n" "add x16, x16, #0x100\n" - ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" + ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n" + ".inst 0x4fa1ea2e // sdot v14.4s, v17.16b, v1.4b[3]\n" "ldr d6, [x16, #0x0]\n" - "ldr x12, [x16, #0x8]\n" - ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + "ldr x21, [x16, #0x8]\n" + ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n" "ldr d0, [x13, #0x0]\n" - ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "ldr d1, [x9, #0x0]\n" + ".inst 0x4fa1ea0f // sdot v15.4s, v16.16b, v1.4b[3]\n" + "ldr d1, [x12, #0x0]\n" "sub x14, x14, #0x10\n" "ldr d7, [x16, #0x10]\n" "cmp x14, #0x20\n" - "ldr x10, [x13, #0x8]\n" - "mov v6.d[1], x12\n" - "ldr x28, [x9, #0x8]\n" - "mov v0.d[1], x10\n" - "ldr x11, [x16, #0x18]\n" - "mov v1.d[1], x28\n" + "ldr x20, [x13, #0x8]\n" + "mov v6.d[1], x21\n" + "ldr x21, [x12, #0x8]\n" + "mov v0.d[1], x20\n" + "ldr x20, [x16, #0x18]\n" + "mov v1.d[1], x21\n" "prfm pldl1keep, [x13, #0x80]\n" - "mov v7.d[1], x11\n" - "prfm pldl1keep, [x9, #0x80]\n" + "mov v7.d[1], x20\n" + "prfm pldl1keep, [x12, #0x80]\n" "bge 51b\n" "52:" // Height 2: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" "add x13, x13, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x16, #0x20]\n" + "ldr q17, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x9, x9, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr q16, [x16, #0x30]\n" + ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n" "sub x14, x14, #0x10\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x16, #0x40]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" + ".inst 0x4f81e22e // sdot v14.4s, v17.16b, v1.4b[0]\n" + "ldr q17, [x16, #0x40]\n" + ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n" "prfm pldl1keep, [x13, #0x80]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x16, #0x50]\n" - ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - "prfm pldl1keep, [x9, #0x80]\n" - ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x16, #0x60]\n" - ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x16, #0x70]\n" - ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x16, #0x80]\n" - ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x16, #0x90]\n" - ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x16, #0xa0]\n" - ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x16, #0xb0]\n" - ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x16, #0xc0]\n" - ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x16, #0xd0]\n" - ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "ldr q6, [x16, #0xe0]\n" - ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - "ldr q7, [x16, #0xf0]\n" - ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + ".inst 0x4f81e20f // sdot v15.4s, v16.16b, v1.4b[0]\n" + "ldr q16, [x16, #0x50]\n" + ".inst 0x4fa0e228 // sdot v8.4s, v17.16b, v0.4b[1]\n" + "prfm pldl1keep, [x12, #0x80]\n" + ".inst 0x4fa1e22c // sdot v12.4s, v17.16b, v1.4b[1]\n" + "ldr q17, [x16, #0x60]\n" + ".inst 0x4fa0e209 // sdot v9.4s, v16.16b, v0.4b[1]\n" + ".inst 0x4fa1e20d // sdot v13.4s, v16.16b, v1.4b[1]\n" + "ldr q16, [x16, #0x70]\n" + ".inst 0x4fa0e22a // sdot v10.4s, v17.16b, v0.4b[1]\n" + ".inst 0x4fa1e22e // sdot v14.4s, v17.16b, v1.4b[1]\n" + "ldr q17, [x16, #0x80]\n" + ".inst 0x4fa0e20b // sdot v11.4s, v16.16b, v0.4b[1]\n" + ".inst 0x4fa1e20f // sdot v15.4s, v16.16b, v1.4b[1]\n" + "ldr q16, [x16, #0x90]\n" + ".inst 0x4f80ea28 // sdot v8.4s, v17.16b, v0.4b[2]\n" + ".inst 0x4f81ea2c // sdot v12.4s, v17.16b, v1.4b[2]\n" + "ldr q17, [x16, #0xa0]\n" + ".inst 0x4f80ea09 // sdot v9.4s, v16.16b, v0.4b[2]\n" + ".inst 0x4f81ea0d // sdot v13.4s, v16.16b, v1.4b[2]\n" + "ldr q16, [x16, #0xb0]\n" + ".inst 0x4f80ea2a // sdot v10.4s, v17.16b, v0.4b[2]\n" + ".inst 0x4f81ea2e // sdot v14.4s, v17.16b, v1.4b[2]\n" + "ldr q17, [x16, #0xc0]\n" + ".inst 0x4f80ea0b // sdot v11.4s, v16.16b, v0.4b[2]\n" + ".inst 0x4f81ea0f // sdot v15.4s, v16.16b, v1.4b[2]\n" + "ldr q16, [x16, #0xd0]\n" + ".inst 0x4fa0ea28 // sdot v8.4s, v17.16b, v0.4b[3]\n" + ".inst 0x4fa1ea2c // sdot v12.4s, v17.16b, v1.4b[3]\n" + "ldr q17, [x16, #0xe0]\n" + ".inst 0x4fa0ea09 // sdot v9.4s, v16.16b, v0.4b[3]\n" + ".inst 0x4fa1ea0d // sdot v13.4s, v16.16b, v1.4b[3]\n" + "ldr q16, [x16, #0xf0]\n" + ".inst 0x4fa0ea2a // sdot v10.4s, v17.16b, v0.4b[3]\n" "add x16, x16, #0x100\n" - ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" - ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" + ".inst 0x4fa1ea2e // sdot v14.4s, v17.16b, v1.4b[3]\n" + ".inst 0x4fa0ea0b // sdot v11.4s, v16.16b, v0.4b[3]\n" + ".inst 0x4fa1ea0f // sdot v15.4s, v16.16b, v1.4b[3]\n" "53:" // Height 2: Multiply loop: Main loop skip "cbz x14, 58f\n" "cmp x14, #0x4\n" "blt 55f\n" "54:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x13], #0x4\n" + "ldr s19, [x13], #0x4\n" "sub x14, x14, #0x4\n" - "ldr s1, [x9], #0x4\n" + "ldr s18, [x12], #0x4\n" "cmp x14, #0x4\n" - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x16, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr q17, [x16, #0x0]\n" + ".inst 0x4f93e228 // sdot v8.4s, v17.16b, v19.4b[0]\n" + "ldr q16, [x16, #0x10]\n" + ".inst 0x4f92e22c // sdot v12.4s, v17.16b, v18.4b[0]\n" + "ldr q17, [x16, #0x20]\n" + ".inst 0x4f93e209 // sdot v9.4s, v16.16b, v19.4b[0]\n" + ".inst 0x4f92e20d // sdot v13.4s, v16.16b, v18.4b[0]\n" + "ldr q16, [x16, #0x30]\n" + ".inst 0x4f93e22a // sdot v10.4s, v17.16b, v19.4b[0]\n" "add x16, x16, #0x40\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" + ".inst 0x4f92e22e // sdot v14.4s, v17.16b, v18.4b[0]\n" + ".inst 0x4f93e20b // sdot v11.4s, v16.16b, v19.4b[0]\n" + ".inst 0x4f92e20f // sdot v15.4s, v16.16b, v18.4b[0]\n" "bge 54b\n" "55:" // Height 2: Multiply loop: Skip odd blocks "cbz x14, 58f\n" "tbz x14, #1, 56f\n" "ldr h0, [x13], #0x2\n" - "ldr h1, [x9], #0x2\n" + "ldr h1, [x12], #0x2\n" "tbz x14, #0, 57f\n" "ld1 { v0.b }[2], [x13]\n" - "ld1 { v1.b }[2], [x9]\n" + "ld1 { v1.b }[2], [x12]\n" "b 57f\n" "56:" // Height 2: Multiply loop: Ragged operand read: partial_1_0 "ldr b0, [x13, #0x0]\n" - "ldr b1, [x9, #0x0]\n" + "ldr b1, [x12, #0x0]\n" "57:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x16, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr q17, [x16, #0x0]\n" + ".inst 0x4f80e228 // sdot v8.4s, v17.16b, v0.4b[0]\n" + "ldr q16, [x16, #0x10]\n" + ".inst 0x4f81e22c // sdot v12.4s, v17.16b, v1.4b[0]\n" + "ldr q17, [x16, #0x20]\n" + ".inst 0x4f80e209 // sdot v9.4s, v16.16b, v0.4b[0]\n" + ".inst 0x4f81e20d // sdot v13.4s, v16.16b, v1.4b[0]\n" + "ldr q16, [x16, #0x30]\n" + ".inst 0x4f80e22a // sdot v10.4s, v17.16b, v0.4b[0]\n" "add x16, x16, #0x40\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" + ".inst 0x4f81e22e // sdot v14.4s, v17.16b, v1.4b[0]\n" + ".inst 0x4f80e20b // sdot v11.4s, v16.16b, v0.4b[0]\n" + ".inst 0x4f81e20f // sdot v15.4s, v16.16b, v1.4b[0]\n" "58:" // Height 2: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x15, x15, #0x1\n" @@ -936,281 +935,281 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "82:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w14, [x20, x15, 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, 83f\n" - "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x13, [x21, #0x0]\n" - "ldr x9, [x21, #0x8]\n" - "ldr x27, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x13, [x20, #0x0]\n" + "ldr x12, [x20, #0x8]\n" + "ldr x11, [x20, #0x10]\n" "cbnz x15, 84f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x13, x13, x20\n" - "add x9, x9, x20\n" - "add x27, x27, x20\n" + "add x12, x12, x20\n" + "add x11, x11, x20\n" "b 84f\n" "83:" // Height 3: setup direct input "mov x13, %x[input_ptr]\n" - "add x9, x13, x20\n" - "add x27, x9, x20\n" + "add x12, x13, x21\n" + "add x11, x12, x21\n" "84:" // Height 3: input setup done "cmp x14, #0x10\n" "blt 87f\n" "ldr q0, [x13, #0x0]\n" "cmp x14, #0x20\n" - "ldr q1, [x9, #0x0]\n" - "ldr q2, [x27, #0x0]\n" + "ldr q1, [x12, #0x0]\n" + "ldr q2, [x11, #0x0]\n" "ldr q6, [x16, #0x0]\n" "ldr q7, [x16, #0x10]\n" "blt 86f\n" "85:" // Height 3: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr x12, [x16, #0x28]\n" + "ldr x21, [x16, #0x28]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr x11, [x16, #0x38]\n" + "ldr x20, [x16, #0x38]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr d6, [x16, #0x20]\n" + "ldr d21, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "mov v6.d[1], x12\n" + "mov v21.d[1], x21\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr x12, [x16, #0x48]\n" + "ldr x21, [x16, #0x48]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "ldr d7, [x16, #0x30]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "ldr x11, [x16, #0x58]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "ldr d6, [x16, #0x40]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr x12, [x16, #0x68]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "ldr d7, [x16, #0x50]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "ldr x11, [x16, #0x78]\n" - ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - "ldr d6, [x16, #0x60]\n" - ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "mov v6.d[1], x12\n" - ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "ldr x12, [x16, #0x88]\n" - ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - "ldr d7, [x16, #0x70]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "ldr x11, [x16, #0x98]\n" - ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" - "ldr d6, [x16, #0x80]\n" - ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "mov v6.d[1], x12\n" - ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "ldr x12, [x16, #0xa8]\n" - ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - "ldr d7, [x16, #0x90]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "ldr x11, [x16, #0xb8]\n" - ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" - "ldr d6, [x16, #0xa0]\n" - ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "ldr x12, [x16, #0xc8]\n" - ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" - "ldr d7, [x16, #0xb0]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "ldr x11, [x16, #0xd8]\n" - ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" - "ldr d6, [x16, #0xc0]\n" - ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "ldr x12, [x16, #0xe8]\n" - ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - "ldr d7, [x16, #0xd0]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "ldr x11, [x16, #0xf8]\n" - ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" - "ldr d6, [x16, #0xe0]\n" - ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "mov v6.d[1], x12\n" - ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" + "ldr d20, [x16, #0x30]\n" + "mov v20.d[1], x20\n" + ".inst 0x4f80e2aa // sdot v10.4s, v21.16b, v0.4b[0]\n" + ".inst 0x4f81e2ae // sdot v14.4s, v21.16b, v1.4b[0]\n" + "ldr x20, [x16, #0x58]\n" + ".inst 0x4f82e2b2 // sdot v18.4s, v21.16b, v2.4b[0]\n" + "ldr d21, [x16, #0x40]\n" + ".inst 0x4f80e28b // sdot v11.4s, v20.16b, v0.4b[0]\n" + "mov v21.d[1], x21\n" + ".inst 0x4f81e28f // sdot v15.4s, v20.16b, v1.4b[0]\n" + "ldr x21, [x16, #0x68]\n" + ".inst 0x4f82e293 // sdot v19.4s, v20.16b, v2.4b[0]\n" + "ldr d20, [x16, #0x50]\n" + "mov v20.d[1], x20\n" + ".inst 0x4fa0e2a8 // sdot v8.4s, v21.16b, v0.4b[1]\n" + ".inst 0x4fa1e2ac // sdot v12.4s, v21.16b, v1.4b[1]\n" + "ldr x20, [x16, #0x78]\n" + ".inst 0x4fa2e2b0 // sdot v16.4s, v21.16b, v2.4b[1]\n" + "ldr d21, [x16, #0x60]\n" + ".inst 0x4fa0e289 // sdot v9.4s, v20.16b, v0.4b[1]\n" + "mov v21.d[1], x21\n" + ".inst 0x4fa1e28d // sdot v13.4s, v20.16b, v1.4b[1]\n" + "ldr x21, [x16, #0x88]\n" + ".inst 0x4fa2e291 // sdot v17.4s, v20.16b, v2.4b[1]\n" + "ldr d20, [x16, #0x70]\n" + "mov v20.d[1], x20\n" + ".inst 0x4fa0e2aa // sdot v10.4s, v21.16b, v0.4b[1]\n" + ".inst 0x4fa1e2ae // sdot v14.4s, v21.16b, v1.4b[1]\n" + "ldr x20, [x16, #0x98]\n" + ".inst 0x4fa2e2b2 // sdot v18.4s, v21.16b, v2.4b[1]\n" + "ldr d21, [x16, #0x80]\n" + ".inst 0x4fa0e28b // sdot v11.4s, v20.16b, v0.4b[1]\n" + "mov v21.d[1], x21\n" + ".inst 0x4fa1e28f // sdot v15.4s, v20.16b, v1.4b[1]\n" + "ldr x21, [x16, #0xa8]\n" + ".inst 0x4fa2e293 // sdot v19.4s, v20.16b, v2.4b[1]\n" + "ldr d20, [x16, #0x90]\n" + "mov v20.d[1], x20\n" + ".inst 0x4f80eaa8 // sdot v8.4s, v21.16b, v0.4b[2]\n" + ".inst 0x4f81eaac // sdot v12.4s, v21.16b, v1.4b[2]\n" + "ldr x20, [x16, #0xb8]\n" + ".inst 0x4f82eab0 // sdot v16.4s, v21.16b, v2.4b[2]\n" + "ldr d21, [x16, #0xa0]\n" + ".inst 0x4f80ea89 // sdot v9.4s, v20.16b, v0.4b[2]\n" + "mov v21.d[1], x21\n" + ".inst 0x4f81ea8d // sdot v13.4s, v20.16b, v1.4b[2]\n" + "ldr x21, [x16, #0xc8]\n" + ".inst 0x4f82ea91 // sdot v17.4s, v20.16b, v2.4b[2]\n" + "ldr d20, [x16, #0xb0]\n" + "mov v20.d[1], x20\n" + ".inst 0x4f80eaaa // sdot v10.4s, v21.16b, v0.4b[2]\n" + ".inst 0x4f81eaae // sdot v14.4s, v21.16b, v1.4b[2]\n" + "ldr x20, [x16, #0xd8]\n" + ".inst 0x4f82eab2 // sdot v18.4s, v21.16b, v2.4b[2]\n" + "ldr d21, [x16, #0xc0]\n" + ".inst 0x4f80ea8b // sdot v11.4s, v20.16b, v0.4b[2]\n" + "mov v21.d[1], x21\n" + ".inst 0x4f81ea8f // sdot v15.4s, v20.16b, v1.4b[2]\n" + "ldr x21, [x16, #0xe8]\n" + ".inst 0x4f82ea93 // sdot v19.4s, v20.16b, v2.4b[2]\n" + "ldr d20, [x16, #0xd0]\n" + "mov v20.d[1], x20\n" + ".inst 0x4fa0eaa8 // sdot v8.4s, v21.16b, v0.4b[3]\n" + ".inst 0x4fa1eaac // sdot v12.4s, v21.16b, v1.4b[3]\n" + "ldr x20, [x16, #0xf8]\n" + ".inst 0x4fa2eab0 // sdot v16.4s, v21.16b, v2.4b[3]\n" + "ldr d21, [x16, #0xe0]\n" + ".inst 0x4fa0ea89 // sdot v9.4s, v20.16b, v0.4b[3]\n" + "mov v21.d[1], x21\n" + ".inst 0x4fa1ea8d // sdot v13.4s, v20.16b, v1.4b[3]\n" "add x13, x13, #0x10\n" - ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - "ldr d7, [x16, #0xf0]\n" - "mov v7.d[1], x11\n" - "add x9, x9, #0x10\n" - "add x27, x27, #0x10\n" + ".inst 0x4fa2ea91 // sdot v17.4s, v20.16b, v2.4b[3]\n" + "ldr d20, [x16, #0xf0]\n" + "mov v20.d[1], x20\n" + "add x12, x12, #0x10\n" + "add x11, x11, #0x10\n" "add x16, x16, #0x100\n" - ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "ldr x12, [x16, #0x8]\n" - ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - "ldr x10, [x13, #0x8]\n" - ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" + ".inst 0x4fa0eaaa // sdot v10.4s, v21.16b, v0.4b[3]\n" + "ldr x20, [x16, #0x8]\n" + ".inst 0x4fa1eaae // sdot v14.4s, v21.16b, v1.4b[3]\n" + "ldr x23, [x13, #0x8]\n" + ".inst 0x4fa2eab2 // sdot v18.4s, v21.16b, v2.4b[3]\n" "ldr d6, [x16, #0x0]\n" - ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + ".inst 0x4fa0ea8b // sdot v11.4s, v20.16b, v0.4b[3]\n" "ldr d0, [x13, #0x0]\n" - ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "ldr d1, [x9, #0x0]\n" - "ldr x28, [x9, #0x8]\n" - ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" - "ldr d2, [x27, #0x0]\n" + ".inst 0x4fa1ea8f // sdot v15.4s, v20.16b, v1.4b[3]\n" + "ldr d1, [x12, #0x0]\n" + "ldr x22, [x12, #0x8]\n" + ".inst 0x4fa2ea93 // sdot v19.4s, v20.16b, v2.4b[3]\n" + "ldr d2, [x11, #0x0]\n" "sub x14, x14, #0x10\n" "ldr d7, [x16, #0x10]\n" "cmp x14, #0x20\n" - "ldr x26, [x27, #0x8]\n" - "mov v6.d[1], x12\n" - "ldr x11, [x16, #0x18]\n" - "mov v0.d[1], x10\n" + "ldr x21, [x11, #0x8]\n" + "mov v6.d[1], x20\n" + "ldr x20, [x16, #0x18]\n" + "mov v0.d[1], x23\n" "prfm pldl1keep, [x13, #0x80]\n" - "mov v1.d[1], x28\n" - "prfm pldl1keep, [x9, #0x80]\n" - "mov v2.d[1], x26\n" - "prfm pldl1keep, [x27, #0x80]\n" - "mov v7.d[1], x11\n" + "mov v1.d[1], x22\n" + "prfm pldl1keep, [x12, #0x80]\n" + "mov v2.d[1], x21\n" + "prfm pldl1keep, [x11, #0x80]\n" + "mov v7.d[1], x20\n" "bge 85b\n" "86:" // Height 3: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" "add x13, x13, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x9, x9, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x16, #0x20]\n" + "ldr q21, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x27, x27, #0x10\n" + "add x11, x11, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" "sub x14, x14, #0x10\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr q20, [x16, #0x30]\n" + ".inst 0x4f80e2aa // sdot v10.4s, v21.16b, v0.4b[0]\n" "prfm pldl1keep, [x13, #0x80]\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x16, #0x40]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "prfm pldl1keep, [x27, #0x80]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x16, #0x50]\n" - ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - "ldr q6, [x16, #0x60]\n" - ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - "ldr q7, [x16, #0x70]\n" - ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" - "ldr q6, [x16, #0x80]\n" - ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - "ldr q7, [x16, #0x90]\n" - ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" - "ldr q6, [x16, #0xa0]\n" - ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" - "ldr q7, [x16, #0xb0]\n" - ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" - "ldr q6, [x16, #0xc0]\n" - ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - "ldr q7, [x16, #0xd0]\n" - ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" - "ldr q6, [x16, #0xe0]\n" - ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - "ldr q7, [x16, #0xf0]\n" - ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + ".inst 0x4f81e2ae // sdot v14.4s, v21.16b, v1.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" + ".inst 0x4f82e2b2 // sdot v18.4s, v21.16b, v2.4b[0]\n" + "ldr q21, [x16, #0x40]\n" + ".inst 0x4f80e28b // sdot v11.4s, v20.16b, v0.4b[0]\n" + "prfm pldl1keep, [x11, #0x80]\n" + ".inst 0x4f81e28f // sdot v15.4s, v20.16b, v1.4b[0]\n" + ".inst 0x4f82e293 // sdot v19.4s, v20.16b, v2.4b[0]\n" + "ldr q20, [x16, #0x50]\n" + ".inst 0x4fa0e2a8 // sdot v8.4s, v21.16b, v0.4b[1]\n" + ".inst 0x4fa1e2ac // sdot v12.4s, v21.16b, v1.4b[1]\n" + ".inst 0x4fa2e2b0 // sdot v16.4s, v21.16b, v2.4b[1]\n" + "ldr q21, [x16, #0x60]\n" + ".inst 0x4fa0e289 // sdot v9.4s, v20.16b, v0.4b[1]\n" + ".inst 0x4fa1e28d // sdot v13.4s, v20.16b, v1.4b[1]\n" + ".inst 0x4fa2e291 // sdot v17.4s, v20.16b, v2.4b[1]\n" + "ldr q20, [x16, #0x70]\n" + ".inst 0x4fa0e2aa // sdot v10.4s, v21.16b, v0.4b[1]\n" + ".inst 0x4fa1e2ae // sdot v14.4s, v21.16b, v1.4b[1]\n" + ".inst 0x4fa2e2b2 // sdot v18.4s, v21.16b, v2.4b[1]\n" + "ldr q21, [x16, #0x80]\n" + ".inst 0x4fa0e28b // sdot v11.4s, v20.16b, v0.4b[1]\n" + ".inst 0x4fa1e28f // sdot v15.4s, v20.16b, v1.4b[1]\n" + ".inst 0x4fa2e293 // sdot v19.4s, v20.16b, v2.4b[1]\n" + "ldr q20, [x16, #0x90]\n" + ".inst 0x4f80eaa8 // sdot v8.4s, v21.16b, v0.4b[2]\n" + ".inst 0x4f81eaac // sdot v12.4s, v21.16b, v1.4b[2]\n" + ".inst 0x4f82eab0 // sdot v16.4s, v21.16b, v2.4b[2]\n" + "ldr q21, [x16, #0xa0]\n" + ".inst 0x4f80ea89 // sdot v9.4s, v20.16b, v0.4b[2]\n" + ".inst 0x4f81ea8d // sdot v13.4s, v20.16b, v1.4b[2]\n" + ".inst 0x4f82ea91 // sdot v17.4s, v20.16b, v2.4b[2]\n" + "ldr q20, [x16, #0xb0]\n" + ".inst 0x4f80eaaa // sdot v10.4s, v21.16b, v0.4b[2]\n" + ".inst 0x4f81eaae // sdot v14.4s, v21.16b, v1.4b[2]\n" + ".inst 0x4f82eab2 // sdot v18.4s, v21.16b, v2.4b[2]\n" + "ldr q21, [x16, #0xc0]\n" + ".inst 0x4f80ea8b // sdot v11.4s, v20.16b, v0.4b[2]\n" + ".inst 0x4f81ea8f // sdot v15.4s, v20.16b, v1.4b[2]\n" + ".inst 0x4f82ea93 // sdot v19.4s, v20.16b, v2.4b[2]\n" + "ldr q20, [x16, #0xd0]\n" + ".inst 0x4fa0eaa8 // sdot v8.4s, v21.16b, v0.4b[3]\n" + ".inst 0x4fa1eaac // sdot v12.4s, v21.16b, v1.4b[3]\n" + ".inst 0x4fa2eab0 // sdot v16.4s, v21.16b, v2.4b[3]\n" + "ldr q21, [x16, #0xe0]\n" + ".inst 0x4fa0ea89 // sdot v9.4s, v20.16b, v0.4b[3]\n" + ".inst 0x4fa1ea8d // sdot v13.4s, v20.16b, v1.4b[3]\n" + ".inst 0x4fa2ea91 // sdot v17.4s, v20.16b, v2.4b[3]\n" + "ldr q20, [x16, #0xf0]\n" + ".inst 0x4fa0eaaa // sdot v10.4s, v21.16b, v0.4b[3]\n" "add x16, x16, #0x100\n" - ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" - ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" - ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" + ".inst 0x4fa1eaae // sdot v14.4s, v21.16b, v1.4b[3]\n" + ".inst 0x4fa2eab2 // sdot v18.4s, v21.16b, v2.4b[3]\n" + ".inst 0x4fa0ea8b // sdot v11.4s, v20.16b, v0.4b[3]\n" + ".inst 0x4fa1ea8f // sdot v15.4s, v20.16b, v1.4b[3]\n" + ".inst 0x4fa2ea93 // sdot v19.4s, v20.16b, v2.4b[3]\n" "87:" // Height 3: Multiply loop: Main loop skip "cbz x14, 92f\n" "cmp x14, #0x4\n" "blt 89f\n" "88:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x13], #0x4\n" + "ldr s24, [x13], #0x4\n" "sub x14, x14, #0x4\n" - "ldr s1, [x9], #0x4\n" + "ldr s23, [x12], #0x4\n" "cmp x14, #0x4\n" - "ldr s2, [x27], #0x4\n" - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x16, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr s22, [x11], #0x4\n" + "ldr q21, [x16, #0x0]\n" + ".inst 0x4f98e2a8 // sdot v8.4s, v21.16b, v24.4b[0]\n" + "ldr q20, [x16, #0x10]\n" + ".inst 0x4f97e2ac // sdot v12.4s, v21.16b, v23.4b[0]\n" + ".inst 0x4f96e2b0 // sdot v16.4s, v21.16b, v22.4b[0]\n" + "ldr q21, [x16, #0x20]\n" + ".inst 0x4f98e289 // sdot v9.4s, v20.16b, v24.4b[0]\n" + ".inst 0x4f97e28d // sdot v13.4s, v20.16b, v23.4b[0]\n" + ".inst 0x4f96e291 // sdot v17.4s, v20.16b, v22.4b[0]\n" + "ldr q20, [x16, #0x30]\n" + ".inst 0x4f98e2aa // sdot v10.4s, v21.16b, v24.4b[0]\n" "add x16, x16, #0x40\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" + ".inst 0x4f97e2ae // sdot v14.4s, v21.16b, v23.4b[0]\n" + ".inst 0x4f96e2b2 // sdot v18.4s, v21.16b, v22.4b[0]\n" + ".inst 0x4f98e28b // sdot v11.4s, v20.16b, v24.4b[0]\n" + ".inst 0x4f97e28f // sdot v15.4s, v20.16b, v23.4b[0]\n" + ".inst 0x4f96e293 // sdot v19.4s, v20.16b, v22.4b[0]\n" "bge 88b\n" "89:" // Height 3: Multiply loop: Skip odd blocks "cbz x14, 92f\n" "tbz x14, #1, 90f\n" "ldr h0, [x13], #0x2\n" - "ldr h1, [x9], #0x2\n" - "ldr h2, [x27], #0x2\n" + "ldr h1, [x12], #0x2\n" + "ldr h2, [x11], #0x2\n" "tbz x14, #0, 91f\n" "ld1 { v0.b }[2], [x13]\n" - "ld1 { v1.b }[2], [x9]\n" - "ld1 { v2.b }[2], [x27]\n" + "ld1 { v1.b }[2], [x12]\n" + "ld1 { v2.b }[2], [x11]\n" "b 91f\n" "90:" // Height 3: Multiply loop: Ragged operand read: partial_1_0 "ldr b0, [x13, #0x0]\n" - "ldr b1, [x9, #0x0]\n" - "ldr b2, [x27, #0x0]\n" + "ldr b1, [x12, #0x0]\n" + "ldr b2, [x11, #0x0]\n" "91:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x16, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr q21, [x16, #0x0]\n" + ".inst 0x4f80e2a8 // sdot v8.4s, v21.16b, v0.4b[0]\n" + "ldr q20, [x16, #0x10]\n" + ".inst 0x4f81e2ac // sdot v12.4s, v21.16b, v1.4b[0]\n" + ".inst 0x4f82e2b0 // sdot v16.4s, v21.16b, v2.4b[0]\n" + "ldr q21, [x16, #0x20]\n" + ".inst 0x4f80e289 // sdot v9.4s, v20.16b, v0.4b[0]\n" + ".inst 0x4f81e28d // sdot v13.4s, v20.16b, v1.4b[0]\n" + ".inst 0x4f82e291 // sdot v17.4s, v20.16b, v2.4b[0]\n" + "ldr q20, [x16, #0x30]\n" + ".inst 0x4f80e2aa // sdot v10.4s, v21.16b, v0.4b[0]\n" "add x16, x16, #0x40\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" + ".inst 0x4f81e2ae // sdot v14.4s, v21.16b, v1.4b[0]\n" + ".inst 0x4f82e2b2 // sdot v18.4s, v21.16b, v2.4b[0]\n" + ".inst 0x4f80e28b // sdot v11.4s, v20.16b, v0.4b[0]\n" + ".inst 0x4f81e28f // sdot v15.4s, v20.16b, v1.4b[0]\n" + ".inst 0x4f82e293 // sdot v19.4s, v20.16b, v2.4b[0]\n" "92:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x15, x15, #0x1\n" @@ -1475,336 +1474,336 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "116:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w14, [x20, x15, 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, 117f\n" - "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x13, [x21, #0x0]\n" - "ldr x9, [x21, #0x8]\n" - "ldr x27, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x13, [x20, #0x0]\n" + "ldr x12, [x20, #0x8]\n" + "ldr x11, [x20, #0x10]\n" + "ldr x10, [x20, #0x18]\n" "cbnz x15, 118f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x13, x13, x20\n" - "add x9, x9, x20\n" - "add x27, x27, x20\n" - "add x25, x25, x20\n" + "add x12, x12, x20\n" + "add x11, x11, x20\n" + "add x10, x10, x20\n" "b 118f\n" "117:" // Height 4: setup direct input "mov x13, %x[input_ptr]\n" - "add x9, x13, x20\n" - "add x27, x9, x20\n" - "add x25, x27, x20\n" + "add x12, x13, x21\n" + "add x11, x12, x21\n" + "add x10, x11, x21\n" "118:" // Height 4: input setup done "cmp x14, #0x10\n" "blt 121f\n" "ldr q0, [x13, #0x0]\n" "cmp x14, #0x20\n" - "ldr q1, [x9, #0x0]\n" - "ldr q2, [x27, #0x0]\n" - "ldr q3, [x25, #0x0]\n" + "ldr q1, [x12, #0x0]\n" + "ldr q2, [x11, #0x0]\n" + "ldr q3, [x10, #0x0]\n" "ldr q6, [x16, #0x0]\n" "ldr q7, [x16, #0x10]\n" "blt 120f\n" "119:" // Height 4: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr x12, [x16, #0x28]\n" + "ldr x21, [x16, #0x28]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr x11, [x16, #0x38]\n" + "ldr x20, [x16, #0x38]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" "add x13, x13, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "ldr d6, [x16, #0x20]\n" + "ldr d25, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "mov v6.d[1], x12\n" + "mov v25.d[1], x21\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr x12, [x16, #0x48]\n" + "ldr x21, [x16, #0x48]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "add x9, x9, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr d7, [x16, #0x30]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "ldr x11, [x16, #0x58]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "add x27, x27, #0x10\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - "ldr d6, [x16, #0x40]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr x12, [x16, #0x68]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "add x25, x25, #0x10\n" - ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - "ldr d7, [x16, #0x50]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "ldr x11, [x16, #0x78]\n" - ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - "ldr x10, [x13, #0x8]\n" - ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" - "ldr d6, [x16, #0x60]\n" - ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "mov v6.d[1], x12\n" - ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "ldr x12, [x16, #0x88]\n" - ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - "ldr x28, [x9, #0x8]\n" - ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" - "ldr d7, [x16, #0x70]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "ldr x11, [x16, #0x98]\n" - ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" - "ldr x26, [x27, #0x8]\n" - ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" - "ldr d6, [x16, #0x80]\n" - ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "mov v6.d[1], x12\n" - ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "ldr x12, [x16, #0xa8]\n" - ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - "ldr x24, [x25, #0x8]\n" - ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" - "ldr d7, [x16, #0x90]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "ldr x11, [x16, #0xb8]\n" - ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" + "ldr d24, [x16, #0x30]\n" + "mov v24.d[1], x20\n" + ".inst 0x4f80e32a // sdot v10.4s, v25.16b, v0.4b[0]\n" + ".inst 0x4f81e32e // sdot v14.4s, v25.16b, v1.4b[0]\n" + "ldr x20, [x16, #0x58]\n" + ".inst 0x4f82e332 // sdot v18.4s, v25.16b, v2.4b[0]\n" + "add x11, x11, #0x10\n" + ".inst 0x4f83e336 // sdot v22.4s, v25.16b, v3.4b[0]\n" + "ldr d25, [x16, #0x40]\n" + ".inst 0x4f80e30b // sdot v11.4s, v24.16b, v0.4b[0]\n" + "mov v25.d[1], x21\n" + ".inst 0x4f81e30f // sdot v15.4s, v24.16b, v1.4b[0]\n" + "ldr x21, [x16, #0x68]\n" + ".inst 0x4f82e313 // sdot v19.4s, v24.16b, v2.4b[0]\n" + "add x10, x10, #0x10\n" + ".inst 0x4f83e317 // sdot v23.4s, v24.16b, v3.4b[0]\n" + "ldr d24, [x16, #0x50]\n" + "mov v24.d[1], x20\n" + ".inst 0x4fa0e328 // sdot v8.4s, v25.16b, v0.4b[1]\n" + ".inst 0x4fa1e32c // sdot v12.4s, v25.16b, v1.4b[1]\n" + "ldr x20, [x16, #0x78]\n" + ".inst 0x4fa2e330 // sdot v16.4s, v25.16b, v2.4b[1]\n" + "ldr x25, [x13, #0x8]\n" + ".inst 0x4fa3e334 // sdot v20.4s, v25.16b, v3.4b[1]\n" + "ldr d25, [x16, #0x60]\n" + ".inst 0x4fa0e309 // sdot v9.4s, v24.16b, v0.4b[1]\n" + "mov v25.d[1], x21\n" + ".inst 0x4fa1e30d // sdot v13.4s, v24.16b, v1.4b[1]\n" + "ldr x21, [x16, #0x88]\n" + ".inst 0x4fa2e311 // sdot v17.4s, v24.16b, v2.4b[1]\n" + "ldr x24, [x12, #0x8]\n" + ".inst 0x4fa3e315 // sdot v21.4s, v24.16b, v3.4b[1]\n" + "ldr d24, [x16, #0x70]\n" + "mov v24.d[1], x20\n" + ".inst 0x4fa0e32a // sdot v10.4s, v25.16b, v0.4b[1]\n" + ".inst 0x4fa1e32e // sdot v14.4s, v25.16b, v1.4b[1]\n" + "ldr x20, [x16, #0x98]\n" + ".inst 0x4fa2e332 // sdot v18.4s, v25.16b, v2.4b[1]\n" + "ldr x23, [x11, #0x8]\n" + ".inst 0x4fa3e336 // sdot v22.4s, v25.16b, v3.4b[1]\n" + "ldr d25, [x16, #0x80]\n" + ".inst 0x4fa0e30b // sdot v11.4s, v24.16b, v0.4b[1]\n" + "mov v25.d[1], x21\n" + ".inst 0x4fa1e30f // sdot v15.4s, v24.16b, v1.4b[1]\n" + "ldr x21, [x16, #0xa8]\n" + ".inst 0x4fa2e313 // sdot v19.4s, v24.16b, v2.4b[1]\n" + "ldr x22, [x10, #0x8]\n" + ".inst 0x4fa3e317 // sdot v23.4s, v24.16b, v3.4b[1]\n" + "ldr d24, [x16, #0x90]\n" + "mov v24.d[1], x20\n" + ".inst 0x4f80eb28 // sdot v8.4s, v25.16b, v0.4b[2]\n" + ".inst 0x4f81eb2c // sdot v12.4s, v25.16b, v1.4b[2]\n" + "ldr x20, [x16, #0xb8]\n" + ".inst 0x4f82eb30 // sdot v16.4s, v25.16b, v2.4b[2]\n" "sub x14, x14, #0x10\n" - ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" - "ldr d6, [x16, #0xa0]\n" - ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "ldr x12, [x16, #0xc8]\n" - ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" + ".inst 0x4f83eb34 // sdot v20.4s, v25.16b, v3.4b[2]\n" + "ldr d25, [x16, #0xa0]\n" + ".inst 0x4f80eb09 // sdot v9.4s, v24.16b, v0.4b[2]\n" + "mov v25.d[1], x21\n" + ".inst 0x4f81eb0d // sdot v13.4s, v24.16b, v1.4b[2]\n" + "ldr x21, [x16, #0xc8]\n" + ".inst 0x4f82eb11 // sdot v17.4s, v24.16b, v2.4b[2]\n" "cmp x14, #0x20\n" - ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" - "ldr d7, [x16, #0xb0]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "ldr x11, [x16, #0xd8]\n" - ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" + ".inst 0x4f83eb15 // sdot v21.4s, v24.16b, v3.4b[2]\n" + "ldr d24, [x16, #0xb0]\n" + "mov v24.d[1], x20\n" + ".inst 0x4f80eb2a // sdot v10.4s, v25.16b, v0.4b[2]\n" + ".inst 0x4f81eb2e // sdot v14.4s, v25.16b, v1.4b[2]\n" + "ldr x20, [x16, #0xd8]\n" + ".inst 0x4f82eb32 // sdot v18.4s, v25.16b, v2.4b[2]\n" "prfm pldl1keep, [x13, #0x80]\n" - ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" - "ldr d6, [x16, #0xc0]\n" - ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "ldr x12, [x16, #0xe8]\n" - ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - "prfm pldl1keep, [x9, #0x80]\n" - ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" - "ldr d7, [x16, #0xd0]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "ldr x11, [x16, #0xf8]\n" - ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" - "prfm pldl1keep, [x27, #0x80]\n" - ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" - "ldr d6, [x16, #0xe0]\n" - ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "mov v6.d[1], x12\n" - ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" - "ldr d7, [x16, #0xf0]\n" - "mov v7.d[1], x11\n" + ".inst 0x4f83eb36 // sdot v22.4s, v25.16b, v3.4b[2]\n" + "ldr d25, [x16, #0xc0]\n" + ".inst 0x4f80eb0b // sdot v11.4s, v24.16b, v0.4b[2]\n" + "mov v25.d[1], x21\n" + ".inst 0x4f81eb0f // sdot v15.4s, v24.16b, v1.4b[2]\n" + "ldr x21, [x16, #0xe8]\n" + ".inst 0x4f82eb13 // sdot v19.4s, v24.16b, v2.4b[2]\n" + "prfm pldl1keep, [x12, #0x80]\n" + ".inst 0x4f83eb17 // sdot v23.4s, v24.16b, v3.4b[2]\n" + "ldr d24, [x16, #0xd0]\n" + "mov v24.d[1], x20\n" + ".inst 0x4fa0eb28 // sdot v8.4s, v25.16b, v0.4b[3]\n" + ".inst 0x4fa1eb2c // sdot v12.4s, v25.16b, v1.4b[3]\n" + "ldr x20, [x16, #0xf8]\n" + ".inst 0x4fa2eb30 // sdot v16.4s, v25.16b, v2.4b[3]\n" + "prfm pldl1keep, [x11, #0x80]\n" + ".inst 0x4fa3eb34 // sdot v20.4s, v25.16b, v3.4b[3]\n" + "ldr d25, [x16, #0xe0]\n" + ".inst 0x4fa0eb09 // sdot v9.4s, v24.16b, v0.4b[3]\n" + "mov v25.d[1], x21\n" + ".inst 0x4fa1eb0d // sdot v13.4s, v24.16b, v1.4b[3]\n" + "prfm pldl1keep, [x10, #0x80]\n" + ".inst 0x4fa2eb11 // sdot v17.4s, v24.16b, v2.4b[3]\n" + ".inst 0x4fa3eb15 // sdot v21.4s, v24.16b, v3.4b[3]\n" + "ldr d24, [x16, #0xf0]\n" + "mov v24.d[1], x20\n" "add x16, x16, #0x100\n" - ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "ldr x12, [x16, #0x8]\n" - ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - "ldr x11, [x16, #0x18]\n" - ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" - ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" + ".inst 0x4fa0eb2a // sdot v10.4s, v25.16b, v0.4b[3]\n" + "ldr x21, [x16, #0x8]\n" + ".inst 0x4fa1eb2e // sdot v14.4s, v25.16b, v1.4b[3]\n" + "ldr x20, [x16, #0x18]\n" + ".inst 0x4fa2eb32 // sdot v18.4s, v25.16b, v2.4b[3]\n" + ".inst 0x4fa3eb36 // sdot v22.4s, v25.16b, v3.4b[3]\n" "ldr d6, [x16, #0x0]\n" - ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + ".inst 0x4fa0eb0b // sdot v11.4s, v24.16b, v0.4b[3]\n" "ldr d0, [x13, #0x0]\n" - ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "ldr d1, [x9, #0x0]\n" - ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" - "ldr d2, [x27, #0x0]\n" - ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" - "ldr d3, [x25, #0x0]\n" + ".inst 0x4fa1eb0f // sdot v15.4s, v24.16b, v1.4b[3]\n" + "ldr d1, [x12, #0x0]\n" + ".inst 0x4fa2eb13 // sdot v19.4s, v24.16b, v2.4b[3]\n" + "ldr d2, [x11, #0x0]\n" + ".inst 0x4fa3eb17 // sdot v23.4s, v24.16b, v3.4b[3]\n" + "ldr d3, [x10, #0x0]\n" "ldr d7, [x16, #0x10]\n" - "mov v6.d[1], x12\n" - "mov v0.d[1], x10\n" - "mov v1.d[1], x28\n" - "mov v2.d[1], x26\n" - "mov v3.d[1], x24\n" - "mov v7.d[1], x11\n" + "mov v6.d[1], x21\n" + "mov v0.d[1], x25\n" + "mov v1.d[1], x24\n" + "mov v2.d[1], x23\n" + "mov v3.d[1], x22\n" + "mov v7.d[1], x20\n" "bge 119b\n" "120:" // Height 4: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" "add x13, x13, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x9, x9, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "add x27, x27, #0x10\n" + "add x11, x11, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x16, #0x20]\n" + "ldr q25, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x25, x25, #0x10\n" + "add x10, x10, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" "sub x14, x14, #0x10\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" "prfm pldl1keep, [x13, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "prfm pldl1keep, [x27, #0x80]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x16, #0x40]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x16, #0x50]\n" - ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" - "ldr q6, [x16, #0x60]\n" - ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" - "ldr q7, [x16, #0x70]\n" - ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" - ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" - "ldr q6, [x16, #0x80]\n" - ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" - "ldr q7, [x16, #0x90]\n" - ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" - ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" - "ldr q6, [x16, #0xa0]\n" - ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" - ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" - "ldr q7, [x16, #0xb0]\n" - ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" - ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" - "ldr q6, [x16, #0xc0]\n" - ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" - "ldr q7, [x16, #0xd0]\n" - ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" - ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" - "ldr q6, [x16, #0xe0]\n" - ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" - "ldr q7, [x16, #0xf0]\n" - ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "ldr q24, [x16, #0x30]\n" + ".inst 0x4f80e32a // sdot v10.4s, v25.16b, v0.4b[0]\n" + "prfm pldl1keep, [x12, #0x80]\n" + ".inst 0x4f81e32e // sdot v14.4s, v25.16b, v1.4b[0]\n" + "prfm pldl1keep, [x11, #0x80]\n" + ".inst 0x4f82e332 // sdot v18.4s, v25.16b, v2.4b[0]\n" + "prfm pldl1keep, [x10, #0x80]\n" + ".inst 0x4f83e336 // sdot v22.4s, v25.16b, v3.4b[0]\n" + "ldr q25, [x16, #0x40]\n" + ".inst 0x4f80e30b // sdot v11.4s, v24.16b, v0.4b[0]\n" + ".inst 0x4f81e30f // sdot v15.4s, v24.16b, v1.4b[0]\n" + ".inst 0x4f82e313 // sdot v19.4s, v24.16b, v2.4b[0]\n" + ".inst 0x4f83e317 // sdot v23.4s, v24.16b, v3.4b[0]\n" + "ldr q24, [x16, #0x50]\n" + ".inst 0x4fa0e328 // sdot v8.4s, v25.16b, v0.4b[1]\n" + ".inst 0x4fa1e32c // sdot v12.4s, v25.16b, v1.4b[1]\n" + ".inst 0x4fa2e330 // sdot v16.4s, v25.16b, v2.4b[1]\n" + ".inst 0x4fa3e334 // sdot v20.4s, v25.16b, v3.4b[1]\n" + "ldr q25, [x16, #0x60]\n" + ".inst 0x4fa0e309 // sdot v9.4s, v24.16b, v0.4b[1]\n" + ".inst 0x4fa1e30d // sdot v13.4s, v24.16b, v1.4b[1]\n" + ".inst 0x4fa2e311 // sdot v17.4s, v24.16b, v2.4b[1]\n" + ".inst 0x4fa3e315 // sdot v21.4s, v24.16b, v3.4b[1]\n" + "ldr q24, [x16, #0x70]\n" + ".inst 0x4fa0e32a // sdot v10.4s, v25.16b, v0.4b[1]\n" + ".inst 0x4fa1e32e // sdot v14.4s, v25.16b, v1.4b[1]\n" + ".inst 0x4fa2e332 // sdot v18.4s, v25.16b, v2.4b[1]\n" + ".inst 0x4fa3e336 // sdot v22.4s, v25.16b, v3.4b[1]\n" + "ldr q25, [x16, #0x80]\n" + ".inst 0x4fa0e30b // sdot v11.4s, v24.16b, v0.4b[1]\n" + ".inst 0x4fa1e30f // sdot v15.4s, v24.16b, v1.4b[1]\n" + ".inst 0x4fa2e313 // sdot v19.4s, v24.16b, v2.4b[1]\n" + ".inst 0x4fa3e317 // sdot v23.4s, v24.16b, v3.4b[1]\n" + "ldr q24, [x16, #0x90]\n" + ".inst 0x4f80eb28 // sdot v8.4s, v25.16b, v0.4b[2]\n" + ".inst 0x4f81eb2c // sdot v12.4s, v25.16b, v1.4b[2]\n" + ".inst 0x4f82eb30 // sdot v16.4s, v25.16b, v2.4b[2]\n" + ".inst 0x4f83eb34 // sdot v20.4s, v25.16b, v3.4b[2]\n" + "ldr q25, [x16, #0xa0]\n" + ".inst 0x4f80eb09 // sdot v9.4s, v24.16b, v0.4b[2]\n" + ".inst 0x4f81eb0d // sdot v13.4s, v24.16b, v1.4b[2]\n" + ".inst 0x4f82eb11 // sdot v17.4s, v24.16b, v2.4b[2]\n" + ".inst 0x4f83eb15 // sdot v21.4s, v24.16b, v3.4b[2]\n" + "ldr q24, [x16, #0xb0]\n" + ".inst 0x4f80eb2a // sdot v10.4s, v25.16b, v0.4b[2]\n" + ".inst 0x4f81eb2e // sdot v14.4s, v25.16b, v1.4b[2]\n" + ".inst 0x4f82eb32 // sdot v18.4s, v25.16b, v2.4b[2]\n" + ".inst 0x4f83eb36 // sdot v22.4s, v25.16b, v3.4b[2]\n" + "ldr q25, [x16, #0xc0]\n" + ".inst 0x4f80eb0b // sdot v11.4s, v24.16b, v0.4b[2]\n" + ".inst 0x4f81eb0f // sdot v15.4s, v24.16b, v1.4b[2]\n" + ".inst 0x4f82eb13 // sdot v19.4s, v24.16b, v2.4b[2]\n" + ".inst 0x4f83eb17 // sdot v23.4s, v24.16b, v3.4b[2]\n" + "ldr q24, [x16, #0xd0]\n" + ".inst 0x4fa0eb28 // sdot v8.4s, v25.16b, v0.4b[3]\n" + ".inst 0x4fa1eb2c // sdot v12.4s, v25.16b, v1.4b[3]\n" + ".inst 0x4fa2eb30 // sdot v16.4s, v25.16b, v2.4b[3]\n" + ".inst 0x4fa3eb34 // sdot v20.4s, v25.16b, v3.4b[3]\n" + "ldr q25, [x16, #0xe0]\n" + ".inst 0x4fa0eb09 // sdot v9.4s, v24.16b, v0.4b[3]\n" + ".inst 0x4fa1eb0d // sdot v13.4s, v24.16b, v1.4b[3]\n" + ".inst 0x4fa2eb11 // sdot v17.4s, v24.16b, v2.4b[3]\n" + ".inst 0x4fa3eb15 // sdot v21.4s, v24.16b, v3.4b[3]\n" + "ldr q24, [x16, #0xf0]\n" + ".inst 0x4fa0eb2a // sdot v10.4s, v25.16b, v0.4b[3]\n" "add x16, x16, #0x100\n" - ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" - ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" - ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" - ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" - ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" + ".inst 0x4fa1eb2e // sdot v14.4s, v25.16b, v1.4b[3]\n" + ".inst 0x4fa2eb32 // sdot v18.4s, v25.16b, v2.4b[3]\n" + ".inst 0x4fa3eb36 // sdot v22.4s, v25.16b, v3.4b[3]\n" + ".inst 0x4fa0eb0b // sdot v11.4s, v24.16b, v0.4b[3]\n" + ".inst 0x4fa1eb0f // sdot v15.4s, v24.16b, v1.4b[3]\n" + ".inst 0x4fa2eb13 // sdot v19.4s, v24.16b, v2.4b[3]\n" + ".inst 0x4fa3eb17 // sdot v23.4s, v24.16b, v3.4b[3]\n" "121:" // Height 4: Multiply loop: Main loop skip "cbz x14, 126f\n" "cmp x14, #0x4\n" "blt 123f\n" "122:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x13], #0x4\n" + "ldr s29, [x13], #0x4\n" "sub x14, x14, #0x4\n" - "ldr s1, [x9], #0x4\n" + "ldr s28, [x12], #0x4\n" "cmp x14, #0x4\n" - "ldr s2, [x27], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x16, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr s27, [x11], #0x4\n" + "ldr s26, [x10], #0x4\n" + "ldr q25, [x16, #0x0]\n" + ".inst 0x4f9de328 // sdot v8.4s, v25.16b, v29.4b[0]\n" + "ldr q24, [x16, #0x10]\n" + ".inst 0x4f9ce32c // sdot v12.4s, v25.16b, v28.4b[0]\n" + ".inst 0x4f9be330 // sdot v16.4s, v25.16b, v27.4b[0]\n" + ".inst 0x4f9ae334 // sdot v20.4s, v25.16b, v26.4b[0]\n" + "ldr q25, [x16, #0x20]\n" + ".inst 0x4f9de309 // sdot v9.4s, v24.16b, v29.4b[0]\n" + ".inst 0x4f9ce30d // sdot v13.4s, v24.16b, v28.4b[0]\n" + ".inst 0x4f9be311 // sdot v17.4s, v24.16b, v27.4b[0]\n" + ".inst 0x4f9ae315 // sdot v21.4s, v24.16b, v26.4b[0]\n" + "ldr q24, [x16, #0x30]\n" + ".inst 0x4f9de32a // sdot v10.4s, v25.16b, v29.4b[0]\n" "add x16, x16, #0x40\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" + ".inst 0x4f9ce32e // sdot v14.4s, v25.16b, v28.4b[0]\n" + ".inst 0x4f9be332 // sdot v18.4s, v25.16b, v27.4b[0]\n" + ".inst 0x4f9ae336 // sdot v22.4s, v25.16b, v26.4b[0]\n" + ".inst 0x4f9de30b // sdot v11.4s, v24.16b, v29.4b[0]\n" + ".inst 0x4f9ce30f // sdot v15.4s, v24.16b, v28.4b[0]\n" + ".inst 0x4f9be313 // sdot v19.4s, v24.16b, v27.4b[0]\n" + ".inst 0x4f9ae317 // sdot v23.4s, v24.16b, v26.4b[0]\n" "bge 122b\n" "123:" // Height 4: Multiply loop: Skip odd blocks "cbz x14, 126f\n" "tbz x14, #1, 124f\n" "ldr h0, [x13], #0x2\n" - "ldr h1, [x9], #0x2\n" - "ldr h2, [x27], #0x2\n" - "ldr h3, [x25], #0x2\n" + "ldr h1, [x12], #0x2\n" + "ldr h2, [x11], #0x2\n" + "ldr h3, [x10], #0x2\n" "tbz x14, #0, 125f\n" "ld1 { v0.b }[2], [x13]\n" - "ld1 { v1.b }[2], [x9]\n" - "ld1 { v2.b }[2], [x27]\n" - "ld1 { v3.b }[2], [x25]\n" + "ld1 { v1.b }[2], [x12]\n" + "ld1 { v2.b }[2], [x11]\n" + "ld1 { v3.b }[2], [x10]\n" "b 125f\n" "124:" // Height 4: Multiply loop: Ragged operand read: partial_1_0 "ldr b0, [x13, #0x0]\n" - "ldr b1, [x9, #0x0]\n" - "ldr b2, [x27, #0x0]\n" - "ldr b3, [x25, #0x0]\n" + "ldr b1, [x12, #0x0]\n" + "ldr b2, [x11, #0x0]\n" + "ldr b3, [x10, #0x0]\n" "125:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x16, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr q25, [x16, #0x0]\n" + ".inst 0x4f80e328 // sdot v8.4s, v25.16b, v0.4b[0]\n" + "ldr q24, [x16, #0x10]\n" + ".inst 0x4f81e32c // sdot v12.4s, v25.16b, v1.4b[0]\n" + ".inst 0x4f82e330 // sdot v16.4s, v25.16b, v2.4b[0]\n" + ".inst 0x4f83e334 // sdot v20.4s, v25.16b, v3.4b[0]\n" + "ldr q25, [x16, #0x20]\n" + ".inst 0x4f80e309 // sdot v9.4s, v24.16b, v0.4b[0]\n" + ".inst 0x4f81e30d // sdot v13.4s, v24.16b, v1.4b[0]\n" + ".inst 0x4f82e311 // sdot v17.4s, v24.16b, v2.4b[0]\n" + ".inst 0x4f83e315 // sdot v21.4s, v24.16b, v3.4b[0]\n" + "ldr q24, [x16, #0x30]\n" + ".inst 0x4f80e32a // sdot v10.4s, v25.16b, v0.4b[0]\n" "add x16, x16, #0x40\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" + ".inst 0x4f81e32e // sdot v14.4s, v25.16b, v1.4b[0]\n" + ".inst 0x4f82e332 // sdot v18.4s, v25.16b, v2.4b[0]\n" + ".inst 0x4f83e336 // sdot v22.4s, v25.16b, v3.4b[0]\n" + ".inst 0x4f80e30b // sdot v11.4s, v24.16b, v0.4b[0]\n" + ".inst 0x4f81e30f // sdot v15.4s, v24.16b, v1.4b[0]\n" + ".inst 0x4f82e313 // sdot v19.4s, v24.16b, v2.4b[0]\n" + ".inst 0x4f83e317 // sdot v23.4s, v24.16b, v3.4b[0]\n" "126:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x15, x15, #0x1\n" @@ -2108,399 +2107,399 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "movi v22.4s, #0x0\n" "movi v23.4s, #0x0\n" "movi v24.4s, #0x0\n" - "movi v25.4s, #0x0\n" - "movi v26.4s, #0x0\n" - "movi v27.4s, #0x0\n" - "149:" // Height 5: setup done - "mov x15, #0x0\n" - "150:" // Height 5: String loop - "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w14, [x20, x15, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" - "tbz %x[flags], #3, 151f\n" - "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x13, [x21, #0x0]\n" - "ldr x9, [x21, #0x8]\n" - "ldr x27, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" - "ldr x23, [x21, #0x20]\n" - "cbnz x15, 152f\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x13, x13, x20\n" - "add x9, x9, x20\n" - "add x27, x27, x20\n" - "add x25, x25, x20\n" - "add x23, x23, x20\n" - "b 152f\n" - "151:" // Height 5: setup direct input - "mov x13, %x[input_ptr]\n" - "add x9, x13, x20\n" - "add x27, x9, x20\n" - "add x25, x27, x20\n" - "add x23, x25, x20\n" - "152:" // Height 5: input setup done - "cmp x14, #0x10\n" - "blt 155f\n" - "ldr q0, [x13, #0x0]\n" - "cmp x14, #0x20\n" - "ldr q1, [x9, #0x0]\n" - "ldr q2, [x27, #0x0]\n" - "ldr q3, [x25, #0x0]\n" - "ldr q4, [x23, #0x0]\n" - "ldr q6, [x16, #0x0]\n" - "ldr q7, [x16, #0x10]\n" - "blt 154f\n" - "153:" // Height 5: Multiply loop: Main loop head - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr x12, [x16, #0x28]\n" - ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr x11, [x16, #0x38]\n" - ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "add x13, x13, #0x10\n" - ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "add x9, x9, #0x10\n" - ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "ldr d6, [x16, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr x12, [x16, #0x48]\n" - ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "add x27, x27, #0x10\n" - ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "add x25, x25, #0x10\n" - ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr d7, [x16, #0x30]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "ldr x11, [x16, #0x58]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "add x23, x23, #0x10\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - "ldr x10, [x13, #0x8]\n" - ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - "ldr d6, [x16, #0x40]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr x12, [x16, #0x68]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "ldr x28, [x9, #0x8]\n" - ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - "ldr x26, [x27, #0x8]\n" - ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" - "ldr d7, [x16, #0x50]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "ldr x11, [x16, #0x78]\n" - ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - "ldr x24, [x25, #0x8]\n" - ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" - "ldr x22, [x23, #0x8]\n" - ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" - "ldr d6, [x16, #0x60]\n" - ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "mov v6.d[1], x12\n" - ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "ldr x12, [x16, #0x88]\n" - ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - "sub x14, x14, #0x10\n" - ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" - "cmp x14, #0x20\n" - ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" - "ldr d7, [x16, #0x70]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "ldr x11, [x16, #0x98]\n" - ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" - "prfm pldl1keep, [x13, #0x80]\n" - ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" - "prfm pldl1keep, [x9, #0x80]\n" - ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" - "ldr d6, [x16, #0x80]\n" - ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "mov v6.d[1], x12\n" - ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "ldr x12, [x16, #0xa8]\n" - ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - "prfm pldl1keep, [x27, #0x80]\n" - ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" - "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" - "ldr d7, [x16, #0x90]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "ldr x11, [x16, #0xb8]\n" - ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" - "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" - ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" - "ldr d6, [x16, #0xa0]\n" - ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "ldr x12, [x16, #0xc8]\n" - ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" - ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" - ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" - "ldr d7, [x16, #0xb0]\n" - "mov v7.d[1], x11\n" - ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "ldr x11, [x16, #0xd8]\n" - ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" - ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" - ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" - "ldr d6, [x16, #0xc0]\n" - ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "mov v6.d[1], x12\n" - ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "ldr x12, [x16, #0xe8]\n" - ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" - ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" - "ldr d7, [x16, #0xd0]\n" - "mov v7.d[1], x11\n" - ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "ldr x11, [x16, #0xf8]\n" - ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" - ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" - ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" - "ldr d6, [x16, #0xe0]\n" - ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "mov v6.d[1], x12\n" - ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" - ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" - "ldr d7, [x16, #0xf0]\n" - "mov v7.d[1], x11\n" + "movi v25.4s, #0x0\n" + "movi v26.4s, #0x0\n" + "movi v27.4s, #0x0\n" + "149:" // Height 5: setup done + "mov x15, #0x0\n" + "150:" // Height 5: String loop + "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr w14, [x20, x15, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "tbz %x[flags], #3, 151f\n" + "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x13, [x20, #0x0]\n" + "ldr x12, [x20, #0x8]\n" + "ldr x11, [x20, #0x10]\n" + "ldr x10, [x20, #0x18]\n" + "ldr x9, [x20, #0x20]\n" + "cbnz x15, 152f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x13, x13, x20\n" + "add x12, x12, x20\n" + "add x11, x11, x20\n" + "add x10, x10, x20\n" + "add x9, x9, x20\n" + "b 152f\n" + "151:" // Height 5: setup direct input + "mov x13, %x[input_ptr]\n" + "add x12, x13, x21\n" + "add x11, x12, x21\n" + "add x10, x11, x21\n" + "add x9, x10, x21\n" + "152:" // Height 5: input setup done + "cmp x14, #0x10\n" + "blt 155f\n" + "ldr q0, [x13, #0x0]\n" + "cmp x14, #0x20\n" + "ldr q1, [x12, #0x0]\n" + "ldr q2, [x11, #0x0]\n" + "ldr q3, [x10, #0x0]\n" + "ldr q4, [x9, #0x0]\n" + "ldr q6, [x16, #0x0]\n" + "ldr q7, [x16, #0x10]\n" + "blt 154f\n" + "153:" // Height 5: Multiply loop: Main loop head + ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" + "ldr x21, [x16, #0x28]\n" + ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" + "ldr x20, [x16, #0x38]\n" + ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" + "add x13, x13, #0x10\n" + ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" + "add x12, x12, #0x10\n" + ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" + "ldr d29, [x16, #0x20]\n" + ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" + "mov v29.d[1], x21\n" + ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" + "ldr x21, [x16, #0x48]\n" + ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" + "add x11, x11, #0x10\n" + ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" + "add x10, x10, #0x10\n" + ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" + "ldr d28, [x16, #0x30]\n" + "mov v28.d[1], x20\n" + ".inst 0x4f80e3aa // sdot v10.4s, v29.16b, v0.4b[0]\n" + ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n" + "ldr x20, [x16, #0x58]\n" + ".inst 0x4f82e3b2 // sdot v18.4s, v29.16b, v2.4b[0]\n" + "add x9, x9, #0x10\n" + ".inst 0x4f83e3b6 // sdot v22.4s, v29.16b, v3.4b[0]\n" + "ldr x26, [x13, #0x8]\n" + ".inst 0x4f84e3ba // sdot v26.4s, v29.16b, v4.4b[0]\n" + "ldr d29, [x16, #0x40]\n" + ".inst 0x4f80e38b // sdot v11.4s, v28.16b, v0.4b[0]\n" + "mov v29.d[1], x21\n" + ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n" + "ldr x21, [x16, #0x68]\n" + ".inst 0x4f82e393 // sdot v19.4s, v28.16b, v2.4b[0]\n" + "ldr x25, [x12, #0x8]\n" + ".inst 0x4f83e397 // sdot v23.4s, v28.16b, v3.4b[0]\n" + "ldr x24, [x11, #0x8]\n" + ".inst 0x4f84e39b // sdot v27.4s, v28.16b, v4.4b[0]\n" + "ldr d28, [x16, #0x50]\n" + "mov v28.d[1], x20\n" + ".inst 0x4fa0e3a8 // sdot v8.4s, v29.16b, v0.4b[1]\n" + ".inst 0x4fa1e3ac // sdot v12.4s, v29.16b, v1.4b[1]\n" + "ldr x20, [x16, #0x78]\n" + ".inst 0x4fa2e3b0 // sdot v16.4s, v29.16b, v2.4b[1]\n" + "ldr x23, [x10, #0x8]\n" + ".inst 0x4fa3e3b4 // sdot v20.4s, v29.16b, v3.4b[1]\n" + "ldr x22, [x9, #0x8]\n" + ".inst 0x4fa4e3b8 // sdot v24.4s, v29.16b, v4.4b[1]\n" + "ldr d29, [x16, #0x60]\n" + ".inst 0x4fa0e389 // sdot v9.4s, v28.16b, v0.4b[1]\n" + "mov v29.d[1], x21\n" + ".inst 0x4fa1e38d // sdot v13.4s, v28.16b, v1.4b[1]\n" + "ldr x21, [x16, #0x88]\n" + ".inst 0x4fa2e391 // sdot v17.4s, v28.16b, v2.4b[1]\n" + "sub x14, x14, #0x10\n" + ".inst 0x4fa3e395 // sdot v21.4s, v28.16b, v3.4b[1]\n" + "cmp x14, #0x20\n" + ".inst 0x4fa4e399 // sdot v25.4s, v28.16b, v4.4b[1]\n" + "ldr d28, [x16, #0x70]\n" + "mov v28.d[1], x20\n" + ".inst 0x4fa0e3aa // sdot v10.4s, v29.16b, v0.4b[1]\n" + ".inst 0x4fa1e3ae // sdot v14.4s, v29.16b, v1.4b[1]\n" + "ldr x20, [x16, #0x98]\n" + ".inst 0x4fa2e3b2 // sdot v18.4s, v29.16b, v2.4b[1]\n" + "prfm pldl1keep, [x13, #0x80]\n" + ".inst 0x4fa3e3b6 // sdot v22.4s, v29.16b, v3.4b[1]\n" + "prfm pldl1keep, [x12, #0x80]\n" + ".inst 0x4fa4e3ba // sdot v26.4s, v29.16b, v4.4b[1]\n" + "ldr d29, [x16, #0x80]\n" + ".inst 0x4fa0e38b // sdot v11.4s, v28.16b, v0.4b[1]\n" + "mov v29.d[1], x21\n" + ".inst 0x4fa1e38f // sdot v15.4s, v28.16b, v1.4b[1]\n" + "ldr x21, [x16, #0xa8]\n" + ".inst 0x4fa2e393 // sdot v19.4s, v28.16b, v2.4b[1]\n" + "prfm pldl1keep, [x11, #0x80]\n" + ".inst 0x4fa3e397 // sdot v23.4s, v28.16b, v3.4b[1]\n" + "prfm pldl1keep, [x10, #0x80]\n" + ".inst 0x4fa4e39b // sdot v27.4s, v28.16b, v4.4b[1]\n" + "ldr d28, [x16, #0x90]\n" + "mov v28.d[1], x20\n" + ".inst 0x4f80eba8 // sdot v8.4s, v29.16b, v0.4b[2]\n" + ".inst 0x4f81ebac // sdot v12.4s, v29.16b, v1.4b[2]\n" + "ldr x20, [x16, #0xb8]\n" + ".inst 0x4f82ebb0 // sdot v16.4s, v29.16b, v2.4b[2]\n" + "prfm pldl1keep, [x9, #0x80]\n" + ".inst 0x4f83ebb4 // sdot v20.4s, v29.16b, v3.4b[2]\n" + ".inst 0x4f84ebb8 // sdot v24.4s, v29.16b, v4.4b[2]\n" + "ldr d29, [x16, #0xa0]\n" + ".inst 0x4f80eb89 // sdot v9.4s, v28.16b, v0.4b[2]\n" + "mov v29.d[1], x21\n" + ".inst 0x4f81eb8d // sdot v13.4s, v28.16b, v1.4b[2]\n" + "ldr x21, [x16, #0xc8]\n" + ".inst 0x4f82eb91 // sdot v17.4s, v28.16b, v2.4b[2]\n" + ".inst 0x4f83eb95 // sdot v21.4s, v28.16b, v3.4b[2]\n" + ".inst 0x4f84eb99 // sdot v25.4s, v28.16b, v4.4b[2]\n" + "ldr d28, [x16, #0xb0]\n" + "mov v28.d[1], x20\n" + ".inst 0x4f80ebaa // sdot v10.4s, v29.16b, v0.4b[2]\n" + ".inst 0x4f81ebae // sdot v14.4s, v29.16b, v1.4b[2]\n" + "ldr x20, [x16, #0xd8]\n" + ".inst 0x4f82ebb2 // sdot v18.4s, v29.16b, v2.4b[2]\n" + ".inst 0x4f83ebb6 // sdot v22.4s, v29.16b, v3.4b[2]\n" + ".inst 0x4f84ebba // sdot v26.4s, v29.16b, v4.4b[2]\n" + "ldr d29, [x16, #0xc0]\n" + ".inst 0x4f80eb8b // sdot v11.4s, v28.16b, v0.4b[2]\n" + "mov v29.d[1], x21\n" + ".inst 0x4f81eb8f // sdot v15.4s, v28.16b, v1.4b[2]\n" + "ldr x21, [x16, #0xe8]\n" + ".inst 0x4f82eb93 // sdot v19.4s, v28.16b, v2.4b[2]\n" + ".inst 0x4f83eb97 // sdot v23.4s, v28.16b, v3.4b[2]\n" + ".inst 0x4f84eb9b // sdot v27.4s, v28.16b, v4.4b[2]\n" + "ldr d28, [x16, #0xd0]\n" + "mov v28.d[1], x20\n" + ".inst 0x4fa0eba8 // sdot v8.4s, v29.16b, v0.4b[3]\n" + ".inst 0x4fa1ebac // sdot v12.4s, v29.16b, v1.4b[3]\n" + "ldr x20, [x16, #0xf8]\n" + ".inst 0x4fa2ebb0 // sdot v16.4s, v29.16b, v2.4b[3]\n" + ".inst 0x4fa3ebb4 // sdot v20.4s, v29.16b, v3.4b[3]\n" + ".inst 0x4fa4ebb8 // sdot v24.4s, v29.16b, v4.4b[3]\n" + "ldr d29, [x16, #0xe0]\n" + ".inst 0x4fa0eb89 // sdot v9.4s, v28.16b, v0.4b[3]\n" + "mov v29.d[1], x21\n" + ".inst 0x4fa1eb8d // sdot v13.4s, v28.16b, v1.4b[3]\n" + ".inst 0x4fa2eb91 // sdot v17.4s, v28.16b, v2.4b[3]\n" + ".inst 0x4fa3eb95 // sdot v21.4s, v28.16b, v3.4b[3]\n" + ".inst 0x4fa4eb99 // sdot v25.4s, v28.16b, v4.4b[3]\n" + "ldr d28, [x16, #0xf0]\n" + "mov v28.d[1], x20\n" "add x16, x16, #0x100\n" - ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "ldr x12, [x16, #0x8]\n" - ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - "ldr x11, [x16, #0x18]\n" - ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" - ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" - ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" + ".inst 0x4fa0ebaa // sdot v10.4s, v29.16b, v0.4b[3]\n" + "ldr x21, [x16, #0x8]\n" + ".inst 0x4fa1ebae // sdot v14.4s, v29.16b, v1.4b[3]\n" + "ldr x20, [x16, #0x18]\n" + ".inst 0x4fa2ebb2 // sdot v18.4s, v29.16b, v2.4b[3]\n" + ".inst 0x4fa3ebb6 // sdot v22.4s, v29.16b, v3.4b[3]\n" + ".inst 0x4fa4ebba // sdot v26.4s, v29.16b, v4.4b[3]\n" "ldr d6, [x16, #0x0]\n" - ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" + ".inst 0x4fa0eb8b // sdot v11.4s, v28.16b, v0.4b[3]\n" "ldr d0, [x13, #0x0]\n" - ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "ldr d1, [x9, #0x0]\n" - ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" - "ldr d2, [x27, #0x0]\n" - ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" - "ldr d3, [x25, #0x0]\n" - ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" - "ldr d4, [x23, #0x0]\n" + ".inst 0x4fa1eb8f // sdot v15.4s, v28.16b, v1.4b[3]\n" + "ldr d1, [x12, #0x0]\n" + ".inst 0x4fa2eb93 // sdot v19.4s, v28.16b, v2.4b[3]\n" + "ldr d2, [x11, #0x0]\n" + ".inst 0x4fa3eb97 // sdot v23.4s, v28.16b, v3.4b[3]\n" + "ldr d3, [x10, #0x0]\n" + ".inst 0x4fa4eb9b // sdot v27.4s, v28.16b, v4.4b[3]\n" + "ldr d4, [x9, #0x0]\n" "ldr d7, [x16, #0x10]\n" - "mov v6.d[1], x12\n" - "mov v0.d[1], x10\n" - "mov v1.d[1], x28\n" - "mov v2.d[1], x26\n" - "mov v3.d[1], x24\n" + "mov v6.d[1], x21\n" + "mov v0.d[1], x26\n" + "mov v1.d[1], x25\n" + "mov v2.d[1], x24\n" + "mov v3.d[1], x23\n" "mov v4.d[1], x22\n" - "mov v7.d[1], x11\n" + "mov v7.d[1], x20\n" "bge 153b\n" "154:" // Height 5: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" "add x13, x13, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x9, x9, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "add x27, x27, #0x10\n" + "add x11, x11, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "add x25, x25, #0x10\n" + "add x10, x10, #0x10\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x16, #0x20]\n" + "ldr q29, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x23, x23, #0x10\n" + "add x9, x9, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" "sub x14, x14, #0x10\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" "prfm pldl1keep, [x13, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "prfm pldl1keep, [x27, #0x80]\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x16, #0x40]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x16, #0x50]\n" - ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" - ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" - ".inst 0x4fa4e0d8 // sdot v24.4s, v6.16b, v4.4b[1]\n" - "ldr q6, [x16, #0x60]\n" - ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" - ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" - "ldr q7, [x16, #0x70]\n" - ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" - ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" - ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" - "ldr q6, [x16, #0x80]\n" - ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" - ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" - ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" - "ldr q7, [x16, #0x90]\n" - ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" - ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" - ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" - "ldr q6, [x16, #0xa0]\n" - ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" - ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" - ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" - "ldr q7, [x16, #0xb0]\n" - ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" - ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" - ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" - "ldr q6, [x16, #0xc0]\n" - ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" - ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" - ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" - "ldr q7, [x16, #0xd0]\n" - ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" - ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" - ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" - "ldr q6, [x16, #0xe0]\n" - ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" - ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" - ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" - ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" - "ldr q7, [x16, #0xf0]\n" - ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" + "ldr q28, [x16, #0x30]\n" + ".inst 0x4f80e3aa // sdot v10.4s, v29.16b, v0.4b[0]\n" + "prfm pldl1keep, [x11, #0x80]\n" + ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n" + "prfm pldl1keep, [x10, #0x80]\n" + ".inst 0x4f82e3b2 // sdot v18.4s, v29.16b, v2.4b[0]\n" + "prfm pldl1keep, [x9, #0x80]\n" + ".inst 0x4f83e3b6 // sdot v22.4s, v29.16b, v3.4b[0]\n" + ".inst 0x4f84e3ba // sdot v26.4s, v29.16b, v4.4b[0]\n" + "ldr q29, [x16, #0x40]\n" + ".inst 0x4f80e38b // sdot v11.4s, v28.16b, v0.4b[0]\n" + ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n" + ".inst 0x4f82e393 // sdot v19.4s, v28.16b, v2.4b[0]\n" + ".inst 0x4f83e397 // sdot v23.4s, v28.16b, v3.4b[0]\n" + ".inst 0x4f84e39b // sdot v27.4s, v28.16b, v4.4b[0]\n" + "ldr q28, [x16, #0x50]\n" + ".inst 0x4fa0e3a8 // sdot v8.4s, v29.16b, v0.4b[1]\n" + ".inst 0x4fa1e3ac // sdot v12.4s, v29.16b, v1.4b[1]\n" + ".inst 0x4fa2e3b0 // sdot v16.4s, v29.16b, v2.4b[1]\n" + ".inst 0x4fa3e3b4 // sdot v20.4s, v29.16b, v3.4b[1]\n" + ".inst 0x4fa4e3b8 // sdot v24.4s, v29.16b, v4.4b[1]\n" + "ldr q29, [x16, #0x60]\n" + ".inst 0x4fa0e389 // sdot v9.4s, v28.16b, v0.4b[1]\n" + ".inst 0x4fa1e38d // sdot v13.4s, v28.16b, v1.4b[1]\n" + ".inst 0x4fa2e391 // sdot v17.4s, v28.16b, v2.4b[1]\n" + ".inst 0x4fa3e395 // sdot v21.4s, v28.16b, v3.4b[1]\n" + ".inst 0x4fa4e399 // sdot v25.4s, v28.16b, v4.4b[1]\n" + "ldr q28, [x16, #0x70]\n" + ".inst 0x4fa0e3aa // sdot v10.4s, v29.16b, v0.4b[1]\n" + ".inst 0x4fa1e3ae // sdot v14.4s, v29.16b, v1.4b[1]\n" + ".inst 0x4fa2e3b2 // sdot v18.4s, v29.16b, v2.4b[1]\n" + ".inst 0x4fa3e3b6 // sdot v22.4s, v29.16b, v3.4b[1]\n" + ".inst 0x4fa4e3ba // sdot v26.4s, v29.16b, v4.4b[1]\n" + "ldr q29, [x16, #0x80]\n" + ".inst 0x4fa0e38b // sdot v11.4s, v28.16b, v0.4b[1]\n" + ".inst 0x4fa1e38f // sdot v15.4s, v28.16b, v1.4b[1]\n" + ".inst 0x4fa2e393 // sdot v19.4s, v28.16b, v2.4b[1]\n" + ".inst 0x4fa3e397 // sdot v23.4s, v28.16b, v3.4b[1]\n" + ".inst 0x4fa4e39b // sdot v27.4s, v28.16b, v4.4b[1]\n" + "ldr q28, [x16, #0x90]\n" + ".inst 0x4f80eba8 // sdot v8.4s, v29.16b, v0.4b[2]\n" + ".inst 0x4f81ebac // sdot v12.4s, v29.16b, v1.4b[2]\n" + ".inst 0x4f82ebb0 // sdot v16.4s, v29.16b, v2.4b[2]\n" + ".inst 0x4f83ebb4 // sdot v20.4s, v29.16b, v3.4b[2]\n" + ".inst 0x4f84ebb8 // sdot v24.4s, v29.16b, v4.4b[2]\n" + "ldr q29, [x16, #0xa0]\n" + ".inst 0x4f80eb89 // sdot v9.4s, v28.16b, v0.4b[2]\n" + ".inst 0x4f81eb8d // sdot v13.4s, v28.16b, v1.4b[2]\n" + ".inst 0x4f82eb91 // sdot v17.4s, v28.16b, v2.4b[2]\n" + ".inst 0x4f83eb95 // sdot v21.4s, v28.16b, v3.4b[2]\n" + ".inst 0x4f84eb99 // sdot v25.4s, v28.16b, v4.4b[2]\n" + "ldr q28, [x16, #0xb0]\n" + ".inst 0x4f80ebaa // sdot v10.4s, v29.16b, v0.4b[2]\n" + ".inst 0x4f81ebae // sdot v14.4s, v29.16b, v1.4b[2]\n" + ".inst 0x4f82ebb2 // sdot v18.4s, v29.16b, v2.4b[2]\n" + ".inst 0x4f83ebb6 // sdot v22.4s, v29.16b, v3.4b[2]\n" + ".inst 0x4f84ebba // sdot v26.4s, v29.16b, v4.4b[2]\n" + "ldr q29, [x16, #0xc0]\n" + ".inst 0x4f80eb8b // sdot v11.4s, v28.16b, v0.4b[2]\n" + ".inst 0x4f81eb8f // sdot v15.4s, v28.16b, v1.4b[2]\n" + ".inst 0x4f82eb93 // sdot v19.4s, v28.16b, v2.4b[2]\n" + ".inst 0x4f83eb97 // sdot v23.4s, v28.16b, v3.4b[2]\n" + ".inst 0x4f84eb9b // sdot v27.4s, v28.16b, v4.4b[2]\n" + "ldr q28, [x16, #0xd0]\n" + ".inst 0x4fa0eba8 // sdot v8.4s, v29.16b, v0.4b[3]\n" + ".inst 0x4fa1ebac // sdot v12.4s, v29.16b, v1.4b[3]\n" + ".inst 0x4fa2ebb0 // sdot v16.4s, v29.16b, v2.4b[3]\n" + ".inst 0x4fa3ebb4 // sdot v20.4s, v29.16b, v3.4b[3]\n" + ".inst 0x4fa4ebb8 // sdot v24.4s, v29.16b, v4.4b[3]\n" + "ldr q29, [x16, #0xe0]\n" + ".inst 0x4fa0eb89 // sdot v9.4s, v28.16b, v0.4b[3]\n" + ".inst 0x4fa1eb8d // sdot v13.4s, v28.16b, v1.4b[3]\n" + ".inst 0x4fa2eb91 // sdot v17.4s, v28.16b, v2.4b[3]\n" + ".inst 0x4fa3eb95 // sdot v21.4s, v28.16b, v3.4b[3]\n" + ".inst 0x4fa4eb99 // sdot v25.4s, v28.16b, v4.4b[3]\n" + "ldr q28, [x16, #0xf0]\n" + ".inst 0x4fa0ebaa // sdot v10.4s, v29.16b, v0.4b[3]\n" "add x16, x16, #0x100\n" - ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" - ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" - ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" - ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" - ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" - ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" - ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" + ".inst 0x4fa1ebae // sdot v14.4s, v29.16b, v1.4b[3]\n" + ".inst 0x4fa2ebb2 // sdot v18.4s, v29.16b, v2.4b[3]\n" + ".inst 0x4fa3ebb6 // sdot v22.4s, v29.16b, v3.4b[3]\n" + ".inst 0x4fa4ebba // sdot v26.4s, v29.16b, v4.4b[3]\n" + ".inst 0x4fa0eb8b // sdot v11.4s, v28.16b, v0.4b[3]\n" + ".inst 0x4fa1eb8f // sdot v15.4s, v28.16b, v1.4b[3]\n" + ".inst 0x4fa2eb93 // sdot v19.4s, v28.16b, v2.4b[3]\n" + ".inst 0x4fa3eb97 // sdot v23.4s, v28.16b, v3.4b[3]\n" + ".inst 0x4fa4eb9b // sdot v27.4s, v28.16b, v4.4b[3]\n" "155:" // Height 5: Multiply loop: Main loop skip "cbz x14, 160f\n" "cmp x14, #0x4\n" "blt 157f\n" "156:" // Height 5: Multiply loop: Odd block loop - "ldr s0, [x13], #0x4\n" + "ldr s2, [x13], #0x4\n" "sub x14, x14, #0x4\n" - "ldr s1, [x9], #0x4\n" + "ldr s1, [x12], #0x4\n" "cmp x14, #0x4\n" - "ldr s2, [x27], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x23], #0x4\n" - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x16, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr s0, [x11], #0x4\n" + "ldr s31, [x10], #0x4\n" + "ldr s30, [x9], #0x4\n" + "ldr q29, [x16, #0x0]\n" + ".inst 0x4f82e3a8 // sdot v8.4s, v29.16b, v2.4b[0]\n" + "ldr q28, [x16, #0x10]\n" + ".inst 0x4f81e3ac // sdot v12.4s, v29.16b, v1.4b[0]\n" + ".inst 0x4f80e3b0 // sdot v16.4s, v29.16b, v0.4b[0]\n" + ".inst 0x4f9fe3b4 // sdot v20.4s, v29.16b, v31.4b[0]\n" + ".inst 0x4f9ee3b8 // sdot v24.4s, v29.16b, v30.4b[0]\n" + "ldr q29, [x16, #0x20]\n" + ".inst 0x4f82e389 // sdot v9.4s, v28.16b, v2.4b[0]\n" + ".inst 0x4f81e38d // sdot v13.4s, v28.16b, v1.4b[0]\n" + ".inst 0x4f80e391 // sdot v17.4s, v28.16b, v0.4b[0]\n" + ".inst 0x4f9fe395 // sdot v21.4s, v28.16b, v31.4b[0]\n" + ".inst 0x4f9ee399 // sdot v25.4s, v28.16b, v30.4b[0]\n" + "ldr q28, [x16, #0x30]\n" + ".inst 0x4f82e3aa // sdot v10.4s, v29.16b, v2.4b[0]\n" "add x16, x16, #0x40\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" + ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n" + ".inst 0x4f80e3b2 // sdot v18.4s, v29.16b, v0.4b[0]\n" + ".inst 0x4f9fe3b6 // sdot v22.4s, v29.16b, v31.4b[0]\n" + ".inst 0x4f9ee3ba // sdot v26.4s, v29.16b, v30.4b[0]\n" + ".inst 0x4f82e38b // sdot v11.4s, v28.16b, v2.4b[0]\n" + ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n" + ".inst 0x4f80e393 // sdot v19.4s, v28.16b, v0.4b[0]\n" + ".inst 0x4f9fe397 // sdot v23.4s, v28.16b, v31.4b[0]\n" + ".inst 0x4f9ee39b // sdot v27.4s, v28.16b, v30.4b[0]\n" "bge 156b\n" "157:" // Height 5: Multiply loop: Skip odd blocks "cbz x14, 160f\n" "tbz x14, #1, 158f\n" "ldr h0, [x13], #0x2\n" - "ldr h1, [x9], #0x2\n" - "ldr h2, [x27], #0x2\n" - "ldr h3, [x25], #0x2\n" - "ldr h4, [x23], #0x2\n" + "ldr h1, [x12], #0x2\n" + "ldr h2, [x11], #0x2\n" + "ldr h3, [x10], #0x2\n" + "ldr h4, [x9], #0x2\n" "tbz x14, #0, 159f\n" "ld1 { v0.b }[2], [x13]\n" - "ld1 { v1.b }[2], [x9]\n" - "ld1 { v2.b }[2], [x27]\n" - "ld1 { v3.b }[2], [x25]\n" - "ld1 { v4.b }[2], [x23]\n" + "ld1 { v1.b }[2], [x12]\n" + "ld1 { v2.b }[2], [x11]\n" + "ld1 { v3.b }[2], [x10]\n" + "ld1 { v4.b }[2], [x9]\n" "b 159f\n" "158:" // Height 5: Multiply loop: Ragged operand read: partial_1_0 "ldr b0, [x13, #0x0]\n" - "ldr b1, [x9, #0x0]\n" - "ldr b2, [x27, #0x0]\n" - "ldr b3, [x25, #0x0]\n" - "ldr b4, [x23, #0x0]\n" + "ldr b1, [x12, #0x0]\n" + "ldr b2, [x11, #0x0]\n" + "ldr b3, [x10, #0x0]\n" + "ldr b4, [x9, #0x0]\n" "159:" // Height 5: Multiply loop: Ragged operand read: Done - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x16, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr q29, [x16, #0x0]\n" + ".inst 0x4f80e3a8 // sdot v8.4s, v29.16b, v0.4b[0]\n" + "ldr q28, [x16, #0x10]\n" + ".inst 0x4f81e3ac // sdot v12.4s, v29.16b, v1.4b[0]\n" + ".inst 0x4f82e3b0 // sdot v16.4s, v29.16b, v2.4b[0]\n" + ".inst 0x4f83e3b4 // sdot v20.4s, v29.16b, v3.4b[0]\n" + ".inst 0x4f84e3b8 // sdot v24.4s, v29.16b, v4.4b[0]\n" + "ldr q29, [x16, #0x20]\n" + ".inst 0x4f80e389 // sdot v9.4s, v28.16b, v0.4b[0]\n" + ".inst 0x4f81e38d // sdot v13.4s, v28.16b, v1.4b[0]\n" + ".inst 0x4f82e391 // sdot v17.4s, v28.16b, v2.4b[0]\n" + ".inst 0x4f83e395 // sdot v21.4s, v28.16b, v3.4b[0]\n" + ".inst 0x4f84e399 // sdot v25.4s, v28.16b, v4.4b[0]\n" + "ldr q28, [x16, #0x30]\n" + ".inst 0x4f80e3aa // sdot v10.4s, v29.16b, v0.4b[0]\n" "add x16, x16, #0x40\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" + ".inst 0x4f81e3ae // sdot v14.4s, v29.16b, v1.4b[0]\n" + ".inst 0x4f82e3b2 // sdot v18.4s, v29.16b, v2.4b[0]\n" + ".inst 0x4f83e3b6 // sdot v22.4s, v29.16b, v3.4b[0]\n" + ".inst 0x4f84e3ba // sdot v26.4s, v29.16b, v4.4b[0]\n" + ".inst 0x4f80e38b // sdot v11.4s, v28.16b, v0.4b[0]\n" + ".inst 0x4f81e38f // sdot v15.4s, v28.16b, v1.4b[0]\n" + ".inst 0x4f82e393 // sdot v19.4s, v28.16b, v2.4b[0]\n" + ".inst 0x4f83e397 // sdot v23.4s, v28.16b, v3.4b[0]\n" + ".inst 0x4f84e39b // sdot v27.4s, v28.16b, v4.4b[0]\n" "160:" // Height 5: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x15, x15, #0x1\n" @@ -2862,98 +2861,98 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "184:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w14, [x20, x15, 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, 185f\n" - "ldr x21, [%x[input_ptr], x15, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x13, [x21, #0x0]\n" - "ldr x9, [x21, #0x8]\n" - "ldr x27, [x21, #0x10]\n" - "ldr x25, [x21, #0x18]\n" - "ldr x23, [x21, #0x20]\n" - "ldr x21, [x21, #0x28]\n" + "ldr x20, [%x[input_ptr], x15, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x13, [x20, #0x0]\n" + "ldr x12, [x20, #0x8]\n" + "ldr x11, [x20, #0x10]\n" + "ldr x10, [x20, #0x18]\n" + "ldr x9, [x20, #0x20]\n" + "ldr x28, [x20, #0x28]\n" "cbnz x15, 186f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x13, x13, x20\n" + "add x12, x12, x20\n" + "add x11, x11, x20\n" + "add x10, x10, x20\n" "add x9, x9, x20\n" - "add x27, x27, x20\n" - "add x25, x25, x20\n" - "add x23, x23, x20\n" - "add x21, x21, x20\n" + "add x28, x28, x20\n" "b 186f\n" "185:" // Height 6: setup direct input "mov x13, %x[input_ptr]\n" - "add x9, x13, x20\n" - "add x27, x9, x20\n" - "add x25, x27, x20\n" - "add x23, x25, x20\n" - "add x21, x23, x20\n" + "add x12, x13, x21\n" + "add x11, x12, x21\n" + "add x10, x11, x21\n" + "add x9, x10, x21\n" + "add x28, x9, x21\n" "186:" // Height 6: input setup done "cmp x14, #0x10\n" "blt 189f\n" "ldr q0, [x13, #0x0]\n" "cmp x14, #0x20\n" - "ldr q1, [x9, #0x0]\n" - "ldr q2, [x27, #0x0]\n" - "ldr q3, [x25, #0x0]\n" - "ldr q4, [x23, #0x0]\n" - "ldr q5, [x21, #0x0]\n" + "ldr q1, [x12, #0x0]\n" + "ldr q2, [x11, #0x0]\n" + "ldr q3, [x10, #0x0]\n" + "ldr q4, [x9, #0x0]\n" + "ldr q5, [x28, #0x0]\n" "ldr q6, [x16, #0x0]\n" "ldr q7, [x16, #0x10]\n" "blt 188f\n" "187:" // Height 6: Multiply loop: Main loop head ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr x12, [x16, #0x28]\n" + "ldr x21, [x16, #0x28]\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "ldr x11, [x16, #0x38]\n" + "ldr x20, [x16, #0x38]\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" "add x13, x13, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "add x9, x9, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "add x27, x27, #0x10\n" + "add x11, x11, #0x10\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" "ldr d6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "mov v6.d[1], x12\n" + "mov v6.d[1], x21\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - "ldr x12, [x16, #0x48]\n" + "ldr x21, [x16, #0x48]\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - "add x25, x25, #0x10\n" + "add x10, x10, #0x10\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "add x23, x23, #0x10\n" + "add x9, x9, #0x10\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "add x21, x21, #0x10\n" + "add x28, x28, #0x10\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" "ldr d7, [x16, #0x30]\n" - "mov v7.d[1], x11\n" + "mov v7.d[1], x20\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "ldr x11, [x16, #0x58]\n" + "ldr x20, [x16, #0x58]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "ldr x10, [x13, #0x8]\n" + "ldr x27, [x13, #0x8]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - "ldr x28, [x9, #0x8]\n" + "ldr x26, [x12, #0x8]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - "ldr x26, [x27, #0x8]\n" + "ldr x25, [x11, #0x8]\n" ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n" "ldr d6, [x16, #0x40]\n" ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - "mov v6.d[1], x12\n" + "mov v6.d[1], x21\n" ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - "ldr x12, [x16, #0x68]\n" + "ldr x21, [x16, #0x68]\n" ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - "ldr x24, [x25, #0x8]\n" + "ldr x24, [x10, #0x8]\n" ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - "ldr x22, [x23, #0x8]\n" + "ldr x23, [x9, #0x8]\n" ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" - "ldr x20, [x21, #0x8]\n" + "ldr x22, [x28, #0x8]\n" ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" "ldr d7, [x16, #0x50]\n" - "mov v7.d[1], x11\n" + "mov v7.d[1], x20\n" ".inst 0x4fa0e0c8 // sdot v8.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0cc // sdot v12.4s, v6.16b, v1.4b[1]\n" - "ldr x11, [x16, #0x78]\n" + "ldr x20, [x16, #0x78]\n" ".inst 0x4fa2e0d0 // sdot v16.4s, v6.16b, v2.4b[1]\n" "sub x14, x14, #0x10\n" ".inst 0x4fa3e0d4 // sdot v20.4s, v6.16b, v3.4b[1]\n" @@ -2963,96 +2962,96 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( ".inst 0x4fa5e0dc // sdot v28.4s, v6.16b, v5.4b[1]\n" "ldr d6, [x16, #0x60]\n" ".inst 0x4fa0e0e9 // sdot v9.4s, v7.16b, v0.4b[1]\n" - "mov v6.d[1], x12\n" + "mov v6.d[1], x21\n" ".inst 0x4fa1e0ed // sdot v13.4s, v7.16b, v1.4b[1]\n" - "ldr x12, [x16, #0x88]\n" + "ldr x21, [x16, #0x88]\n" ".inst 0x4fa2e0f1 // sdot v17.4s, v7.16b, v2.4b[1]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4fa3e0f5 // sdot v21.4s, v7.16b, v3.4b[1]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x11, #0x80]\n" ".inst 0x4fa4e0f9 // sdot v25.4s, v7.16b, v4.4b[1]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x10, #0x80]\n" ".inst 0x4fa5e0fd // sdot v29.4s, v7.16b, v5.4b[1]\n" "ldr d7, [x16, #0x70]\n" - "mov v7.d[1], x11\n" + "mov v7.d[1], x20\n" ".inst 0x4fa0e0ca // sdot v10.4s, v6.16b, v0.4b[1]\n" ".inst 0x4fa1e0ce // sdot v14.4s, v6.16b, v1.4b[1]\n" - "ldr x11, [x16, #0x98]\n" + "ldr x20, [x16, #0x98]\n" ".inst 0x4fa2e0d2 // sdot v18.4s, v6.16b, v2.4b[1]\n" - "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x9, #0x80]\n" ".inst 0x4fa3e0d6 // sdot v22.4s, v6.16b, v3.4b[1]\n" - "prfm pldl1keep, [x21, #0x80]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4fa4e0da // sdot v26.4s, v6.16b, v4.4b[1]\n" ".inst 0x4fa5e0de // sdot v30.4s, v6.16b, v5.4b[1]\n" "ldr d6, [x16, #0x80]\n" ".inst 0x4fa0e0eb // sdot v11.4s, v7.16b, v0.4b[1]\n" - "mov v6.d[1], x12\n" + "mov v6.d[1], x21\n" ".inst 0x4fa1e0ef // sdot v15.4s, v7.16b, v1.4b[1]\n" - "ldr x12, [x16, #0xa8]\n" + "ldr x21, [x16, #0xa8]\n" ".inst 0x4fa2e0f3 // sdot v19.4s, v7.16b, v2.4b[1]\n" ".inst 0x4fa3e0f7 // sdot v23.4s, v7.16b, v3.4b[1]\n" ".inst 0x4fa4e0fb // sdot v27.4s, v7.16b, v4.4b[1]\n" ".inst 0x4fa5e0ff // sdot v31.4s, v7.16b, v5.4b[1]\n" "ldr d7, [x16, #0x90]\n" - "mov v7.d[1], x11\n" + "mov v7.d[1], x20\n" ".inst 0x4f80e8c8 // sdot v8.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8cc // sdot v12.4s, v6.16b, v1.4b[2]\n" - "ldr x11, [x16, #0xb8]\n" + "ldr x20, [x16, #0xb8]\n" ".inst 0x4f82e8d0 // sdot v16.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d4 // sdot v20.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8d8 // sdot v24.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8dc // sdot v28.4s, v6.16b, v5.4b[2]\n" "ldr d6, [x16, #0xa0]\n" ".inst 0x4f80e8e9 // sdot v9.4s, v7.16b, v0.4b[2]\n" - "mov v6.d[1], x12\n" + "mov v6.d[1], x21\n" ".inst 0x4f81e8ed // sdot v13.4s, v7.16b, v1.4b[2]\n" - "ldr x12, [x16, #0xc8]\n" + "ldr x21, [x16, #0xc8]\n" ".inst 0x4f82e8f1 // sdot v17.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f5 // sdot v21.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8fd // sdot v29.4s, v7.16b, v5.4b[2]\n" "ldr d7, [x16, #0xb0]\n" - "mov v7.d[1], x11\n" + "mov v7.d[1], x20\n" ".inst 0x4f80e8ca // sdot v10.4s, v6.16b, v0.4b[2]\n" ".inst 0x4f81e8ce // sdot v14.4s, v6.16b, v1.4b[2]\n" - "ldr x11, [x16, #0xd8]\n" + "ldr x20, [x16, #0xd8]\n" ".inst 0x4f82e8d2 // sdot v18.4s, v6.16b, v2.4b[2]\n" ".inst 0x4f83e8d6 // sdot v22.4s, v6.16b, v3.4b[2]\n" ".inst 0x4f84e8da // sdot v26.4s, v6.16b, v4.4b[2]\n" ".inst 0x4f85e8de // sdot v30.4s, v6.16b, v5.4b[2]\n" "ldr d6, [x16, #0xc0]\n" ".inst 0x4f80e8eb // sdot v11.4s, v7.16b, v0.4b[2]\n" - "mov v6.d[1], x12\n" + "mov v6.d[1], x21\n" ".inst 0x4f81e8ef // sdot v15.4s, v7.16b, v1.4b[2]\n" - "ldr x12, [x16, #0xe8]\n" + "ldr x21, [x16, #0xe8]\n" ".inst 0x4f82e8f3 // sdot v19.4s, v7.16b, v2.4b[2]\n" ".inst 0x4f83e8f7 // sdot v23.4s, v7.16b, v3.4b[2]\n" ".inst 0x4f84e8fb // sdot v27.4s, v7.16b, v4.4b[2]\n" ".inst 0x4f85e8ff // sdot v31.4s, v7.16b, v5.4b[2]\n" "ldr d7, [x16, #0xd0]\n" - "mov v7.d[1], x11\n" + "mov v7.d[1], x20\n" ".inst 0x4fa0e8c8 // sdot v8.4s, v6.16b, v0.4b[3]\n" ".inst 0x4fa1e8cc // sdot v12.4s, v6.16b, v1.4b[3]\n" - "ldr x11, [x16, #0xf8]\n" + "ldr x20, [x16, #0xf8]\n" ".inst 0x4fa2e8d0 // sdot v16.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d4 // sdot v20.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8d8 // sdot v24.4s, v6.16b, v4.4b[3]\n" ".inst 0x4fa5e8dc // sdot v28.4s, v6.16b, v5.4b[3]\n" "ldr d6, [x16, #0xe0]\n" ".inst 0x4fa0e8e9 // sdot v9.4s, v7.16b, v0.4b[3]\n" - "mov v6.d[1], x12\n" + "mov v6.d[1], x21\n" ".inst 0x4fa1e8ed // sdot v13.4s, v7.16b, v1.4b[3]\n" ".inst 0x4fa2e8f1 // sdot v17.4s, v7.16b, v2.4b[3]\n" ".inst 0x4fa3e8f5 // sdot v21.4s, v7.16b, v3.4b[3]\n" ".inst 0x4fa4e8f9 // sdot v25.4s, v7.16b, v4.4b[3]\n" ".inst 0x4fa5e8fd // sdot v29.4s, v7.16b, v5.4b[3]\n" "ldr d7, [x16, #0xf0]\n" - "mov v7.d[1], x11\n" + "mov v7.d[1], x20\n" "add x16, x16, #0x100\n" ".inst 0x4fa0e8ca // sdot v10.4s, v6.16b, v0.4b[3]\n" - "ldr x12, [x16, #0x8]\n" + "ldr x21, [x16, #0x8]\n" ".inst 0x4fa1e8ce // sdot v14.4s, v6.16b, v1.4b[3]\n" - "ldr x11, [x16, #0x18]\n" + "ldr x20, [x16, #0x18]\n" ".inst 0x4fa2e8d2 // sdot v18.4s, v6.16b, v2.4b[3]\n" ".inst 0x4fa3e8d6 // sdot v22.4s, v6.16b, v3.4b[3]\n" ".inst 0x4fa4e8da // sdot v26.4s, v6.16b, v4.4b[3]\n" @@ -3061,56 +3060,56 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( ".inst 0x4fa0e8eb // sdot v11.4s, v7.16b, v0.4b[3]\n" "ldr d0, [x13, #0x0]\n" ".inst 0x4fa1e8ef // sdot v15.4s, v7.16b, v1.4b[3]\n" - "ldr d1, [x9, #0x0]\n" + "ldr d1, [x12, #0x0]\n" ".inst 0x4fa2e8f3 // sdot v19.4s, v7.16b, v2.4b[3]\n" - "ldr d2, [x27, #0x0]\n" + "ldr d2, [x11, #0x0]\n" ".inst 0x4fa3e8f7 // sdot v23.4s, v7.16b, v3.4b[3]\n" - "ldr d3, [x25, #0x0]\n" + "ldr d3, [x10, #0x0]\n" ".inst 0x4fa4e8fb // sdot v27.4s, v7.16b, v4.4b[3]\n" - "ldr d4, [x23, #0x0]\n" + "ldr d4, [x9, #0x0]\n" ".inst 0x4fa5e8ff // sdot v31.4s, v7.16b, v5.4b[3]\n" - "ldr d5, [x21, #0x0]\n" + "ldr d5, [x28, #0x0]\n" "ldr d7, [x16, #0x10]\n" - "mov v6.d[1], x12\n" - "mov v0.d[1], x10\n" - "mov v1.d[1], x28\n" - "mov v2.d[1], x26\n" + "mov v6.d[1], x21\n" + "mov v0.d[1], x27\n" + "mov v1.d[1], x26\n" + "mov v2.d[1], x25\n" "mov v3.d[1], x24\n" - "mov v4.d[1], x22\n" - "mov v5.d[1], x20\n" - "mov v7.d[1], x11\n" + "mov v4.d[1], x23\n" + "mov v5.d[1], x22\n" + "mov v7.d[1], x20\n" "bge 187b\n" "188:" // Height 6: Multiply loop: Single iteration only ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" "add x13, x13, #0x10\n" ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - "add x9, x9, #0x10\n" + "add x12, x12, #0x10\n" ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - "add x27, x27, #0x10\n" + "add x11, x11, #0x10\n" ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - "add x25, x25, #0x10\n" + "add x10, x10, #0x10\n" ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - "add x23, x23, #0x10\n" + "add x9, x9, #0x10\n" ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" "ldr q6, [x16, #0x20]\n" ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - "add x21, x21, #0x10\n" + "add x28, x28, #0x10\n" ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" "sub x14, x14, #0x10\n" ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" "prfm pldl1keep, [x13, #0x80]\n" ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - "prfm pldl1keep, [x9, #0x80]\n" + "prfm pldl1keep, [x12, #0x80]\n" ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - "prfm pldl1keep, [x27, #0x80]\n" + "prfm pldl1keep, [x11, #0x80]\n" ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" "ldr q7, [x16, #0x30]\n" ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x10, #0x80]\n" ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x9, #0x80]\n" ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - "prfm pldl1keep, [x21, #0x80]\n" + "prfm pldl1keep, [x28, #0x80]\n" ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n" @@ -3210,98 +3209,98 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "cmp x14, #0x4\n" "blt 191f\n" "190:" // Height 6: Multiply loop: Odd block loop - "ldr s0, [x13], #0x4\n" + "ldr s7, [x13], #0x4\n" "sub x14, x14, #0x4\n" - "ldr s1, [x9], #0x4\n" + "ldr s6, [x12], #0x4\n" "cmp x14, #0x4\n" - "ldr s2, [x27], #0x4\n" - "ldr s3, [x25], #0x4\n" - "ldr s4, [x23], #0x4\n" - "ldr s5, [x21], #0x4\n" - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" - "ldr q6, [x16, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr s5, [x11], #0x4\n" + "ldr s4, [x10], #0x4\n" + "ldr s3, [x9], #0x4\n" + "ldr s2, [x28], #0x4\n" + "ldr q1, [x16, #0x0]\n" + ".inst 0x4f87e028 // sdot v8.4s, v1.16b, v7.4b[0]\n" + "ldr q0, [x16, #0x10]\n" + ".inst 0x4f86e02c // sdot v12.4s, v1.16b, v6.4b[0]\n" + ".inst 0x4f85e030 // sdot v16.4s, v1.16b, v5.4b[0]\n" + ".inst 0x4f84e034 // sdot v20.4s, v1.16b, v4.4b[0]\n" + ".inst 0x4f83e038 // sdot v24.4s, v1.16b, v3.4b[0]\n" + ".inst 0x4f82e03c // sdot v28.4s, v1.16b, v2.4b[0]\n" + "ldr q1, [x16, #0x20]\n" + ".inst 0x4f87e009 // sdot v9.4s, v0.16b, v7.4b[0]\n" + ".inst 0x4f86e00d // sdot v13.4s, v0.16b, v6.4b[0]\n" + ".inst 0x4f85e011 // sdot v17.4s, v0.16b, v5.4b[0]\n" + ".inst 0x4f84e015 // sdot v21.4s, v0.16b, v4.4b[0]\n" + ".inst 0x4f83e019 // sdot v25.4s, v0.16b, v3.4b[0]\n" + ".inst 0x4f82e01d // sdot v29.4s, v0.16b, v2.4b[0]\n" + "ldr q0, [x16, #0x30]\n" + ".inst 0x4f87e02a // sdot v10.4s, v1.16b, v7.4b[0]\n" "add x16, x16, #0x40\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" - ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" + ".inst 0x4f86e02e // sdot v14.4s, v1.16b, v6.4b[0]\n" + ".inst 0x4f85e032 // sdot v18.4s, v1.16b, v5.4b[0]\n" + ".inst 0x4f84e036 // sdot v22.4s, v1.16b, v4.4b[0]\n" + ".inst 0x4f83e03a // sdot v26.4s, v1.16b, v3.4b[0]\n" + ".inst 0x4f82e03e // sdot v30.4s, v1.16b, v2.4b[0]\n" + ".inst 0x4f87e00b // sdot v11.4s, v0.16b, v7.4b[0]\n" + ".inst 0x4f86e00f // sdot v15.4s, v0.16b, v6.4b[0]\n" + ".inst 0x4f85e013 // sdot v19.4s, v0.16b, v5.4b[0]\n" + ".inst 0x4f84e017 // sdot v23.4s, v0.16b, v4.4b[0]\n" + ".inst 0x4f83e01b // sdot v27.4s, v0.16b, v3.4b[0]\n" + ".inst 0x4f82e01f // sdot v31.4s, v0.16b, v2.4b[0]\n" "bge 190b\n" "191:" // Height 6: Multiply loop: Skip odd blocks "cbz x14, 194f\n" "tbz x14, #1, 192f\n" "ldr h0, [x13], #0x2\n" - "ldr h1, [x9], #0x2\n" - "ldr h2, [x27], #0x2\n" - "ldr h3, [x25], #0x2\n" - "ldr h4, [x23], #0x2\n" - "ldr h5, [x21], #0x2\n" + "ldr h1, [x12], #0x2\n" + "ldr h2, [x11], #0x2\n" + "ldr h3, [x10], #0x2\n" + "ldr h4, [x9], #0x2\n" + "ldr h5, [x28], #0x2\n" "tbz x14, #0, 193f\n" "ld1 { v0.b }[2], [x13]\n" - "ld1 { v1.b }[2], [x9]\n" - "ld1 { v2.b }[2], [x27]\n" - "ld1 { v3.b }[2], [x25]\n" - "ld1 { v4.b }[2], [x23]\n" - "ld1 { v5.b }[2], [x21]\n" + "ld1 { v1.b }[2], [x12]\n" + "ld1 { v2.b }[2], [x11]\n" + "ld1 { v3.b }[2], [x10]\n" + "ld1 { v4.b }[2], [x9]\n" + "ld1 { v5.b }[2], [x28]\n" "b 193f\n" "192:" // Height 6: Multiply loop: Ragged operand read: partial_1_0 "ldr b0, [x13, #0x0]\n" - "ldr b1, [x9, #0x0]\n" - "ldr b2, [x27, #0x0]\n" - "ldr b3, [x25, #0x0]\n" - "ldr b4, [x23, #0x0]\n" - "ldr b5, [x21, #0x0]\n" + "ldr b1, [x12, #0x0]\n" + "ldr b2, [x11, #0x0]\n" + "ldr b3, [x10, #0x0]\n" + "ldr b4, [x9, #0x0]\n" + "ldr b5, [x28, #0x0]\n" "193:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q6, [x16, #0x0]\n" - ".inst 0x4f80e0c8 // sdot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q7, [x16, #0x10]\n" - ".inst 0x4f81e0cc // sdot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d0 // sdot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d4 // sdot v20.4s, v6.16b, v3.4b[0]\n" - ".inst 0x4f84e0d8 // sdot v24.4s, v6.16b, v4.4b[0]\n" - ".inst 0x4f85e0dc // sdot v28.4s, v6.16b, v5.4b[0]\n" - "ldr q6, [x16, #0x20]\n" - ".inst 0x4f80e0e9 // sdot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ed // sdot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f1 // sdot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f5 // sdot v21.4s, v7.16b, v3.4b[0]\n" - ".inst 0x4f84e0f9 // sdot v25.4s, v7.16b, v4.4b[0]\n" - ".inst 0x4f85e0fd // sdot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q7, [x16, #0x30]\n" - ".inst 0x4f80e0ca // sdot v10.4s, v6.16b, v0.4b[0]\n" + "ldr q7, [x16, #0x0]\n" + ".inst 0x4f80e0e8 // sdot v8.4s, v7.16b, v0.4b[0]\n" + "ldr q6, [x16, #0x10]\n" + ".inst 0x4f81e0ec // sdot v12.4s, v7.16b, v1.4b[0]\n" + ".inst 0x4f82e0f0 // sdot v16.4s, v7.16b, v2.4b[0]\n" + ".inst 0x4f83e0f4 // sdot v20.4s, v7.16b, v3.4b[0]\n" + ".inst 0x4f84e0f8 // sdot v24.4s, v7.16b, v4.4b[0]\n" + ".inst 0x4f85e0fc // sdot v28.4s, v7.16b, v5.4b[0]\n" + "ldr q7, [x16, #0x20]\n" + ".inst 0x4f80e0c9 // sdot v9.4s, v6.16b, v0.4b[0]\n" + ".inst 0x4f81e0cd // sdot v13.4s, v6.16b, v1.4b[0]\n" + ".inst 0x4f82e0d1 // sdot v17.4s, v6.16b, v2.4b[0]\n" + ".inst 0x4f83e0d5 // sdot v21.4s, v6.16b, v3.4b[0]\n" + ".inst 0x4f84e0d9 // sdot v25.4s, v6.16b, v4.4b[0]\n" + ".inst 0x4f85e0dd // sdot v29.4s, v6.16b, v5.4b[0]\n" + "ldr q6, [x16, #0x30]\n" + ".inst 0x4f80e0ea // sdot v10.4s, v7.16b, v0.4b[0]\n" "add x16, x16, #0x40\n" - ".inst 0x4f81e0ce // sdot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x4f82e0d2 // sdot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x4f83e0d6 // sdot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x4f84e0da // sdot v26.4s, v6.16b, v4.4b[0]\n" - ".inst 0x4f85e0de // sdot v30.4s, v6.16b, v5.4b[0]\n" - ".inst 0x4f80e0eb // sdot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x4f81e0ef // sdot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x4f82e0f3 // sdot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x4f83e0f7 // sdot v23.4s, v7.16b, v3.4b[0]\n" - ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" - ".inst 0x4f85e0ff // sdot v31.4s, v7.16b, v5.4b[0]\n" + ".inst 0x4f81e0ee // sdot v14.4s, v7.16b, v1.4b[0]\n" + ".inst 0x4f82e0f2 // sdot v18.4s, v7.16b, v2.4b[0]\n" + ".inst 0x4f83e0f6 // sdot v22.4s, v7.16b, v3.4b[0]\n" + ".inst 0x4f84e0fa // sdot v26.4s, v7.16b, v4.4b[0]\n" + ".inst 0x4f85e0fe // sdot v30.4s, v7.16b, v5.4b[0]\n" + ".inst 0x4f80e0cb // sdot v11.4s, v6.16b, v0.4b[0]\n" + ".inst 0x4f81e0cf // sdot v15.4s, v6.16b, v1.4b[0]\n" + ".inst 0x4f82e0d3 // sdot v19.4s, v6.16b, v2.4b[0]\n" + ".inst 0x4f83e0d7 // sdot v23.4s, v6.16b, v3.4b[0]\n" + ".inst 0x4f84e0db // sdot v27.4s, v6.16b, v4.4b[0]\n" + ".inst 0x4f85e0df // sdot v31.4s, v6.16b, v5.4b[0]\n" "194:" // Height 6: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" "add x15, x15, #0x1\n" @@ -3488,7 +3487,6 @@ void a64_hybrid_s8s32_dot_6x16_a55 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "206:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [flags] "r" (flags), [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", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" -- cgit v1.2.1