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_u8u32_dot_6x16/generic.cpp | 1810 ++++++++++---------- 1 file changed, 904 insertions(+), 906 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp index 38131cfd4b..849c680843 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_u8u32_dot_6x16/generic.cpp @@ -77,7 +77,6 @@ void a64_hybrid_u8u32_dot_6x16 ( 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_u8u32_dot_6x16 ( "14:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 15f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" "cbnz x28, 16f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -186,37 +185,37 @@ void a64_hybrid_u8u32_dot_6x16 ( "blt 18f\n" "17:" // Height 1: Multiply loop: Main loop head ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - "ldr q7, [x10, #0xf0]\n" + "ldr q16, [x10, #0x30]\n" + ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" + "ldr q17, [x10, #0x40]\n" + ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" + "ldr q16, [x10, #0x50]\n" + ".inst 0x6fa0e228 // udot v8.4s, v17.16b, v0.4b[1]\n" + "ldr q17, [x10, #0x60]\n" + ".inst 0x6fa0e209 // udot v9.4s, v16.16b, v0.4b[1]\n" + "ldr q16, [x10, #0x70]\n" + ".inst 0x6fa0e22a // udot v10.4s, v17.16b, v0.4b[1]\n" + "ldr q17, [x10, #0x80]\n" + ".inst 0x6fa0e20b // udot v11.4s, v16.16b, v0.4b[1]\n" + "ldr q16, [x10, #0x90]\n" + ".inst 0x6f80ea28 // udot v8.4s, v17.16b, v0.4b[2]\n" + "ldr q17, [x10, #0xa0]\n" + ".inst 0x6f80ea09 // udot v9.4s, v16.16b, v0.4b[2]\n" + "ldr q16, [x10, #0xb0]\n" + ".inst 0x6f80ea2a // udot v10.4s, v17.16b, v0.4b[2]\n" + "ldr q17, [x10, #0xc0]\n" + ".inst 0x6f80ea0b // udot v11.4s, v16.16b, v0.4b[2]\n" + "ldr q16, [x10, #0xd0]\n" + ".inst 0x6fa0ea28 // udot v8.4s, v17.16b, v0.4b[3]\n" + "ldr q17, [x10, #0xe0]\n" + ".inst 0x6fa0ea09 // udot v9.4s, v16.16b, v0.4b[3]\n" + "ldr q16, [x10, #0xf0]\n" "sub x27, x27, #0x10\n" "add x26, x26, #0x10\n" - ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" + ".inst 0x6fa0ea2a // udot v10.4s, v17.16b, v0.4b[3]\n" + ".inst 0x6fa0ea0b // udot v11.4s, v16.16b, v0.4b[3]\n" "ldr q0, [x26, #0x0]\n" "cmp x27, #0x20\n" "add x10, x10, #0x100\n" @@ -226,37 +225,37 @@ void a64_hybrid_u8u32_dot_6x16 ( "bge 17b\n" "18:" // Height 1: Multiply loop: Single iteration only ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - "ldr q7, [x10, #0xf0]\n" + "ldr q16, [x10, #0x30]\n" + ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" + "ldr q17, [x10, #0x40]\n" + ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" + "ldr q16, [x10, #0x50]\n" + ".inst 0x6fa0e228 // udot v8.4s, v17.16b, v0.4b[1]\n" + "ldr q17, [x10, #0x60]\n" + ".inst 0x6fa0e209 // udot v9.4s, v16.16b, v0.4b[1]\n" + "ldr q16, [x10, #0x70]\n" + ".inst 0x6fa0e22a // udot v10.4s, v17.16b, v0.4b[1]\n" + "ldr q17, [x10, #0x80]\n" + ".inst 0x6fa0e20b // udot v11.4s, v16.16b, v0.4b[1]\n" + "ldr q16, [x10, #0x90]\n" + ".inst 0x6f80ea28 // udot v8.4s, v17.16b, v0.4b[2]\n" + "ldr q17, [x10, #0xa0]\n" + ".inst 0x6f80ea09 // udot v9.4s, v16.16b, v0.4b[2]\n" + "ldr q16, [x10, #0xb0]\n" + ".inst 0x6f80ea2a // udot v10.4s, v17.16b, v0.4b[2]\n" + "ldr q17, [x10, #0xc0]\n" + ".inst 0x6f80ea0b // udot v11.4s, v16.16b, v0.4b[2]\n" + "ldr q16, [x10, #0xd0]\n" + ".inst 0x6fa0ea28 // udot v8.4s, v17.16b, v0.4b[3]\n" + "ldr q17, [x10, #0xe0]\n" + ".inst 0x6fa0ea09 // udot v9.4s, v16.16b, v0.4b[3]\n" + "ldr q16, [x10, #0xf0]\n" "add x26, x26, #0x10\n" "sub x27, x27, #0x10\n" - ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" + ".inst 0x6fa0ea2a // udot v10.4s, v17.16b, v0.4b[3]\n" + ".inst 0x6fa0ea0b // udot v11.4s, v16.16b, v0.4b[3]\n" "prfm pldl1keep, [x26, #0x80]\n" "add x10, x10, #0x100\n" "19:" // Height 1: Multiply loop: Main loop skip @@ -264,17 +263,17 @@ void a64_hybrid_u8u32_dot_6x16 ( "cmp x27, #0x4\n" "blt 21f\n" "20:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr q6, [x10, #0x0]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + "ldr s18, [x26], #0x4\n" + "ldr q16, [x10, #0x0]\n" + ".inst 0x6f92e208 // udot v8.4s, v16.16b, v18.4b[0]\n" "sub x27, x27, #0x4\n" - "ldr q7, [x10, #0x10]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "ldr q16, [x10, #0x10]\n" + "ldr q17, [x10, #0x20]\n" + ".inst 0x6f92e209 // udot v9.4s, v16.16b, v18.4b[0]\n" "cmp x27, #0x4\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + "ldr q16, [x10, #0x30]\n" + ".inst 0x6f92e22a // udot v10.4s, v17.16b, v18.4b[0]\n" + ".inst 0x6f92e20b // udot v11.4s, v16.16b, v18.4b[0]\n" "add x10, x10, #0x40\n" "bge 20b\n" "21:" // Height 1: Multiply loop: Skip odd blocks @@ -287,14 +286,14 @@ void a64_hybrid_u8u32_dot_6x16 ( "22:" // Height 1: Multiply loop: Ragged operand read: partial_1_0 "ldr b0, [x26, #0x0]\n" "23:" // Height 1: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + "ldr q17, [x10, #0x0]\n" + "ldr q16, [x10, #0x10]\n" + ".inst 0x6f80e228 // udot v8.4s, v17.16b, v0.4b[0]\n" + ".inst 0x6f80e209 // udot v9.4s, v16.16b, v0.4b[0]\n" + "ldr q17, [x10, #0x20]\n" + "ldr q16, [x10, #0x30]\n" + ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" + ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" "add x10, x10, #0x40\n" "24:" // Height 1: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -465,12 +464,12 @@ void a64_hybrid_u8u32_dot_6x16 ( "48:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 49f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" "cbnz x28, 50f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -478,7 +477,7 @@ void a64_hybrid_u8u32_dot_6x16 ( "b 50f\n" "49:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" + "add x25, x26, x21\n" "50:" // Height 2: input setup done "cmp x27, #0x10\n" "blt 53f\n" @@ -491,137 +490,137 @@ void a64_hybrid_u8u32_dot_6x16 ( "51:" // Height 2: Multiply loop: Main loop head ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "sub x27, x27, #0x10\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q16, [x10, #0x30]\n" "add x26, x26, #0x10\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x10, #0x40]\n" + ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" + ".inst 0x6f81e22e // udot v14.4s, v17.16b, v1.4b[0]\n" + "ldr q17, [x10, #0x40]\n" "add x25, x25, #0x10\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x10, #0x50]\n" + ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" + ".inst 0x6f81e20f // udot v15.4s, v16.16b, v1.4b[0]\n" + "ldr q16, [x10, #0x50]\n" "cmp x27, #0x20\n" - ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x10, #0x60]\n" + ".inst 0x6fa0e228 // udot v8.4s, v17.16b, v0.4b[1]\n" + ".inst 0x6fa1e22c // udot v12.4s, v17.16b, v1.4b[1]\n" + "ldr q17, [x10, #0x60]\n" "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x10, #0x70]\n" + ".inst 0x6fa0e209 // udot v9.4s, v16.16b, v0.4b[1]\n" + ".inst 0x6fa1e20d // udot v13.4s, v16.16b, v1.4b[1]\n" + "ldr q16, [x10, #0x70]\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" - "ldr q7, [x10, #0xf0]\n" + ".inst 0x6fa0e22a // udot v10.4s, v17.16b, v0.4b[1]\n" + ".inst 0x6fa1e22e // udot v14.4s, v17.16b, v1.4b[1]\n" + "ldr q17, [x10, #0x80]\n" + ".inst 0x6fa0e20b // udot v11.4s, v16.16b, v0.4b[1]\n" + ".inst 0x6fa1e20f // udot v15.4s, v16.16b, v1.4b[1]\n" + "ldr q16, [x10, #0x90]\n" + ".inst 0x6f80ea28 // udot v8.4s, v17.16b, v0.4b[2]\n" + ".inst 0x6f81ea2c // udot v12.4s, v17.16b, v1.4b[2]\n" + "ldr q17, [x10, #0xa0]\n" + ".inst 0x6f80ea09 // udot v9.4s, v16.16b, v0.4b[2]\n" + ".inst 0x6f81ea0d // udot v13.4s, v16.16b, v1.4b[2]\n" + "ldr q16, [x10, #0xb0]\n" + ".inst 0x6f80ea2a // udot v10.4s, v17.16b, v0.4b[2]\n" + ".inst 0x6f81ea2e // udot v14.4s, v17.16b, v1.4b[2]\n" + "ldr q17, [x10, #0xc0]\n" + ".inst 0x6f80ea0b // udot v11.4s, v16.16b, v0.4b[2]\n" + ".inst 0x6f81ea0f // udot v15.4s, v16.16b, v1.4b[2]\n" + "ldr q16, [x10, #0xd0]\n" + ".inst 0x6fa0ea28 // udot v8.4s, v17.16b, v0.4b[3]\n" + ".inst 0x6fa1ea2c // udot v12.4s, v17.16b, v1.4b[3]\n" + "ldr q17, [x10, #0xe0]\n" + ".inst 0x6fa0ea09 // udot v9.4s, v16.16b, v0.4b[3]\n" + ".inst 0x6fa1ea0d // udot v13.4s, v16.16b, v1.4b[3]\n" + "ldr q16, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" + ".inst 0x6fa0ea2a // udot v10.4s, v17.16b, v0.4b[3]\n" + ".inst 0x6fa1ea2e // udot v14.4s, v17.16b, v1.4b[3]\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" + ".inst 0x6fa0ea0b // udot v11.4s, v16.16b, v0.4b[3]\n" "ldr q0, [x26, #0x0]\n" - ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" + ".inst 0x6fa1ea0f // udot v15.4s, v16.16b, v1.4b[3]\n" "ldr q1, [x25, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 51b\n" "52:" // Height 2: Multiply loop: Single iteration only ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "add x26, x26, #0x10\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q16, [x10, #0x30]\n" "add x25, x25, #0x10\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x10, #0x40]\n" + ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" + ".inst 0x6f81e22e // udot v14.4s, v17.16b, v1.4b[0]\n" + "ldr q17, [x10, #0x40]\n" "sub x27, x27, #0x10\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x10, #0x50]\n" + ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" + ".inst 0x6f81e20f // udot v15.4s, v16.16b, v1.4b[0]\n" + "ldr q16, [x10, #0x50]\n" "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x10, #0x60]\n" + ".inst 0x6fa0e228 // udot v8.4s, v17.16b, v0.4b[1]\n" + ".inst 0x6fa1e22c // udot v12.4s, v17.16b, v1.4b[1]\n" + "ldr q17, [x10, #0x60]\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" - "ldr q7, [x10, #0xf0]\n" + ".inst 0x6fa0e209 // udot v9.4s, v16.16b, v0.4b[1]\n" + ".inst 0x6fa1e20d // udot v13.4s, v16.16b, v1.4b[1]\n" + "ldr q16, [x10, #0x70]\n" + ".inst 0x6fa0e22a // udot v10.4s, v17.16b, v0.4b[1]\n" + ".inst 0x6fa1e22e // udot v14.4s, v17.16b, v1.4b[1]\n" + "ldr q17, [x10, #0x80]\n" + ".inst 0x6fa0e20b // udot v11.4s, v16.16b, v0.4b[1]\n" + ".inst 0x6fa1e20f // udot v15.4s, v16.16b, v1.4b[1]\n" + "ldr q16, [x10, #0x90]\n" + ".inst 0x6f80ea28 // udot v8.4s, v17.16b, v0.4b[2]\n" + ".inst 0x6f81ea2c // udot v12.4s, v17.16b, v1.4b[2]\n" + "ldr q17, [x10, #0xa0]\n" + ".inst 0x6f80ea09 // udot v9.4s, v16.16b, v0.4b[2]\n" + ".inst 0x6f81ea0d // udot v13.4s, v16.16b, v1.4b[2]\n" + "ldr q16, [x10, #0xb0]\n" + ".inst 0x6f80ea2a // udot v10.4s, v17.16b, v0.4b[2]\n" + ".inst 0x6f81ea2e // udot v14.4s, v17.16b, v1.4b[2]\n" + "ldr q17, [x10, #0xc0]\n" + ".inst 0x6f80ea0b // udot v11.4s, v16.16b, v0.4b[2]\n" + ".inst 0x6f81ea0f // udot v15.4s, v16.16b, v1.4b[2]\n" + "ldr q16, [x10, #0xd0]\n" + ".inst 0x6fa0ea28 // udot v8.4s, v17.16b, v0.4b[3]\n" + ".inst 0x6fa1ea2c // udot v12.4s, v17.16b, v1.4b[3]\n" + "ldr q17, [x10, #0xe0]\n" + ".inst 0x6fa0ea09 // udot v9.4s, v16.16b, v0.4b[3]\n" + ".inst 0x6fa1ea0d // udot v13.4s, v16.16b, v1.4b[3]\n" + "ldr q16, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" + ".inst 0x6fa0ea2a // udot v10.4s, v17.16b, v0.4b[3]\n" + ".inst 0x6fa1ea2e // udot v14.4s, v17.16b, v1.4b[3]\n" + ".inst 0x6fa0ea0b // udot v11.4s, v16.16b, v0.4b[3]\n" + ".inst 0x6fa1ea0f // udot v15.4s, v16.16b, v1.4b[3]\n" "53:" // Height 2: Multiply loop: Main loop skip "cbz x27, 58f\n" "cmp x27, #0x4\n" "blt 55f\n" "54:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr s1, [x25], #0x4\n" + "ldr s19, [x26], #0x4\n" + "ldr s18, [x25], #0x4\n" "sub x27, x27, #0x4\n" "cmp x27, #0x4\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + "ldr q17, [x10, #0x0]\n" + "ldr q16, [x10, #0x10]\n" + ".inst 0x6f93e228 // udot v8.4s, v17.16b, v19.4b[0]\n" + ".inst 0x6f92e22c // udot v12.4s, v17.16b, v18.4b[0]\n" + "ldr q17, [x10, #0x20]\n" + ".inst 0x6f93e209 // udot v9.4s, v16.16b, v19.4b[0]\n" + ".inst 0x6f92e20d // udot v13.4s, v16.16b, v18.4b[0]\n" + "ldr q16, [x10, #0x30]\n" + ".inst 0x6f93e22a // udot v10.4s, v17.16b, v19.4b[0]\n" + ".inst 0x6f92e22e // udot v14.4s, v17.16b, v18.4b[0]\n" "add x10, x10, #0x40\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" + ".inst 0x6f93e20b // udot v11.4s, v16.16b, v19.4b[0]\n" + ".inst 0x6f92e20f // udot v15.4s, v16.16b, v18.4b[0]\n" "bge 54b\n" "55:" // Height 2: Multiply loop: Skip odd blocks "cbz x27, 58f\n" @@ -636,19 +635,19 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr b0, [x26, #0x0]\n" "ldr b1, [x25, #0x0]\n" "57:" // Height 2: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - "ldr q7, [x10, #0x30]\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + "ldr q17, [x10, #0x0]\n" + "ldr q16, [x10, #0x10]\n" + ".inst 0x6f80e228 // udot v8.4s, v17.16b, v0.4b[0]\n" + ".inst 0x6f81e22c // udot v12.4s, v17.16b, v1.4b[0]\n" + "ldr q17, [x10, #0x20]\n" + ".inst 0x6f80e209 // udot v9.4s, v16.16b, v0.4b[0]\n" + ".inst 0x6f81e20d // udot v13.4s, v16.16b, v1.4b[0]\n" + "ldr q16, [x10, #0x30]\n" + ".inst 0x6f80e22a // udot v10.4s, v17.16b, v0.4b[0]\n" + ".inst 0x6f81e22e // udot v14.4s, v17.16b, v1.4b[0]\n" "add x10, x10, #0x40\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" + ".inst 0x6f80e20b // udot v11.4s, v16.16b, v0.4b[0]\n" + ".inst 0x6f81e20f // udot 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 x28, x28, #0x1\n" @@ -866,13 +865,13 @@ void a64_hybrid_u8u32_dot_6x16 ( "82:" // Height 3: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 83f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" "cbnz x28, 84f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -881,8 +880,8 @@ void a64_hybrid_u8u32_dot_6x16 ( "b 84f\n" "83:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" "84:" // Height 3: input setup done "cmp x27, #0x10\n" "blt 87f\n" @@ -899,75 +898,75 @@ void a64_hybrid_u8u32_dot_6x16 ( "sub x27, x27, #0x10\n" "add x26, x26, #0x10\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q21, [x10, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" "add x25, x25, #0x10\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q20, [x10, #0x30]\n" "add x24, x24, #0x10\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + ".inst 0x6f80e2aa // udot v10.4s, v21.16b, v0.4b[0]\n" + ".inst 0x6f81e2ae // udot v14.4s, v21.16b, v1.4b[0]\n" "cmp x27, #0x20\n" "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + ".inst 0x6f82e2b2 // udot v18.4s, v21.16b, v2.4b[0]\n" + "ldr q21, [x10, #0x40]\n" + ".inst 0x6f80e28b // udot v11.4s, v20.16b, v0.4b[0]\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x10, #0x50]\n" + ".inst 0x6f81e28f // udot v15.4s, v20.16b, v1.4b[0]\n" + ".inst 0x6f82e293 // udot v19.4s, v20.16b, v2.4b[0]\n" + "ldr q20, [x10, #0x50]\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" - ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" - "ldr q7, [x10, #0xf0]\n" + ".inst 0x6fa0e2a8 // udot v8.4s, v21.16b, v0.4b[1]\n" + ".inst 0x6fa1e2ac // udot v12.4s, v21.16b, v1.4b[1]\n" + ".inst 0x6fa2e2b0 // udot v16.4s, v21.16b, v2.4b[1]\n" + "ldr q21, [x10, #0x60]\n" + ".inst 0x6fa0e289 // udot v9.4s, v20.16b, v0.4b[1]\n" + ".inst 0x6fa1e28d // udot v13.4s, v20.16b, v1.4b[1]\n" + ".inst 0x6fa2e291 // udot v17.4s, v20.16b, v2.4b[1]\n" + "ldr q20, [x10, #0x70]\n" + ".inst 0x6fa0e2aa // udot v10.4s, v21.16b, v0.4b[1]\n" + ".inst 0x6fa1e2ae // udot v14.4s, v21.16b, v1.4b[1]\n" + ".inst 0x6fa2e2b2 // udot v18.4s, v21.16b, v2.4b[1]\n" + "ldr q21, [x10, #0x80]\n" + ".inst 0x6fa0e28b // udot v11.4s, v20.16b, v0.4b[1]\n" + ".inst 0x6fa1e28f // udot v15.4s, v20.16b, v1.4b[1]\n" + ".inst 0x6fa2e293 // udot v19.4s, v20.16b, v2.4b[1]\n" + "ldr q20, [x10, #0x90]\n" + ".inst 0x6f80eaa8 // udot v8.4s, v21.16b, v0.4b[2]\n" + ".inst 0x6f81eaac // udot v12.4s, v21.16b, v1.4b[2]\n" + ".inst 0x6f82eab0 // udot v16.4s, v21.16b, v2.4b[2]\n" + "ldr q21, [x10, #0xa0]\n" + ".inst 0x6f80ea89 // udot v9.4s, v20.16b, v0.4b[2]\n" + ".inst 0x6f81ea8d // udot v13.4s, v20.16b, v1.4b[2]\n" + ".inst 0x6f82ea91 // udot v17.4s, v20.16b, v2.4b[2]\n" + "ldr q20, [x10, #0xb0]\n" + ".inst 0x6f80eaaa // udot v10.4s, v21.16b, v0.4b[2]\n" + ".inst 0x6f81eaae // udot v14.4s, v21.16b, v1.4b[2]\n" + ".inst 0x6f82eab2 // udot v18.4s, v21.16b, v2.4b[2]\n" + "ldr q21, [x10, #0xc0]\n" + ".inst 0x6f80ea8b // udot v11.4s, v20.16b, v0.4b[2]\n" + ".inst 0x6f81ea8f // udot v15.4s, v20.16b, v1.4b[2]\n" + ".inst 0x6f82ea93 // udot v19.4s, v20.16b, v2.4b[2]\n" + "ldr q20, [x10, #0xd0]\n" + ".inst 0x6fa0eaa8 // udot v8.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa1eaac // udot v12.4s, v21.16b, v1.4b[3]\n" + ".inst 0x6fa2eab0 // udot v16.4s, v21.16b, v2.4b[3]\n" + "ldr q21, [x10, #0xe0]\n" + ".inst 0x6fa0ea89 // udot v9.4s, v20.16b, v0.4b[3]\n" + ".inst 0x6fa1ea8d // udot v13.4s, v20.16b, v1.4b[3]\n" + ".inst 0x6fa2ea91 // udot v17.4s, v20.16b, v2.4b[3]\n" + "ldr q20, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" + ".inst 0x6fa0eaaa // udot v10.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa1eaae // udot v14.4s, v21.16b, v1.4b[3]\n" + ".inst 0x6fa2eab2 // udot v18.4s, v21.16b, v2.4b[3]\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" + ".inst 0x6fa0ea8b // udot v11.4s, v20.16b, v0.4b[3]\n" "ldr q0, [x26, #0x0]\n" - ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" + ".inst 0x6fa1ea8f // udot v15.4s, v20.16b, v1.4b[3]\n" "ldr q1, [x25, #0x0]\n" - ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" + ".inst 0x6fa2ea93 // udot v19.4s, v20.16b, v2.4b[3]\n" "ldr q2, [x24, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 85b\n" @@ -977,98 +976,98 @@ void a64_hybrid_u8u32_dot_6x16 ( "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q21, [x10, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" "add x24, x24, #0x10\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q20, [x10, #0x30]\n" "sub x27, x27, #0x10\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + ".inst 0x6f80e2aa // udot v10.4s, v21.16b, v0.4b[0]\n" + ".inst 0x6f81e2ae // udot v14.4s, v21.16b, v1.4b[0]\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" + ".inst 0x6f82e2b2 // udot v18.4s, v21.16b, v2.4b[0]\n" + "ldr q21, [x10, #0x40]\n" + ".inst 0x6f80e28b // udot v11.4s, v20.16b, v0.4b[0]\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" - ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" - "ldr q7, [x10, #0xf0]\n" + ".inst 0x6f81e28f // udot v15.4s, v20.16b, v1.4b[0]\n" + ".inst 0x6f82e293 // udot v19.4s, v20.16b, v2.4b[0]\n" + "ldr q20, [x10, #0x50]\n" + ".inst 0x6fa0e2a8 // udot v8.4s, v21.16b, v0.4b[1]\n" + ".inst 0x6fa1e2ac // udot v12.4s, v21.16b, v1.4b[1]\n" + ".inst 0x6fa2e2b0 // udot v16.4s, v21.16b, v2.4b[1]\n" + "ldr q21, [x10, #0x60]\n" + ".inst 0x6fa0e289 // udot v9.4s, v20.16b, v0.4b[1]\n" + ".inst 0x6fa1e28d // udot v13.4s, v20.16b, v1.4b[1]\n" + ".inst 0x6fa2e291 // udot v17.4s, v20.16b, v2.4b[1]\n" + "ldr q20, [x10, #0x70]\n" + ".inst 0x6fa0e2aa // udot v10.4s, v21.16b, v0.4b[1]\n" + ".inst 0x6fa1e2ae // udot v14.4s, v21.16b, v1.4b[1]\n" + ".inst 0x6fa2e2b2 // udot v18.4s, v21.16b, v2.4b[1]\n" + "ldr q21, [x10, #0x80]\n" + ".inst 0x6fa0e28b // udot v11.4s, v20.16b, v0.4b[1]\n" + ".inst 0x6fa1e28f // udot v15.4s, v20.16b, v1.4b[1]\n" + ".inst 0x6fa2e293 // udot v19.4s, v20.16b, v2.4b[1]\n" + "ldr q20, [x10, #0x90]\n" + ".inst 0x6f80eaa8 // udot v8.4s, v21.16b, v0.4b[2]\n" + ".inst 0x6f81eaac // udot v12.4s, v21.16b, v1.4b[2]\n" + ".inst 0x6f82eab0 // udot v16.4s, v21.16b, v2.4b[2]\n" + "ldr q21, [x10, #0xa0]\n" + ".inst 0x6f80ea89 // udot v9.4s, v20.16b, v0.4b[2]\n" + ".inst 0x6f81ea8d // udot v13.4s, v20.16b, v1.4b[2]\n" + ".inst 0x6f82ea91 // udot v17.4s, v20.16b, v2.4b[2]\n" + "ldr q20, [x10, #0xb0]\n" + ".inst 0x6f80eaaa // udot v10.4s, v21.16b, v0.4b[2]\n" + ".inst 0x6f81eaae // udot v14.4s, v21.16b, v1.4b[2]\n" + ".inst 0x6f82eab2 // udot v18.4s, v21.16b, v2.4b[2]\n" + "ldr q21, [x10, #0xc0]\n" + ".inst 0x6f80ea8b // udot v11.4s, v20.16b, v0.4b[2]\n" + ".inst 0x6f81ea8f // udot v15.4s, v20.16b, v1.4b[2]\n" + ".inst 0x6f82ea93 // udot v19.4s, v20.16b, v2.4b[2]\n" + "ldr q20, [x10, #0xd0]\n" + ".inst 0x6fa0eaa8 // udot v8.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa1eaac // udot v12.4s, v21.16b, v1.4b[3]\n" + ".inst 0x6fa2eab0 // udot v16.4s, v21.16b, v2.4b[3]\n" + "ldr q21, [x10, #0xe0]\n" + ".inst 0x6fa0ea89 // udot v9.4s, v20.16b, v0.4b[3]\n" + ".inst 0x6fa1ea8d // udot v13.4s, v20.16b, v1.4b[3]\n" + ".inst 0x6fa2ea91 // udot v17.4s, v20.16b, v2.4b[3]\n" + "ldr q20, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" - ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" - ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" + ".inst 0x6fa0eaaa // udot v10.4s, v21.16b, v0.4b[3]\n" + ".inst 0x6fa1eaae // udot v14.4s, v21.16b, v1.4b[3]\n" + ".inst 0x6fa2eab2 // udot v18.4s, v21.16b, v2.4b[3]\n" + ".inst 0x6fa0ea8b // udot v11.4s, v20.16b, v0.4b[3]\n" + ".inst 0x6fa1ea8f // udot v15.4s, v20.16b, v1.4b[3]\n" + ".inst 0x6fa2ea93 // udot v19.4s, v20.16b, v2.4b[3]\n" "87:" // Height 3: Multiply loop: Main loop skip "cbz x27, 92f\n" "cmp x27, #0x4\n" "blt 89f\n" "88:" // Height 3: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr s1, [x25], #0x4\n" + "ldr s24, [x26], #0x4\n" + "ldr s23, [x25], #0x4\n" "sub x27, x27, #0x4\n" "cmp x27, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr q6, [x10, #0x0]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr s22, [x24], #0x4\n" + "ldr q21, [x10, #0x0]\n" + ".inst 0x6f98e2a8 // udot v8.4s, v21.16b, v24.4b[0]\n" + ".inst 0x6f97e2ac // udot v12.4s, v21.16b, v23.4b[0]\n" + "ldr q20, [x10, #0x10]\n" + ".inst 0x6f96e2b0 // udot v16.4s, v21.16b, v22.4b[0]\n" + "ldr q21, [x10, #0x20]\n" + ".inst 0x6f98e289 // udot v9.4s, v20.16b, v24.4b[0]\n" + ".inst 0x6f97e28d // udot v13.4s, v20.16b, v23.4b[0]\n" + ".inst 0x6f96e291 // udot v17.4s, v20.16b, v22.4b[0]\n" + "ldr q20, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" + ".inst 0x6f98e2aa // udot v10.4s, v21.16b, v24.4b[0]\n" + ".inst 0x6f97e2ae // udot v14.4s, v21.16b, v23.4b[0]\n" + ".inst 0x6f96e2b2 // udot v18.4s, v21.16b, v22.4b[0]\n" + ".inst 0x6f98e28b // udot v11.4s, v20.16b, v24.4b[0]\n" + ".inst 0x6f97e28f // udot v15.4s, v20.16b, v23.4b[0]\n" + ".inst 0x6f96e293 // udot v19.4s, v20.16b, v22.4b[0]\n" "bge 88b\n" "89:" // Height 3: Multiply loop: Skip odd blocks "cbz x27, 92f\n" @@ -1086,23 +1085,23 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr b1, [x25, #0x0]\n" "ldr b2, [x24, #0x0]\n" "91:" // Height 3: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q21, [x10, #0x0]\n" + "ldr q20, [x10, #0x10]\n" + ".inst 0x6f80e2a8 // udot v8.4s, v21.16b, v0.4b[0]\n" + ".inst 0x6f81e2ac // udot v12.4s, v21.16b, v1.4b[0]\n" + ".inst 0x6f82e2b0 // udot v16.4s, v21.16b, v2.4b[0]\n" + "ldr q21, [x10, #0x20]\n" + ".inst 0x6f80e289 // udot v9.4s, v20.16b, v0.4b[0]\n" + ".inst 0x6f81e28d // udot v13.4s, v20.16b, v1.4b[0]\n" + ".inst 0x6f82e291 // udot v17.4s, v20.16b, v2.4b[0]\n" + "ldr q20, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" + ".inst 0x6f80e2aa // udot v10.4s, v21.16b, v0.4b[0]\n" + ".inst 0x6f81e2ae // udot v14.4s, v21.16b, v1.4b[0]\n" + ".inst 0x6f82e2b2 // udot v18.4s, v21.16b, v2.4b[0]\n" + ".inst 0x6f80e28b // udot v11.4s, v20.16b, v0.4b[0]\n" + ".inst 0x6f81e28f // udot v15.4s, v20.16b, v1.4b[0]\n" + ".inst 0x6f82e293 // udot 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 x28, x28, #0x1\n" @@ -1367,14 +1366,14 @@ void a64_hybrid_u8u32_dot_6x16 ( "116:" // Height 4: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 117f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" "cbnz x28, 118f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -1384,9 +1383,9 @@ void a64_hybrid_u8u32_dot_6x16 ( "b 118f\n" "117:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" - "add x23, x24, x20\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" + "add x23, x24, x21\n" "118:" // Height 4: input setup done "cmp x27, #0x10\n" "blt 121f\n" @@ -1405,7 +1404,7 @@ void a64_hybrid_u8u32_dot_6x16 ( "add x26, x26, #0x10\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q25, [x10, #0x20]\n" "add x25, x25, #0x10\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" @@ -1413,85 +1412,85 @@ void a64_hybrid_u8u32_dot_6x16 ( "add x23, x23, #0x10\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q24, [x10, #0x30]\n" "cmp x27, #0x20\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + ".inst 0x6f80e32a // udot v10.4s, v25.16b, v0.4b[0]\n" + ".inst 0x6f81e32e // udot v14.4s, v25.16b, v1.4b[0]\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x10, #0x40]\n" + ".inst 0x6f82e332 // udot v18.4s, v25.16b, v2.4b[0]\n" + ".inst 0x6f83e336 // udot v22.4s, v25.16b, v3.4b[0]\n" + "ldr q25, [x10, #0x40]\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" + ".inst 0x6f80e30b // udot v11.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f81e30f // udot v15.4s, v24.16b, v1.4b[0]\n" "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" - ".inst 0x6fa3e0d4 // udot v20.4s, v6.16b, v3.4b[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" - ".inst 0x6fa3e0f5 // udot v21.4s, v7.16b, v3.4b[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" - ".inst 0x6fa3e0d6 // udot v22.4s, v6.16b, v3.4b[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" - ".inst 0x6fa3e0f7 // udot v23.4s, v7.16b, v3.4b[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" - ".inst 0x6f83e8d4 // udot v20.4s, v6.16b, v3.4b[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" - ".inst 0x6f83e8f5 // udot v21.4s, v7.16b, v3.4b[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" - ".inst 0x6f83e8d6 // udot v22.4s, v6.16b, v3.4b[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" - ".inst 0x6f83e8f7 // udot v23.4s, v7.16b, v3.4b[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" - ".inst 0x6fa3e8d4 // udot v20.4s, v6.16b, v3.4b[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" - ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" - ".inst 0x6fa3e8f5 // udot v21.4s, v7.16b, v3.4b[3]\n" - "ldr q7, [x10, #0xf0]\n" + ".inst 0x6f82e313 // udot v19.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f83e317 // udot v23.4s, v24.16b, v3.4b[0]\n" + "ldr q24, [x10, #0x50]\n" + ".inst 0x6fa0e328 // udot v8.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa1e32c // udot v12.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa2e330 // udot v16.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa3e334 // udot v20.4s, v25.16b, v3.4b[1]\n" + "ldr q25, [x10, #0x60]\n" + ".inst 0x6fa0e309 // udot v9.4s, v24.16b, v0.4b[1]\n" + ".inst 0x6fa1e30d // udot v13.4s, v24.16b, v1.4b[1]\n" + ".inst 0x6fa2e311 // udot v17.4s, v24.16b, v2.4b[1]\n" + ".inst 0x6fa3e315 // udot v21.4s, v24.16b, v3.4b[1]\n" + "ldr q24, [x10, #0x70]\n" + ".inst 0x6fa0e32a // udot v10.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa1e32e // udot v14.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa2e332 // udot v18.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa3e336 // udot v22.4s, v25.16b, v3.4b[1]\n" + "ldr q25, [x10, #0x80]\n" + ".inst 0x6fa0e30b // udot v11.4s, v24.16b, v0.4b[1]\n" + ".inst 0x6fa1e30f // udot v15.4s, v24.16b, v1.4b[1]\n" + ".inst 0x6fa2e313 // udot v19.4s, v24.16b, v2.4b[1]\n" + ".inst 0x6fa3e317 // udot v23.4s, v24.16b, v3.4b[1]\n" + "ldr q24, [x10, #0x90]\n" + ".inst 0x6f80eb28 // udot v8.4s, v25.16b, v0.4b[2]\n" + ".inst 0x6f81eb2c // udot v12.4s, v25.16b, v1.4b[2]\n" + ".inst 0x6f82eb30 // udot v16.4s, v25.16b, v2.4b[2]\n" + ".inst 0x6f83eb34 // udot v20.4s, v25.16b, v3.4b[2]\n" + "ldr q25, [x10, #0xa0]\n" + ".inst 0x6f80eb09 // udot v9.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f81eb0d // udot v13.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f82eb11 // udot v17.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f83eb15 // udot v21.4s, v24.16b, v3.4b[2]\n" + "ldr q24, [x10, #0xb0]\n" + ".inst 0x6f80eb2a // udot v10.4s, v25.16b, v0.4b[2]\n" + ".inst 0x6f81eb2e // udot v14.4s, v25.16b, v1.4b[2]\n" + ".inst 0x6f82eb32 // udot v18.4s, v25.16b, v2.4b[2]\n" + ".inst 0x6f83eb36 // udot v22.4s, v25.16b, v3.4b[2]\n" + "ldr q25, [x10, #0xc0]\n" + ".inst 0x6f80eb0b // udot v11.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f81eb0f // udot v15.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f82eb13 // udot v19.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f83eb17 // udot v23.4s, v24.16b, v3.4b[2]\n" + "ldr q24, [x10, #0xd0]\n" + ".inst 0x6fa0eb28 // udot v8.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa1eb2c // udot v12.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa2eb30 // udot v16.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa3eb34 // udot v20.4s, v25.16b, v3.4b[3]\n" + "ldr q25, [x10, #0xe0]\n" + ".inst 0x6fa0eb09 // udot v9.4s, v24.16b, v0.4b[3]\n" + ".inst 0x6fa1eb0d // udot v13.4s, v24.16b, v1.4b[3]\n" + ".inst 0x6fa2eb11 // udot v17.4s, v24.16b, v2.4b[3]\n" + ".inst 0x6fa3eb15 // udot v21.4s, v24.16b, v3.4b[3]\n" + "ldr q24, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" - ".inst 0x6fa3e8d6 // udot v22.4s, v6.16b, v3.4b[3]\n" + ".inst 0x6fa0eb2a // udot v10.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa1eb2e // udot v14.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa2eb32 // udot v18.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa3eb36 // udot v22.4s, v25.16b, v3.4b[3]\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" + ".inst 0x6fa0eb0b // udot v11.4s, v24.16b, v0.4b[3]\n" "ldr q0, [x26, #0x0]\n" - ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" + ".inst 0x6fa1eb0f // udot v15.4s, v24.16b, v1.4b[3]\n" "ldr q1, [x25, #0x0]\n" - ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" + ".inst 0x6fa2eb13 // udot v19.4s, v24.16b, v2.4b[3]\n" "ldr q2, [x24, #0x0]\n" - ".inst 0x6fa3e8f7 // udot v23.4s, v7.16b, v3.4b[3]\n" + ".inst 0x6fa3eb17 // udot v23.4s, v24.16b, v3.4b[3]\n" "ldr q3, [x23, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 119b\n" @@ -1502,7 +1501,7 @@ void a64_hybrid_u8u32_dot_6x16 ( "add x25, x25, #0x10\n" ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q25, [x10, #0x20]\n" "add x24, x24, #0x10\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" @@ -1510,112 +1509,112 @@ void a64_hybrid_u8u32_dot_6x16 ( "sub x27, x27, #0x10\n" ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q24, [x10, #0x30]\n" "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + ".inst 0x6f80e32a // udot v10.4s, v25.16b, v0.4b[0]\n" + ".inst 0x6f81e32e // udot v14.4s, v25.16b, v1.4b[0]\n" "prfm pldl1keep, [x25, #0x80]\n" "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x10, #0x40]\n" + ".inst 0x6f82e332 // udot v18.4s, v25.16b, v2.4b[0]\n" + ".inst 0x6f83e336 // udot v22.4s, v25.16b, v3.4b[0]\n" + "ldr q25, [x10, #0x40]\n" "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" - ".inst 0x6fa3e0d4 // udot v20.4s, v6.16b, v3.4b[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" - ".inst 0x6fa3e0f5 // udot v21.4s, v7.16b, v3.4b[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" - ".inst 0x6fa3e0d6 // udot v22.4s, v6.16b, v3.4b[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" - ".inst 0x6fa3e0f7 // udot v23.4s, v7.16b, v3.4b[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" - ".inst 0x6f83e8d4 // udot v20.4s, v6.16b, v3.4b[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" - ".inst 0x6f83e8f5 // udot v21.4s, v7.16b, v3.4b[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" - ".inst 0x6f83e8d6 // udot v22.4s, v6.16b, v3.4b[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" - ".inst 0x6f83e8f7 // udot v23.4s, v7.16b, v3.4b[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" - ".inst 0x6fa3e8d4 // udot v20.4s, v6.16b, v3.4b[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" - ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" - ".inst 0x6fa3e8f5 // udot v21.4s, v7.16b, v3.4b[3]\n" - "ldr q7, [x10, #0xf0]\n" + ".inst 0x6f80e30b // udot v11.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f81e30f // udot v15.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f82e313 // udot v19.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f83e317 // udot v23.4s, v24.16b, v3.4b[0]\n" + "ldr q24, [x10, #0x50]\n" + ".inst 0x6fa0e328 // udot v8.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa1e32c // udot v12.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa2e330 // udot v16.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa3e334 // udot v20.4s, v25.16b, v3.4b[1]\n" + "ldr q25, [x10, #0x60]\n" + ".inst 0x6fa0e309 // udot v9.4s, v24.16b, v0.4b[1]\n" + ".inst 0x6fa1e30d // udot v13.4s, v24.16b, v1.4b[1]\n" + ".inst 0x6fa2e311 // udot v17.4s, v24.16b, v2.4b[1]\n" + ".inst 0x6fa3e315 // udot v21.4s, v24.16b, v3.4b[1]\n" + "ldr q24, [x10, #0x70]\n" + ".inst 0x6fa0e32a // udot v10.4s, v25.16b, v0.4b[1]\n" + ".inst 0x6fa1e32e // udot v14.4s, v25.16b, v1.4b[1]\n" + ".inst 0x6fa2e332 // udot v18.4s, v25.16b, v2.4b[1]\n" + ".inst 0x6fa3e336 // udot v22.4s, v25.16b, v3.4b[1]\n" + "ldr q25, [x10, #0x80]\n" + ".inst 0x6fa0e30b // udot v11.4s, v24.16b, v0.4b[1]\n" + ".inst 0x6fa1e30f // udot v15.4s, v24.16b, v1.4b[1]\n" + ".inst 0x6fa2e313 // udot v19.4s, v24.16b, v2.4b[1]\n" + ".inst 0x6fa3e317 // udot v23.4s, v24.16b, v3.4b[1]\n" + "ldr q24, [x10, #0x90]\n" + ".inst 0x6f80eb28 // udot v8.4s, v25.16b, v0.4b[2]\n" + ".inst 0x6f81eb2c // udot v12.4s, v25.16b, v1.4b[2]\n" + ".inst 0x6f82eb30 // udot v16.4s, v25.16b, v2.4b[2]\n" + ".inst 0x6f83eb34 // udot v20.4s, v25.16b, v3.4b[2]\n" + "ldr q25, [x10, #0xa0]\n" + ".inst 0x6f80eb09 // udot v9.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f81eb0d // udot v13.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f82eb11 // udot v17.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f83eb15 // udot v21.4s, v24.16b, v3.4b[2]\n" + "ldr q24, [x10, #0xb0]\n" + ".inst 0x6f80eb2a // udot v10.4s, v25.16b, v0.4b[2]\n" + ".inst 0x6f81eb2e // udot v14.4s, v25.16b, v1.4b[2]\n" + ".inst 0x6f82eb32 // udot v18.4s, v25.16b, v2.4b[2]\n" + ".inst 0x6f83eb36 // udot v22.4s, v25.16b, v3.4b[2]\n" + "ldr q25, [x10, #0xc0]\n" + ".inst 0x6f80eb0b // udot v11.4s, v24.16b, v0.4b[2]\n" + ".inst 0x6f81eb0f // udot v15.4s, v24.16b, v1.4b[2]\n" + ".inst 0x6f82eb13 // udot v19.4s, v24.16b, v2.4b[2]\n" + ".inst 0x6f83eb17 // udot v23.4s, v24.16b, v3.4b[2]\n" + "ldr q24, [x10, #0xd0]\n" + ".inst 0x6fa0eb28 // udot v8.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa1eb2c // udot v12.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa2eb30 // udot v16.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa3eb34 // udot v20.4s, v25.16b, v3.4b[3]\n" + "ldr q25, [x10, #0xe0]\n" + ".inst 0x6fa0eb09 // udot v9.4s, v24.16b, v0.4b[3]\n" + ".inst 0x6fa1eb0d // udot v13.4s, v24.16b, v1.4b[3]\n" + ".inst 0x6fa2eb11 // udot v17.4s, v24.16b, v2.4b[3]\n" + ".inst 0x6fa3eb15 // udot v21.4s, v24.16b, v3.4b[3]\n" + "ldr q24, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" - ".inst 0x6fa3e8d6 // udot v22.4s, v6.16b, v3.4b[3]\n" - ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" - ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" - ".inst 0x6fa3e8f7 // udot v23.4s, v7.16b, v3.4b[3]\n" + ".inst 0x6fa0eb2a // udot v10.4s, v25.16b, v0.4b[3]\n" + ".inst 0x6fa1eb2e // udot v14.4s, v25.16b, v1.4b[3]\n" + ".inst 0x6fa2eb32 // udot v18.4s, v25.16b, v2.4b[3]\n" + ".inst 0x6fa3eb36 // udot v22.4s, v25.16b, v3.4b[3]\n" + ".inst 0x6fa0eb0b // udot v11.4s, v24.16b, v0.4b[3]\n" + ".inst 0x6fa1eb0f // udot v15.4s, v24.16b, v1.4b[3]\n" + ".inst 0x6fa2eb13 // udot v19.4s, v24.16b, v2.4b[3]\n" + ".inst 0x6fa3eb17 // udot v23.4s, v24.16b, v3.4b[3]\n" "121:" // Height 4: Multiply loop: Main loop skip "cbz x27, 126f\n" "cmp x27, #0x4\n" "blt 123f\n" "122:" // Height 4: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr s1, [x25], #0x4\n" + "ldr s29, [x26], #0x4\n" + "ldr s28, [x25], #0x4\n" "sub x27, x27, #0x4\n" "cmp x27, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr s27, [x24], #0x4\n" + "ldr s26, [x23], #0x4\n" + "ldr q25, [x10, #0x0]\n" + "ldr q24, [x10, #0x10]\n" + ".inst 0x6f9de328 // udot v8.4s, v25.16b, v29.4b[0]\n" + ".inst 0x6f9ce32c // udot v12.4s, v25.16b, v28.4b[0]\n" + ".inst 0x6f9be330 // udot v16.4s, v25.16b, v27.4b[0]\n" + ".inst 0x6f9ae334 // udot v20.4s, v25.16b, v26.4b[0]\n" + "ldr q25, [x10, #0x20]\n" + ".inst 0x6f9de309 // udot v9.4s, v24.16b, v29.4b[0]\n" + ".inst 0x6f9ce30d // udot v13.4s, v24.16b, v28.4b[0]\n" + ".inst 0x6f9be311 // udot v17.4s, v24.16b, v27.4b[0]\n" + ".inst 0x6f9ae315 // udot v21.4s, v24.16b, v26.4b[0]\n" + "ldr q24, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" + ".inst 0x6f9de32a // udot v10.4s, v25.16b, v29.4b[0]\n" + ".inst 0x6f9ce32e // udot v14.4s, v25.16b, v28.4b[0]\n" + ".inst 0x6f9be332 // udot v18.4s, v25.16b, v27.4b[0]\n" + ".inst 0x6f9ae336 // udot v22.4s, v25.16b, v26.4b[0]\n" + ".inst 0x6f9de30b // udot v11.4s, v24.16b, v29.4b[0]\n" + ".inst 0x6f9ce30f // udot v15.4s, v24.16b, v28.4b[0]\n" + ".inst 0x6f9be313 // udot v19.4s, v24.16b, v27.4b[0]\n" + ".inst 0x6f9ae317 // udot v23.4s, v24.16b, v26.4b[0]\n" "bge 122b\n" "123:" // Height 4: Multiply loop: Skip odd blocks "cbz x27, 126f\n" @@ -1636,27 +1635,27 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr b2, [x24, #0x0]\n" "ldr b3, [x23, #0x0]\n" "125:" // Height 4: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q25, [x10, #0x0]\n" + "ldr q24, [x10, #0x10]\n" + ".inst 0x6f80e328 // udot v8.4s, v25.16b, v0.4b[0]\n" + ".inst 0x6f81e32c // udot v12.4s, v25.16b, v1.4b[0]\n" + ".inst 0x6f82e330 // udot v16.4s, v25.16b, v2.4b[0]\n" + ".inst 0x6f83e334 // udot v20.4s, v25.16b, v3.4b[0]\n" + "ldr q25, [x10, #0x20]\n" + ".inst 0x6f80e309 // udot v9.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f81e30d // udot v13.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f82e311 // udot v17.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f83e315 // udot v21.4s, v24.16b, v3.4b[0]\n" + "ldr q24, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" + ".inst 0x6f80e32a // udot v10.4s, v25.16b, v0.4b[0]\n" + ".inst 0x6f81e32e // udot v14.4s, v25.16b, v1.4b[0]\n" + ".inst 0x6f82e332 // udot v18.4s, v25.16b, v2.4b[0]\n" + ".inst 0x6f83e336 // udot v22.4s, v25.16b, v3.4b[0]\n" + ".inst 0x6f80e30b // udot v11.4s, v24.16b, v0.4b[0]\n" + ".inst 0x6f81e30f // udot v15.4s, v24.16b, v1.4b[0]\n" + ".inst 0x6f82e313 // udot v19.4s, v24.16b, v2.4b[0]\n" + ".inst 0x6f83e317 // udot 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 x28, x28, #0x1\n" @@ -1960,162 +1959,162 @@ void a64_hybrid_u8u32_dot_6x16 ( "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 x28, #0x0\n" - "150:" // Height 5: String loop - "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" - "tbz %x[flags], #3, 151f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" - "cbnz x28, 152f\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x26, x26, x20\n" - "add x25, x25, x20\n" - "add x24, x24, x20\n" - "add x23, x23, x20\n" - "add x22, x22, x20\n" - "b 152f\n" - "151:" // Height 5: setup direct input - "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" - "add x23, x24, x20\n" - "add x22, x23, x20\n" - "152:" // Height 5: input setup done - "cmp x27, #0x10\n" - "blt 155f\n" - "ldr q0, [x26, #0x0]\n" - "ldr q1, [x25, #0x0]\n" - "cmp x27, #0x20\n" - "ldr q2, [x24, #0x0]\n" - "ldr q3, [x23, #0x0]\n" - "ldr q4, [x22, #0x0]\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - "blt 154f\n" - "153:" // Height 5: Multiply loop: Main loop head - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - "sub x27, x27, #0x10\n" - "add x26, x26, #0x10\n" - ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" - "add x25, x25, #0x10\n" - "add x24, x24, #0x10\n" - ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - "add x23, x23, #0x10\n" - ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - "add x22, x22, #0x10\n" - "cmp x27, #0x20\n" - ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x10, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" - "prfm pldl1keep, [x24, #0x80]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "prfm pldl1keep, [x22, #0x80]\n" - ".inst 0x6f84e0da // udot v26.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" - ".inst 0x6fa3e0d4 // udot v20.4s, v6.16b, v3.4b[1]\n" - ".inst 0x6fa4e0d8 // udot v24.4s, v6.16b, v4.4b[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" - ".inst 0x6fa3e0f5 // udot v21.4s, v7.16b, v3.4b[1]\n" - ".inst 0x6fa4e0f9 // udot v25.4s, v7.16b, v4.4b[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" - ".inst 0x6fa3e0d6 // udot v22.4s, v6.16b, v3.4b[1]\n" - ".inst 0x6fa4e0da // udot v26.4s, v6.16b, v4.4b[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" - ".inst 0x6fa3e0f7 // udot v23.4s, v7.16b, v3.4b[1]\n" - ".inst 0x6fa4e0fb // udot v27.4s, v7.16b, v4.4b[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" - ".inst 0x6f83e8d4 // udot v20.4s, v6.16b, v3.4b[2]\n" - ".inst 0x6f84e8d8 // udot v24.4s, v6.16b, v4.4b[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" - ".inst 0x6f83e8f5 // udot v21.4s, v7.16b, v3.4b[2]\n" - ".inst 0x6f84e8f9 // udot v25.4s, v7.16b, v4.4b[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" - ".inst 0x6f83e8d6 // udot v22.4s, v6.16b, v3.4b[2]\n" - ".inst 0x6f84e8da // udot v26.4s, v6.16b, v4.4b[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" - ".inst 0x6f83e8f7 // udot v23.4s, v7.16b, v3.4b[2]\n" - ".inst 0x6f84e8fb // udot v27.4s, v7.16b, v4.4b[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" - ".inst 0x6fa3e8d4 // udot v20.4s, v6.16b, v3.4b[3]\n" - ".inst 0x6fa4e8d8 // udot v24.4s, v6.16b, v4.4b[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" - ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" - ".inst 0x6fa3e8f5 // udot v21.4s, v7.16b, v3.4b[3]\n" - ".inst 0x6fa4e8f9 // udot v25.4s, v7.16b, v4.4b[3]\n" - "ldr q7, [x10, #0xf0]\n" + "movi v25.4s, #0x0\n" + "movi v26.4s, #0x0\n" + "movi v27.4s, #0x0\n" + "149:" // Height 5: setup done + "mov x28, #0x0\n" + "150:" // Height 5: String loop + "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "tbz %x[flags], #3, 151f\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" + "cbnz x28, 152f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20\n" + "add x25, x25, x20\n" + "add x24, x24, x20\n" + "add x23, x23, x20\n" + "add x22, x22, x20\n" + "b 152f\n" + "151:" // Height 5: setup direct input + "mov x26, %x[input_ptr]\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" + "add x23, x24, x21\n" + "add x22, x23, x21\n" + "152:" // Height 5: input setup done + "cmp x27, #0x10\n" + "blt 155f\n" + "ldr q0, [x26, #0x0]\n" + "ldr q1, [x25, #0x0]\n" + "cmp x27, #0x20\n" + "ldr q2, [x24, #0x0]\n" + "ldr q3, [x23, #0x0]\n" + "ldr q4, [x22, #0x0]\n" + "ldr q6, [x10, #0x0]\n" + "ldr q7, [x10, #0x10]\n" + "blt 154f\n" + "153:" // Height 5: Multiply loop: Main loop head + ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" + ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" + "sub x27, x27, #0x10\n" + "add x26, x26, #0x10\n" + ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" + ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" + "add x25, x25, #0x10\n" + "add x24, x24, #0x10\n" + ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" + "ldr q29, [x10, #0x20]\n" + ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" + "add x23, x23, #0x10\n" + ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" + ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" + "add x22, x22, #0x10\n" + "cmp x27, #0x20\n" + ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" + ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" + "ldr q28, [x10, #0x30]\n" + "prfm pldl1keep, [x26, #0x80]\n" + ".inst 0x6f80e3aa // udot v10.4s, v29.16b, v0.4b[0]\n" + ".inst 0x6f81e3ae // udot v14.4s, v29.16b, v1.4b[0]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" + ".inst 0x6f82e3b2 // udot v18.4s, v29.16b, v2.4b[0]\n" + ".inst 0x6f83e3b6 // udot v22.4s, v29.16b, v3.4b[0]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" + ".inst 0x6f84e3ba // udot v26.4s, v29.16b, v4.4b[0]\n" + "ldr q29, [x10, #0x40]\n" + ".inst 0x6f80e38b // udot v11.4s, v28.16b, v0.4b[0]\n" + ".inst 0x6f81e38f // udot v15.4s, v28.16b, v1.4b[0]\n" + ".inst 0x6f82e393 // udot v19.4s, v28.16b, v2.4b[0]\n" + ".inst 0x6f83e397 // udot v23.4s, v28.16b, v3.4b[0]\n" + ".inst 0x6f84e39b // udot v27.4s, v28.16b, v4.4b[0]\n" + "ldr q28, [x10, #0x50]\n" + ".inst 0x6fa0e3a8 // udot v8.4s, v29.16b, v0.4b[1]\n" + ".inst 0x6fa1e3ac // udot v12.4s, v29.16b, v1.4b[1]\n" + ".inst 0x6fa2e3b0 // udot v16.4s, v29.16b, v2.4b[1]\n" + ".inst 0x6fa3e3b4 // udot v20.4s, v29.16b, v3.4b[1]\n" + ".inst 0x6fa4e3b8 // udot v24.4s, v29.16b, v4.4b[1]\n" + "ldr q29, [x10, #0x60]\n" + ".inst 0x6fa0e389 // udot v9.4s, v28.16b, v0.4b[1]\n" + ".inst 0x6fa1e38d // udot v13.4s, v28.16b, v1.4b[1]\n" + ".inst 0x6fa2e391 // udot v17.4s, v28.16b, v2.4b[1]\n" + ".inst 0x6fa3e395 // udot v21.4s, v28.16b, v3.4b[1]\n" + ".inst 0x6fa4e399 // udot v25.4s, v28.16b, v4.4b[1]\n" + "ldr q28, [x10, #0x70]\n" + ".inst 0x6fa0e3aa // udot v10.4s, v29.16b, v0.4b[1]\n" + ".inst 0x6fa1e3ae // udot v14.4s, v29.16b, v1.4b[1]\n" + ".inst 0x6fa2e3b2 // udot v18.4s, v29.16b, v2.4b[1]\n" + ".inst 0x6fa3e3b6 // udot v22.4s, v29.16b, v3.4b[1]\n" + ".inst 0x6fa4e3ba // udot v26.4s, v29.16b, v4.4b[1]\n" + "ldr q29, [x10, #0x80]\n" + ".inst 0x6fa0e38b // udot v11.4s, v28.16b, v0.4b[1]\n" + ".inst 0x6fa1e38f // udot v15.4s, v28.16b, v1.4b[1]\n" + ".inst 0x6fa2e393 // udot v19.4s, v28.16b, v2.4b[1]\n" + ".inst 0x6fa3e397 // udot v23.4s, v28.16b, v3.4b[1]\n" + ".inst 0x6fa4e39b // udot v27.4s, v28.16b, v4.4b[1]\n" + "ldr q28, [x10, #0x90]\n" + ".inst 0x6f80eba8 // udot v8.4s, v29.16b, v0.4b[2]\n" + ".inst 0x6f81ebac // udot v12.4s, v29.16b, v1.4b[2]\n" + ".inst 0x6f82ebb0 // udot v16.4s, v29.16b, v2.4b[2]\n" + ".inst 0x6f83ebb4 // udot v20.4s, v29.16b, v3.4b[2]\n" + ".inst 0x6f84ebb8 // udot v24.4s, v29.16b, v4.4b[2]\n" + "ldr q29, [x10, #0xa0]\n" + ".inst 0x6f80eb89 // udot v9.4s, v28.16b, v0.4b[2]\n" + ".inst 0x6f81eb8d // udot v13.4s, v28.16b, v1.4b[2]\n" + ".inst 0x6f82eb91 // udot v17.4s, v28.16b, v2.4b[2]\n" + ".inst 0x6f83eb95 // udot v21.4s, v28.16b, v3.4b[2]\n" + ".inst 0x6f84eb99 // udot v25.4s, v28.16b, v4.4b[2]\n" + "ldr q28, [x10, #0xb0]\n" + ".inst 0x6f80ebaa // udot v10.4s, v29.16b, v0.4b[2]\n" + ".inst 0x6f81ebae // udot v14.4s, v29.16b, v1.4b[2]\n" + ".inst 0x6f82ebb2 // udot v18.4s, v29.16b, v2.4b[2]\n" + ".inst 0x6f83ebb6 // udot v22.4s, v29.16b, v3.4b[2]\n" + ".inst 0x6f84ebba // udot v26.4s, v29.16b, v4.4b[2]\n" + "ldr q29, [x10, #0xc0]\n" + ".inst 0x6f80eb8b // udot v11.4s, v28.16b, v0.4b[2]\n" + ".inst 0x6f81eb8f // udot v15.4s, v28.16b, v1.4b[2]\n" + ".inst 0x6f82eb93 // udot v19.4s, v28.16b, v2.4b[2]\n" + ".inst 0x6f83eb97 // udot v23.4s, v28.16b, v3.4b[2]\n" + ".inst 0x6f84eb9b // udot v27.4s, v28.16b, v4.4b[2]\n" + "ldr q28, [x10, #0xd0]\n" + ".inst 0x6fa0eba8 // udot v8.4s, v29.16b, v0.4b[3]\n" + ".inst 0x6fa1ebac // udot v12.4s, v29.16b, v1.4b[3]\n" + ".inst 0x6fa2ebb0 // udot v16.4s, v29.16b, v2.4b[3]\n" + ".inst 0x6fa3ebb4 // udot v20.4s, v29.16b, v3.4b[3]\n" + ".inst 0x6fa4ebb8 // udot v24.4s, v29.16b, v4.4b[3]\n" + "ldr q29, [x10, #0xe0]\n" + ".inst 0x6fa0eb89 // udot v9.4s, v28.16b, v0.4b[3]\n" + ".inst 0x6fa1eb8d // udot v13.4s, v28.16b, v1.4b[3]\n" + ".inst 0x6fa2eb91 // udot v17.4s, v28.16b, v2.4b[3]\n" + ".inst 0x6fa3eb95 // udot v21.4s, v28.16b, v3.4b[3]\n" + ".inst 0x6fa4eb99 // udot v25.4s, v28.16b, v4.4b[3]\n" + "ldr q28, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" - ".inst 0x6fa3e8d6 // udot v22.4s, v6.16b, v3.4b[3]\n" - ".inst 0x6fa4e8da // udot v26.4s, v6.16b, v4.4b[3]\n" + ".inst 0x6fa0ebaa // udot v10.4s, v29.16b, v0.4b[3]\n" + ".inst 0x6fa1ebae // udot v14.4s, v29.16b, v1.4b[3]\n" + ".inst 0x6fa2ebb2 // udot v18.4s, v29.16b, v2.4b[3]\n" + ".inst 0x6fa3ebb6 // udot v22.4s, v29.16b, v3.4b[3]\n" + ".inst 0x6fa4ebba // udot v26.4s, v29.16b, v4.4b[3]\n" "ldr q6, [x10, #0x0]\n" - ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" + ".inst 0x6fa0eb8b // udot v11.4s, v28.16b, v0.4b[3]\n" "ldr q0, [x26, #0x0]\n" - ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" + ".inst 0x6fa1eb8f // udot v15.4s, v28.16b, v1.4b[3]\n" "ldr q1, [x25, #0x0]\n" - ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" + ".inst 0x6fa2eb93 // udot v19.4s, v28.16b, v2.4b[3]\n" "ldr q2, [x24, #0x0]\n" - ".inst 0x6fa3e8f7 // udot v23.4s, v7.16b, v3.4b[3]\n" + ".inst 0x6fa3eb97 // udot v23.4s, v28.16b, v3.4b[3]\n" "ldr q3, [x23, #0x0]\n" - ".inst 0x6fa4e8fb // udot v27.4s, v7.16b, v4.4b[3]\n" + ".inst 0x6fa4eb9b // udot v27.4s, v28.16b, v4.4b[3]\n" "ldr q4, [x22, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 153b\n" @@ -2129,7 +2128,7 @@ void a64_hybrid_u8u32_dot_6x16 ( "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q29, [x10, #0x20]\n" ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" "add x22, x22, #0x10\n" ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" @@ -2138,131 +2137,131 @@ void a64_hybrid_u8u32_dot_6x16 ( "prfm pldl1keep, [x26, #0x80]\n" ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q28, [x10, #0x30]\n" "prfm pldl1keep, [x25, #0x80]\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" + ".inst 0x6f80e3aa // udot v10.4s, v29.16b, v0.4b[0]\n" + ".inst 0x6f81e3ae // udot v14.4s, v29.16b, v1.4b[0]\n" "prfm pldl1keep, [x24, #0x80]\n" "prfm pldl1keep, [x23, #0x80]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" + ".inst 0x6f82e3b2 // udot v18.4s, v29.16b, v2.4b[0]\n" + ".inst 0x6f83e3b6 // udot v22.4s, v29.16b, v3.4b[0]\n" "prfm pldl1keep, [x22, #0x80]\n" - ".inst 0x6f84e0da // udot v26.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x10, #0x40]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x10, #0x50]\n" - ".inst 0x6fa0e0c8 // udot v8.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0cc // udot v12.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d0 // udot v16.4s, v6.16b, v2.4b[1]\n" - ".inst 0x6fa3e0d4 // udot v20.4s, v6.16b, v3.4b[1]\n" - ".inst 0x6fa4e0d8 // udot v24.4s, v6.16b, v4.4b[1]\n" - "ldr q6, [x10, #0x60]\n" - ".inst 0x6fa0e0e9 // udot v9.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ed // udot v13.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f1 // udot v17.4s, v7.16b, v2.4b[1]\n" - ".inst 0x6fa3e0f5 // udot v21.4s, v7.16b, v3.4b[1]\n" - ".inst 0x6fa4e0f9 // udot v25.4s, v7.16b, v4.4b[1]\n" - "ldr q7, [x10, #0x70]\n" - ".inst 0x6fa0e0ca // udot v10.4s, v6.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ce // udot v14.4s, v6.16b, v1.4b[1]\n" - ".inst 0x6fa2e0d2 // udot v18.4s, v6.16b, v2.4b[1]\n" - ".inst 0x6fa3e0d6 // udot v22.4s, v6.16b, v3.4b[1]\n" - ".inst 0x6fa4e0da // udot v26.4s, v6.16b, v4.4b[1]\n" - "ldr q6, [x10, #0x80]\n" - ".inst 0x6fa0e0eb // udot v11.4s, v7.16b, v0.4b[1]\n" - ".inst 0x6fa1e0ef // udot v15.4s, v7.16b, v1.4b[1]\n" - ".inst 0x6fa2e0f3 // udot v19.4s, v7.16b, v2.4b[1]\n" - ".inst 0x6fa3e0f7 // udot v23.4s, v7.16b, v3.4b[1]\n" - ".inst 0x6fa4e0fb // udot v27.4s, v7.16b, v4.4b[1]\n" - "ldr q7, [x10, #0x90]\n" - ".inst 0x6f80e8c8 // udot v8.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8cc // udot v12.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d0 // udot v16.4s, v6.16b, v2.4b[2]\n" - ".inst 0x6f83e8d4 // udot v20.4s, v6.16b, v3.4b[2]\n" - ".inst 0x6f84e8d8 // udot v24.4s, v6.16b, v4.4b[2]\n" - "ldr q6, [x10, #0xa0]\n" - ".inst 0x6f80e8e9 // udot v9.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ed // udot v13.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f1 // udot v17.4s, v7.16b, v2.4b[2]\n" - ".inst 0x6f83e8f5 // udot v21.4s, v7.16b, v3.4b[2]\n" - ".inst 0x6f84e8f9 // udot v25.4s, v7.16b, v4.4b[2]\n" - "ldr q7, [x10, #0xb0]\n" - ".inst 0x6f80e8ca // udot v10.4s, v6.16b, v0.4b[2]\n" - ".inst 0x6f81e8ce // udot v14.4s, v6.16b, v1.4b[2]\n" - ".inst 0x6f82e8d2 // udot v18.4s, v6.16b, v2.4b[2]\n" - ".inst 0x6f83e8d6 // udot v22.4s, v6.16b, v3.4b[2]\n" - ".inst 0x6f84e8da // udot v26.4s, v6.16b, v4.4b[2]\n" - "ldr q6, [x10, #0xc0]\n" - ".inst 0x6f80e8eb // udot v11.4s, v7.16b, v0.4b[2]\n" - ".inst 0x6f81e8ef // udot v15.4s, v7.16b, v1.4b[2]\n" - ".inst 0x6f82e8f3 // udot v19.4s, v7.16b, v2.4b[2]\n" - ".inst 0x6f83e8f7 // udot v23.4s, v7.16b, v3.4b[2]\n" - ".inst 0x6f84e8fb // udot v27.4s, v7.16b, v4.4b[2]\n" - "ldr q7, [x10, #0xd0]\n" - ".inst 0x6fa0e8c8 // udot v8.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8cc // udot v12.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d0 // udot v16.4s, v6.16b, v2.4b[3]\n" - ".inst 0x6fa3e8d4 // udot v20.4s, v6.16b, v3.4b[3]\n" - ".inst 0x6fa4e8d8 // udot v24.4s, v6.16b, v4.4b[3]\n" - "ldr q6, [x10, #0xe0]\n" - ".inst 0x6fa0e8e9 // udot v9.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ed // udot v13.4s, v7.16b, v1.4b[3]\n" - ".inst 0x6fa2e8f1 // udot v17.4s, v7.16b, v2.4b[3]\n" - ".inst 0x6fa3e8f5 // udot v21.4s, v7.16b, v3.4b[3]\n" - ".inst 0x6fa4e8f9 // udot v25.4s, v7.16b, v4.4b[3]\n" - "ldr q7, [x10, #0xf0]\n" + ".inst 0x6f84e3ba // udot v26.4s, v29.16b, v4.4b[0]\n" + "ldr q29, [x10, #0x40]\n" + ".inst 0x6f80e38b // udot v11.4s, v28.16b, v0.4b[0]\n" + ".inst 0x6f81e38f // udot v15.4s, v28.16b, v1.4b[0]\n" + ".inst 0x6f82e393 // udot v19.4s, v28.16b, v2.4b[0]\n" + ".inst 0x6f83e397 // udot v23.4s, v28.16b, v3.4b[0]\n" + ".inst 0x6f84e39b // udot v27.4s, v28.16b, v4.4b[0]\n" + "ldr q28, [x10, #0x50]\n" + ".inst 0x6fa0e3a8 // udot v8.4s, v29.16b, v0.4b[1]\n" + ".inst 0x6fa1e3ac // udot v12.4s, v29.16b, v1.4b[1]\n" + ".inst 0x6fa2e3b0 // udot v16.4s, v29.16b, v2.4b[1]\n" + ".inst 0x6fa3e3b4 // udot v20.4s, v29.16b, v3.4b[1]\n" + ".inst 0x6fa4e3b8 // udot v24.4s, v29.16b, v4.4b[1]\n" + "ldr q29, [x10, #0x60]\n" + ".inst 0x6fa0e389 // udot v9.4s, v28.16b, v0.4b[1]\n" + ".inst 0x6fa1e38d // udot v13.4s, v28.16b, v1.4b[1]\n" + ".inst 0x6fa2e391 // udot v17.4s, v28.16b, v2.4b[1]\n" + ".inst 0x6fa3e395 // udot v21.4s, v28.16b, v3.4b[1]\n" + ".inst 0x6fa4e399 // udot v25.4s, v28.16b, v4.4b[1]\n" + "ldr q28, [x10, #0x70]\n" + ".inst 0x6fa0e3aa // udot v10.4s, v29.16b, v0.4b[1]\n" + ".inst 0x6fa1e3ae // udot v14.4s, v29.16b, v1.4b[1]\n" + ".inst 0x6fa2e3b2 // udot v18.4s, v29.16b, v2.4b[1]\n" + ".inst 0x6fa3e3b6 // udot v22.4s, v29.16b, v3.4b[1]\n" + ".inst 0x6fa4e3ba // udot v26.4s, v29.16b, v4.4b[1]\n" + "ldr q29, [x10, #0x80]\n" + ".inst 0x6fa0e38b // udot v11.4s, v28.16b, v0.4b[1]\n" + ".inst 0x6fa1e38f // udot v15.4s, v28.16b, v1.4b[1]\n" + ".inst 0x6fa2e393 // udot v19.4s, v28.16b, v2.4b[1]\n" + ".inst 0x6fa3e397 // udot v23.4s, v28.16b, v3.4b[1]\n" + ".inst 0x6fa4e39b // udot v27.4s, v28.16b, v4.4b[1]\n" + "ldr q28, [x10, #0x90]\n" + ".inst 0x6f80eba8 // udot v8.4s, v29.16b, v0.4b[2]\n" + ".inst 0x6f81ebac // udot v12.4s, v29.16b, v1.4b[2]\n" + ".inst 0x6f82ebb0 // udot v16.4s, v29.16b, v2.4b[2]\n" + ".inst 0x6f83ebb4 // udot v20.4s, v29.16b, v3.4b[2]\n" + ".inst 0x6f84ebb8 // udot v24.4s, v29.16b, v4.4b[2]\n" + "ldr q29, [x10, #0xa0]\n" + ".inst 0x6f80eb89 // udot v9.4s, v28.16b, v0.4b[2]\n" + ".inst 0x6f81eb8d // udot v13.4s, v28.16b, v1.4b[2]\n" + ".inst 0x6f82eb91 // udot v17.4s, v28.16b, v2.4b[2]\n" + ".inst 0x6f83eb95 // udot v21.4s, v28.16b, v3.4b[2]\n" + ".inst 0x6f84eb99 // udot v25.4s, v28.16b, v4.4b[2]\n" + "ldr q28, [x10, #0xb0]\n" + ".inst 0x6f80ebaa // udot v10.4s, v29.16b, v0.4b[2]\n" + ".inst 0x6f81ebae // udot v14.4s, v29.16b, v1.4b[2]\n" + ".inst 0x6f82ebb2 // udot v18.4s, v29.16b, v2.4b[2]\n" + ".inst 0x6f83ebb6 // udot v22.4s, v29.16b, v3.4b[2]\n" + ".inst 0x6f84ebba // udot v26.4s, v29.16b, v4.4b[2]\n" + "ldr q29, [x10, #0xc0]\n" + ".inst 0x6f80eb8b // udot v11.4s, v28.16b, v0.4b[2]\n" + ".inst 0x6f81eb8f // udot v15.4s, v28.16b, v1.4b[2]\n" + ".inst 0x6f82eb93 // udot v19.4s, v28.16b, v2.4b[2]\n" + ".inst 0x6f83eb97 // udot v23.4s, v28.16b, v3.4b[2]\n" + ".inst 0x6f84eb9b // udot v27.4s, v28.16b, v4.4b[2]\n" + "ldr q28, [x10, #0xd0]\n" + ".inst 0x6fa0eba8 // udot v8.4s, v29.16b, v0.4b[3]\n" + ".inst 0x6fa1ebac // udot v12.4s, v29.16b, v1.4b[3]\n" + ".inst 0x6fa2ebb0 // udot v16.4s, v29.16b, v2.4b[3]\n" + ".inst 0x6fa3ebb4 // udot v20.4s, v29.16b, v3.4b[3]\n" + ".inst 0x6fa4ebb8 // udot v24.4s, v29.16b, v4.4b[3]\n" + "ldr q29, [x10, #0xe0]\n" + ".inst 0x6fa0eb89 // udot v9.4s, v28.16b, v0.4b[3]\n" + ".inst 0x6fa1eb8d // udot v13.4s, v28.16b, v1.4b[3]\n" + ".inst 0x6fa2eb91 // udot v17.4s, v28.16b, v2.4b[3]\n" + ".inst 0x6fa3eb95 // udot v21.4s, v28.16b, v3.4b[3]\n" + ".inst 0x6fa4eb99 // udot v25.4s, v28.16b, v4.4b[3]\n" + "ldr q28, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - ".inst 0x6fa0e8ca // udot v10.4s, v6.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ce // udot v14.4s, v6.16b, v1.4b[3]\n" - ".inst 0x6fa2e8d2 // udot v18.4s, v6.16b, v2.4b[3]\n" - ".inst 0x6fa3e8d6 // udot v22.4s, v6.16b, v3.4b[3]\n" - ".inst 0x6fa4e8da // udot v26.4s, v6.16b, v4.4b[3]\n" - ".inst 0x6fa0e8eb // udot v11.4s, v7.16b, v0.4b[3]\n" - ".inst 0x6fa1e8ef // udot v15.4s, v7.16b, v1.4b[3]\n" - ".inst 0x6fa2e8f3 // udot v19.4s, v7.16b, v2.4b[3]\n" - ".inst 0x6fa3e8f7 // udot v23.4s, v7.16b, v3.4b[3]\n" - ".inst 0x6fa4e8fb // udot v27.4s, v7.16b, v4.4b[3]\n" + ".inst 0x6fa0ebaa // udot v10.4s, v29.16b, v0.4b[3]\n" + ".inst 0x6fa1ebae // udot v14.4s, v29.16b, v1.4b[3]\n" + ".inst 0x6fa2ebb2 // udot v18.4s, v29.16b, v2.4b[3]\n" + ".inst 0x6fa3ebb6 // udot v22.4s, v29.16b, v3.4b[3]\n" + ".inst 0x6fa4ebba // udot v26.4s, v29.16b, v4.4b[3]\n" + ".inst 0x6fa0eb8b // udot v11.4s, v28.16b, v0.4b[3]\n" + ".inst 0x6fa1eb8f // udot v15.4s, v28.16b, v1.4b[3]\n" + ".inst 0x6fa2eb93 // udot v19.4s, v28.16b, v2.4b[3]\n" + ".inst 0x6fa3eb97 // udot v23.4s, v28.16b, v3.4b[3]\n" + ".inst 0x6fa4eb9b // udot v27.4s, v28.16b, v4.4b[3]\n" "155:" // Height 5: Multiply loop: Main loop skip "cbz x27, 160f\n" "cmp x27, #0x4\n" "blt 157f\n" "156:" // Height 5: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" + "ldr s2, [x26], #0x4\n" "ldr s1, [x25], #0x4\n" "sub x27, x27, #0x4\n" "cmp x27, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr s4, [x22], #0x4\n" - "ldr q6, [x10, #0x0]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" - ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr s0, [x24], #0x4\n" + "ldr s31, [x23], #0x4\n" + "ldr s30, [x22], #0x4\n" + "ldr q29, [x10, #0x0]\n" + ".inst 0x6f82e3a8 // udot v8.4s, v29.16b, v2.4b[0]\n" + ".inst 0x6f81e3ac // udot v12.4s, v29.16b, v1.4b[0]\n" + "ldr q28, [x10, #0x10]\n" + ".inst 0x6f80e3b0 // udot v16.4s, v29.16b, v0.4b[0]\n" + ".inst 0x6f9fe3b4 // udot v20.4s, v29.16b, v31.4b[0]\n" + ".inst 0x6f9ee3b8 // udot v24.4s, v29.16b, v30.4b[0]\n" + "ldr q29, [x10, #0x20]\n" + ".inst 0x6f82e389 // udot v9.4s, v28.16b, v2.4b[0]\n" + ".inst 0x6f81e38d // udot v13.4s, v28.16b, v1.4b[0]\n" + ".inst 0x6f80e391 // udot v17.4s, v28.16b, v0.4b[0]\n" + ".inst 0x6f9fe395 // udot v21.4s, v28.16b, v31.4b[0]\n" + ".inst 0x6f9ee399 // udot v25.4s, v28.16b, v30.4b[0]\n" + "ldr q28, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x6f84e0da // udot v26.4s, v6.16b, v4.4b[0]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" + ".inst 0x6f82e3aa // udot v10.4s, v29.16b, v2.4b[0]\n" + ".inst 0x6f81e3ae // udot v14.4s, v29.16b, v1.4b[0]\n" + ".inst 0x6f80e3b2 // udot v18.4s, v29.16b, v0.4b[0]\n" + ".inst 0x6f9fe3b6 // udot v22.4s, v29.16b, v31.4b[0]\n" + ".inst 0x6f9ee3ba // udot v26.4s, v29.16b, v30.4b[0]\n" + ".inst 0x6f82e38b // udot v11.4s, v28.16b, v2.4b[0]\n" + ".inst 0x6f81e38f // udot v15.4s, v28.16b, v1.4b[0]\n" + ".inst 0x6f80e393 // udot v19.4s, v28.16b, v0.4b[0]\n" + ".inst 0x6f9fe397 // udot v23.4s, v28.16b, v31.4b[0]\n" + ".inst 0x6f9ee39b // udot v27.4s, v28.16b, v30.4b[0]\n" "bge 156b\n" "157:" // Height 5: Multiply loop: Skip odd blocks "cbz x27, 160f\n" @@ -2286,31 +2285,31 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr b3, [x23, #0x0]\n" "ldr b4, [x22, #0x0]\n" "159:" // Height 5: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" - ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q29, [x10, #0x0]\n" + "ldr q28, [x10, #0x10]\n" + ".inst 0x6f80e3a8 // udot v8.4s, v29.16b, v0.4b[0]\n" + ".inst 0x6f81e3ac // udot v12.4s, v29.16b, v1.4b[0]\n" + ".inst 0x6f82e3b0 // udot v16.4s, v29.16b, v2.4b[0]\n" + ".inst 0x6f83e3b4 // udot v20.4s, v29.16b, v3.4b[0]\n" + ".inst 0x6f84e3b8 // udot v24.4s, v29.16b, v4.4b[0]\n" + "ldr q29, [x10, #0x20]\n" + ".inst 0x6f80e389 // udot v9.4s, v28.16b, v0.4b[0]\n" + ".inst 0x6f81e38d // udot v13.4s, v28.16b, v1.4b[0]\n" + ".inst 0x6f82e391 // udot v17.4s, v28.16b, v2.4b[0]\n" + ".inst 0x6f83e395 // udot v21.4s, v28.16b, v3.4b[0]\n" + ".inst 0x6f84e399 // udot v25.4s, v28.16b, v4.4b[0]\n" + "ldr q28, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x6f84e0da // udot v26.4s, v6.16b, v4.4b[0]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" + ".inst 0x6f80e3aa // udot v10.4s, v29.16b, v0.4b[0]\n" + ".inst 0x6f81e3ae // udot v14.4s, v29.16b, v1.4b[0]\n" + ".inst 0x6f82e3b2 // udot v18.4s, v29.16b, v2.4b[0]\n" + ".inst 0x6f83e3b6 // udot v22.4s, v29.16b, v3.4b[0]\n" + ".inst 0x6f84e3ba // udot v26.4s, v29.16b, v4.4b[0]\n" + ".inst 0x6f80e38b // udot v11.4s, v28.16b, v0.4b[0]\n" + ".inst 0x6f81e38f // udot v15.4s, v28.16b, v1.4b[0]\n" + ".inst 0x6f82e393 // udot v19.4s, v28.16b, v2.4b[0]\n" + ".inst 0x6f83e397 // udot v23.4s, v28.16b, v3.4b[0]\n" + ".inst 0x6f84e39b // udot 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 x28, x28, #0x1\n" @@ -2672,16 +2671,16 @@ void a64_hybrid_u8u32_dot_6x16 ( "184:" // Height 6: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 185f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" - "ldr x21, [x21, #0x28]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" + "ldr x21, [x20, #0x28]\n" "cbnz x28, 186f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20\n" @@ -2693,11 +2692,11 @@ void a64_hybrid_u8u32_dot_6x16 ( "b 186f\n" "185:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20\n" - "add x24, x25, x20\n" - "add x23, x24, x20\n" - "add x22, x23, x20\n" - "add x21, x22, x20\n" + "add x25, x26, x21\n" + "add x24, x25, x21\n" + "add x23, x24, x21\n" + "add x22, x23, x21\n" + "add x21, x22, x21\n" "186:" // Height 6: input setup done "cmp x27, #0x10\n" "blt 189f\n" @@ -2976,43 +2975,43 @@ void a64_hybrid_u8u32_dot_6x16 ( "cmp x27, #0x4\n" "blt 191f\n" "190:" // Height 6: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr s1, [x25], #0x4\n" + "ldr s7, [x26], #0x4\n" + "ldr s6, [x25], #0x4\n" "sub x27, x27, #0x4\n" "cmp x27, #0x4\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr s4, [x22], #0x4\n" - "ldr s5, [x21], #0x4\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" - ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" - ".inst 0x6f85e0dc // udot v28.4s, v6.16b, v5.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" - ".inst 0x6f85e0fd // udot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr s5, [x24], #0x4\n" + "ldr s4, [x23], #0x4\n" + "ldr s3, [x22], #0x4\n" + "ldr s2, [x21], #0x4\n" + "ldr q1, [x10, #0x0]\n" + "ldr q0, [x10, #0x10]\n" + ".inst 0x6f87e028 // udot v8.4s, v1.16b, v7.4b[0]\n" + ".inst 0x6f86e02c // udot v12.4s, v1.16b, v6.4b[0]\n" + ".inst 0x6f85e030 // udot v16.4s, v1.16b, v5.4b[0]\n" + ".inst 0x6f84e034 // udot v20.4s, v1.16b, v4.4b[0]\n" + ".inst 0x6f83e038 // udot v24.4s, v1.16b, v3.4b[0]\n" + ".inst 0x6f82e03c // udot v28.4s, v1.16b, v2.4b[0]\n" + "ldr q1, [x10, #0x20]\n" + ".inst 0x6f87e009 // udot v9.4s, v0.16b, v7.4b[0]\n" + ".inst 0x6f86e00d // udot v13.4s, v0.16b, v6.4b[0]\n" + ".inst 0x6f85e011 // udot v17.4s, v0.16b, v5.4b[0]\n" + ".inst 0x6f84e015 // udot v21.4s, v0.16b, v4.4b[0]\n" + ".inst 0x6f83e019 // udot v25.4s, v0.16b, v3.4b[0]\n" + ".inst 0x6f82e01d // udot v29.4s, v0.16b, v2.4b[0]\n" + "ldr q0, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x6f84e0da // udot v26.4s, v6.16b, v4.4b[0]\n" - ".inst 0x6f85e0de // udot v30.4s, v6.16b, v5.4b[0]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" - ".inst 0x6f85e0ff // udot v31.4s, v7.16b, v5.4b[0]\n" + ".inst 0x6f87e02a // udot v10.4s, v1.16b, v7.4b[0]\n" + ".inst 0x6f86e02e // udot v14.4s, v1.16b, v6.4b[0]\n" + ".inst 0x6f85e032 // udot v18.4s, v1.16b, v5.4b[0]\n" + ".inst 0x6f84e036 // udot v22.4s, v1.16b, v4.4b[0]\n" + ".inst 0x6f83e03a // udot v26.4s, v1.16b, v3.4b[0]\n" + ".inst 0x6f82e03e // udot v30.4s, v1.16b, v2.4b[0]\n" + ".inst 0x6f87e00b // udot v11.4s, v0.16b, v7.4b[0]\n" + ".inst 0x6f86e00f // udot v15.4s, v0.16b, v6.4b[0]\n" + ".inst 0x6f85e013 // udot v19.4s, v0.16b, v5.4b[0]\n" + ".inst 0x6f84e017 // udot v23.4s, v0.16b, v4.4b[0]\n" + ".inst 0x6f83e01b // udot v27.4s, v0.16b, v3.4b[0]\n" + ".inst 0x6f82e01f // udot v31.4s, v0.16b, v2.4b[0]\n" "bge 190b\n" "191:" // Height 6: Multiply loop: Skip odd blocks "cbz x27, 194f\n" @@ -3039,35 +3038,35 @@ void a64_hybrid_u8u32_dot_6x16 ( "ldr b4, [x22, #0x0]\n" "ldr b5, [x21, #0x0]\n" "193:" // Height 6: Multiply loop: Ragged operand read: Done - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - ".inst 0x6f80e0c8 // udot v8.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0cc // udot v12.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d0 // udot v16.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d4 // udot v20.4s, v6.16b, v3.4b[0]\n" - ".inst 0x6f84e0d8 // udot v24.4s, v6.16b, v4.4b[0]\n" - ".inst 0x6f85e0dc // udot v28.4s, v6.16b, v5.4b[0]\n" - "ldr q6, [x10, #0x20]\n" - ".inst 0x6f80e0e9 // udot v9.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ed // udot v13.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f1 // udot v17.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f5 // udot v21.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f84e0f9 // udot v25.4s, v7.16b, v4.4b[0]\n" - ".inst 0x6f85e0fd // udot v29.4s, v7.16b, v5.4b[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q7, [x10, #0x0]\n" + "ldr q6, [x10, #0x10]\n" + ".inst 0x6f80e0e8 // udot v8.4s, v7.16b, v0.4b[0]\n" + ".inst 0x6f81e0ec // udot v12.4s, v7.16b, v1.4b[0]\n" + ".inst 0x6f82e0f0 // udot v16.4s, v7.16b, v2.4b[0]\n" + ".inst 0x6f83e0f4 // udot v20.4s, v7.16b, v3.4b[0]\n" + ".inst 0x6f84e0f8 // udot v24.4s, v7.16b, v4.4b[0]\n" + ".inst 0x6f85e0fc // udot v28.4s, v7.16b, v5.4b[0]\n" + "ldr q7, [x10, #0x20]\n" + ".inst 0x6f80e0c9 // udot v9.4s, v6.16b, v0.4b[0]\n" + ".inst 0x6f81e0cd // udot v13.4s, v6.16b, v1.4b[0]\n" + ".inst 0x6f82e0d1 // udot v17.4s, v6.16b, v2.4b[0]\n" + ".inst 0x6f83e0d5 // udot v21.4s, v6.16b, v3.4b[0]\n" + ".inst 0x6f84e0d9 // udot v25.4s, v6.16b, v4.4b[0]\n" + ".inst 0x6f85e0dd // udot v29.4s, v6.16b, v5.4b[0]\n" + "ldr q6, [x10, #0x30]\n" "add x10, x10, #0x40\n" - ".inst 0x6f80e0ca // udot v10.4s, v6.16b, v0.4b[0]\n" - ".inst 0x6f81e0ce // udot v14.4s, v6.16b, v1.4b[0]\n" - ".inst 0x6f82e0d2 // udot v18.4s, v6.16b, v2.4b[0]\n" - ".inst 0x6f83e0d6 // udot v22.4s, v6.16b, v3.4b[0]\n" - ".inst 0x6f84e0da // udot v26.4s, v6.16b, v4.4b[0]\n" - ".inst 0x6f85e0de // udot v30.4s, v6.16b, v5.4b[0]\n" - ".inst 0x6f80e0eb // udot v11.4s, v7.16b, v0.4b[0]\n" - ".inst 0x6f81e0ef // udot v15.4s, v7.16b, v1.4b[0]\n" - ".inst 0x6f82e0f3 // udot v19.4s, v7.16b, v2.4b[0]\n" - ".inst 0x6f83e0f7 // udot v23.4s, v7.16b, v3.4b[0]\n" - ".inst 0x6f84e0fb // udot v27.4s, v7.16b, v4.4b[0]\n" - ".inst 0x6f85e0ff // udot v31.4s, v7.16b, v5.4b[0]\n" + ".inst 0x6f80e0ea // udot v10.4s, v7.16b, v0.4b[0]\n" + ".inst 0x6f81e0ee // udot v14.4s, v7.16b, v1.4b[0]\n" + ".inst 0x6f82e0f2 // udot v18.4s, v7.16b, v2.4b[0]\n" + ".inst 0x6f83e0f6 // udot v22.4s, v7.16b, v3.4b[0]\n" + ".inst 0x6f84e0fa // udot v26.4s, v7.16b, v4.4b[0]\n" + ".inst 0x6f85e0fe // udot v30.4s, v7.16b, v5.4b[0]\n" + ".inst 0x6f80e0cb // udot v11.4s, v6.16b, v0.4b[0]\n" + ".inst 0x6f81e0cf // udot v15.4s, v6.16b, v1.4b[0]\n" + ".inst 0x6f82e0d3 // udot v19.4s, v6.16b, v2.4b[0]\n" + ".inst 0x6f83e0d7 // udot v23.4s, v6.16b, v3.4b[0]\n" + ".inst 0x6f84e0db // udot v27.4s, v6.16b, v4.4b[0]\n" + ".inst 0x6f85e0df // udot 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 x28, x28, #0x1\n" @@ -3254,7 +3253,6 @@ void a64_hybrid_u8u32_dot_6x16 ( "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", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" -- cgit v1.2.1