diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp | 1746 |
1 files changed, 872 insertions, 874 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp index c5e4388aa9..bb84a50282 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp32_mla_6x16/generic.cpp @@ -92,7 +92,6 @@ void a64_hybrid_fp32_mla_6x16 ( break; } __asm__ __volatile__( - "1:" // Row loop "cmp %x[M], #0x6\n" "bge 166f\n" @@ -189,11 +188,11 @@ void a64_hybrid_fp32_mla_6x16 ( "15:" // Height 1: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 16f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" "cbnz x28, 17f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -210,37 +209,37 @@ void a64_hybrid_fp32_mla_6x16 ( "blt 19f\n" "18:" // Height 1: Multiply loop: Main loop head "fmla v8.4s, v6.4s, v0.s[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" - "ldr q7, [x10, #0x30]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "ldr q7, [x10, #0xf0]\n" + "ldr q16, [x10, #0x30]\n" + "fmla v10.4s, v17.4s, v0.s[0]\n" + "ldr q17, [x10, #0x40]\n" + "fmla v11.4s, v16.4s, v0.s[0]\n" + "ldr q16, [x10, #0x50]\n" + "fmla v8.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x10, #0x60]\n" + "fmla v9.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x10, #0x70]\n" + "fmla v10.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x10, #0x80]\n" + "fmla v11.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x10, #0x90]\n" + "fmla v8.4s, v17.4s, v0.s[2]\n" + "ldr q17, [x10, #0xa0]\n" + "fmla v9.4s, v16.4s, v0.s[2]\n" + "ldr q16, [x10, #0xb0]\n" + "fmla v10.4s, v17.4s, v0.s[2]\n" + "ldr q17, [x10, #0xc0]\n" + "fmla v11.4s, v16.4s, v0.s[2]\n" + "ldr q16, [x10, #0xd0]\n" + "fmla v8.4s, v17.4s, v0.s[3]\n" + "ldr q17, [x10, #0xe0]\n" + "fmla v9.4s, v16.4s, v0.s[3]\n" + "ldr q16, [x10, #0xf0]\n" "sub x27, x27, #0x4\n" "add x26, x26, #0x10\n" - "fmla v10.4s, v6.4s, v0.s[3]\n" - "fmla v11.4s, v7.4s, v0.s[3]\n" + "fmla v10.4s, v17.4s, v0.s[3]\n" + "fmla v11.4s, v16.4s, v0.s[3]\n" "ldr q0, [x26, #0x0]\n" "cmp x27, #0x8\n" "add x10, x10, #0x100\n" @@ -250,52 +249,52 @@ void a64_hybrid_fp32_mla_6x16 ( "bge 18b\n" "19:" // Height 1: Multiply loop: Single iteration only "fmla v8.4s, v6.4s, v0.s[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" - "ldr q7, [x10, #0x30]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "ldr q7, [x10, #0xf0]\n" + "ldr q16, [x10, #0x30]\n" + "fmla v10.4s, v17.4s, v0.s[0]\n" + "ldr q17, [x10, #0x40]\n" + "fmla v11.4s, v16.4s, v0.s[0]\n" + "ldr q16, [x10, #0x50]\n" + "fmla v8.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x10, #0x60]\n" + "fmla v9.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x10, #0x70]\n" + "fmla v10.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x10, #0x80]\n" + "fmla v11.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x10, #0x90]\n" + "fmla v8.4s, v17.4s, v0.s[2]\n" + "ldr q17, [x10, #0xa0]\n" + "fmla v9.4s, v16.4s, v0.s[2]\n" + "ldr q16, [x10, #0xb0]\n" + "fmla v10.4s, v17.4s, v0.s[2]\n" + "ldr q17, [x10, #0xc0]\n" + "fmla v11.4s, v16.4s, v0.s[2]\n" + "ldr q16, [x10, #0xd0]\n" + "fmla v8.4s, v17.4s, v0.s[3]\n" + "ldr q17, [x10, #0xe0]\n" + "fmla v9.4s, v16.4s, v0.s[3]\n" + "ldr q16, [x10, #0xf0]\n" "add x26, x26, #0x10\n" "sub x27, x27, #0x4\n" - "fmla v10.4s, v6.4s, v0.s[3]\n" - "fmla v11.4s, v7.4s, v0.s[3]\n" + "fmla v10.4s, v17.4s, v0.s[3]\n" + "fmla v11.4s, v16.4s, v0.s[3]\n" "prfm pldl1keep, [x26, #0x80]\n" "add x10, x10, #0x100\n" "20:" // Height 1: Multiply loop: Main loop skip "cbz x27, 22f\n" "21:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr q6, [x10, #0x0]\n" - "fmla v8.4s, v6.4s, v0.s[0]\n" + "ldr s18, [x26], #0x4\n" + "ldr q16, [x10, #0x0]\n" + "fmla v8.4s, v16.4s, v18.s[0]\n" "sub x27, x27, #0x1\n" - "ldr q7, [x10, #0x10]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.4s, v7.4s, v0.s[0]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "ldr q7, [x10, #0x30]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" + "ldr q17, [x10, #0x10]\n" + "ldr q16, [x10, #0x20]\n" + "fmla v9.4s, v17.4s, v18.s[0]\n" + "fmla v10.4s, v16.4s, v18.s[0]\n" + "ldr q16, [x10, #0x30]\n" + "fmla v11.4s, v16.4s, v18.s[0]\n" "add x10, x10, #0x40\n" "cbnz x27, 21b\n" "22:" // Height 1: Multiply loop: No odd multiplies @@ -306,17 +305,17 @@ void a64_hybrid_fp32_mla_6x16 ( "prfm pstl1keep, [x9, #0x0]\n" "tbz %x[flags], #1, 23f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" + "ld1r { v16.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v17.4s\n" + "fmin v9.4s, v9.4s, v17.4s\n" + "fmin v10.4s, v10.4s, v17.4s\n" + "fmin v11.4s, v11.4s, v17.4s\n" + "fmax v8.4s, v8.4s, v16.4s\n" + "fmax v9.4s, v9.4s, v16.4s\n" + "fmax v10.4s, v10.4s, v16.4s\n" + "fmax v11.4s, v11.4s, v16.4s\n" "23:" // Height 1: No activation "cmp x11, #0x10\n" "bge 32f\n" @@ -494,12 +493,12 @@ void a64_hybrid_fp32_mla_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, LSL #2\n" @@ -507,7 +506,7 @@ void a64_hybrid_fp32_mla_6x16 ( "b 50f\n" "49:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" "50:" // Height 2: input setup done "cmp x27, #0x4\n" "blt 53f\n" @@ -520,134 +519,134 @@ void a64_hybrid_fp32_mla_6x16 ( "51:" // Height 2: Multiply loop: Main loop head "fmla v8.4s, v6.4s, v0.s[0]\n" "fmla v12.4s, v6.4s, v1.s[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "sub x27, x27, #0x4\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "fmla v13.4s, v7.4s, v1.s[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q16, [x10, #0x30]\n" "add x26, x26, #0x10\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" - "ldr q6, [x10, #0x40]\n" + "fmla v10.4s, v17.4s, v0.s[0]\n" + "fmla v14.4s, v17.4s, v1.s[0]\n" + "ldr q17, [x10, #0x40]\n" "add x25, x25, #0x10\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "ldr q7, [x10, #0x50]\n" + "fmla v11.4s, v16.4s, v0.s[0]\n" + "fmla v15.4s, v16.4s, v1.s[0]\n" + "ldr q16, [x10, #0x50]\n" "cmp x27, #0x8\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v12.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x10, #0x60]\n" + "fmla v8.4s, v17.4s, v0.s[1]\n" + "fmla v12.4s, v17.4s, v1.s[1]\n" + "ldr q17, [x10, #0x60]\n" "prfm pldl1keep, [x26, #0x80]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v13.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x10, #0x70]\n" + "fmla v9.4s, v16.4s, v0.s[1]\n" + "fmla v13.4s, v16.4s, v1.s[1]\n" + "ldr q16, [x10, #0x70]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "fmla v12.4s, v6.4s, v1.s[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "fmla v13.4s, v7.4s, v1.s[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v14.4s, v6.4s, v1.s[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v12.4s, v6.4s, v1.s[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "ldr q7, [x10, #0xf0]\n" + "fmla v10.4s, v17.4s, v0.s[1]\n" + "fmla v14.4s, v17.4s, v1.s[1]\n" + "ldr q17, [x10, #0x80]\n" + "fmla v11.4s, v16.4s, v0.s[1]\n" + "fmla v15.4s, v16.4s, v1.s[1]\n" + "ldr q16, [x10, #0x90]\n" + "fmla v8.4s, v17.4s, v0.s[2]\n" + "fmla v12.4s, v17.4s, v1.s[2]\n" + "ldr q17, [x10, #0xa0]\n" + "fmla v9.4s, v16.4s, v0.s[2]\n" + "fmla v13.4s, v16.4s, v1.s[2]\n" + "ldr q16, [x10, #0xb0]\n" + "fmla v10.4s, v17.4s, v0.s[2]\n" + "fmla v14.4s, v17.4s, v1.s[2]\n" + "ldr q17, [x10, #0xc0]\n" + "fmla v11.4s, v16.4s, v0.s[2]\n" + "fmla v15.4s, v16.4s, v1.s[2]\n" + "ldr q16, [x10, #0xd0]\n" + "fmla v8.4s, v17.4s, v0.s[3]\n" + "fmla v12.4s, v17.4s, v1.s[3]\n" + "ldr q17, [x10, #0xe0]\n" + "fmla v9.4s, v16.4s, v0.s[3]\n" + "fmla v13.4s, v16.4s, v1.s[3]\n" + "ldr q16, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - "fmla v10.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" + "fmla v10.4s, v17.4s, v0.s[3]\n" + "fmla v14.4s, v17.4s, v1.s[3]\n" "ldr q6, [x10, #0x0]\n" - "fmla v11.4s, v7.4s, v0.s[3]\n" + "fmla v11.4s, v16.4s, v0.s[3]\n" "ldr q0, [x26, #0x0]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" + "fmla v15.4s, v16.4s, v1.s[3]\n" "ldr q1, [x25, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 51b\n" "52:" // Height 2: Multiply loop: Single iteration only "fmla v8.4s, v6.4s, v0.s[0]\n" "fmla v12.4s, v6.4s, v1.s[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "add x26, x26, #0x10\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "fmla v13.4s, v7.4s, v1.s[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q16, [x10, #0x30]\n" "add x25, x25, #0x10\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" - "ldr q6, [x10, #0x40]\n" + "fmla v10.4s, v17.4s, v0.s[0]\n" + "fmla v14.4s, v17.4s, v1.s[0]\n" + "ldr q17, [x10, #0x40]\n" "sub x27, x27, #0x4\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "ldr q7, [x10, #0x50]\n" + "fmla v11.4s, v16.4s, v0.s[0]\n" + "fmla v15.4s, v16.4s, v1.s[0]\n" + "ldr q16, [x10, #0x50]\n" "prfm pldl1keep, [x26, #0x80]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v12.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x10, #0x60]\n" + "fmla v8.4s, v17.4s, v0.s[1]\n" + "fmla v12.4s, v17.4s, v1.s[1]\n" + "ldr q17, [x10, #0x60]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v13.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "fmla v12.4s, v6.4s, v1.s[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "fmla v13.4s, v7.4s, v1.s[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v14.4s, v6.4s, v1.s[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v12.4s, v6.4s, v1.s[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "ldr q7, [x10, #0xf0]\n" + "fmla v9.4s, v16.4s, v0.s[1]\n" + "fmla v13.4s, v16.4s, v1.s[1]\n" + "ldr q16, [x10, #0x70]\n" + "fmla v10.4s, v17.4s, v0.s[1]\n" + "fmla v14.4s, v17.4s, v1.s[1]\n" + "ldr q17, [x10, #0x80]\n" + "fmla v11.4s, v16.4s, v0.s[1]\n" + "fmla v15.4s, v16.4s, v1.s[1]\n" + "ldr q16, [x10, #0x90]\n" + "fmla v8.4s, v17.4s, v0.s[2]\n" + "fmla v12.4s, v17.4s, v1.s[2]\n" + "ldr q17, [x10, #0xa0]\n" + "fmla v9.4s, v16.4s, v0.s[2]\n" + "fmla v13.4s, v16.4s, v1.s[2]\n" + "ldr q16, [x10, #0xb0]\n" + "fmla v10.4s, v17.4s, v0.s[2]\n" + "fmla v14.4s, v17.4s, v1.s[2]\n" + "ldr q17, [x10, #0xc0]\n" + "fmla v11.4s, v16.4s, v0.s[2]\n" + "fmla v15.4s, v16.4s, v1.s[2]\n" + "ldr q16, [x10, #0xd0]\n" + "fmla v8.4s, v17.4s, v0.s[3]\n" + "fmla v12.4s, v17.4s, v1.s[3]\n" + "ldr q17, [x10, #0xe0]\n" + "fmla v9.4s, v16.4s, v0.s[3]\n" + "fmla v13.4s, v16.4s, v1.s[3]\n" + "ldr q16, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - "fmla v10.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" - "fmla v11.4s, v7.4s, v0.s[3]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" + "fmla v10.4s, v17.4s, v0.s[3]\n" + "fmla v14.4s, v17.4s, v1.s[3]\n" + "fmla v11.4s, v16.4s, v0.s[3]\n" + "fmla v15.4s, v16.4s, v1.s[3]\n" "53:" // Height 2: Multiply loop: Main loop skip "cbz x27, 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, #0x1\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - "fmla v8.4s, v6.4s, v0.s[0]\n" - "fmla v12.4s, v6.4s, v1.s[0]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.4s, v7.4s, v0.s[0]\n" - "fmla v13.4s, v7.4s, v1.s[0]\n" - "ldr q7, [x10, #0x30]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" + "ldr q17, [x10, #0x0]\n" + "ldr q16, [x10, #0x10]\n" + "fmla v8.4s, v17.4s, v19.s[0]\n" + "fmla v12.4s, v17.4s, v18.s[0]\n" + "ldr q17, [x10, #0x20]\n" + "fmla v9.4s, v16.4s, v19.s[0]\n" + "fmla v13.4s, v16.4s, v18.s[0]\n" + "ldr q16, [x10, #0x30]\n" + "fmla v10.4s, v17.4s, v19.s[0]\n" + "fmla v14.4s, v17.4s, v18.s[0]\n" "add x10, x10, #0x40\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" + "fmla v11.4s, v16.4s, v19.s[0]\n" + "fmla v15.4s, v16.4s, v18.s[0]\n" "cbnz x27, 54b\n" "55:" // Height 2: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -660,25 +659,25 @@ void a64_hybrid_fp32_mla_6x16 ( "prfm pstl1keep, [x25, #0x0]\n" "tbz %x[flags], #1, 56f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v17.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" + "ld1r { v16.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v17.4s\n" + "fmin v9.4s, v9.4s, v17.4s\n" + "fmin v10.4s, v10.4s, v17.4s\n" + "fmin v11.4s, v11.4s, v17.4s\n" + "fmin v12.4s, v12.4s, v17.4s\n" + "fmin v13.4s, v13.4s, v17.4s\n" + "fmin v14.4s, v14.4s, v17.4s\n" + "fmin v15.4s, v15.4s, v17.4s\n" + "fmax v8.4s, v8.4s, v16.4s\n" + "fmax v9.4s, v9.4s, v16.4s\n" + "fmax v10.4s, v10.4s, v16.4s\n" + "fmax v11.4s, v11.4s, v16.4s\n" + "fmax v12.4s, v12.4s, v16.4s\n" + "fmax v13.4s, v13.4s, v16.4s\n" + "fmax v14.4s, v14.4s, v16.4s\n" + "fmax v15.4s, v15.4s, v16.4s\n" "56:" // Height 2: No activation "cmp x11, #0x10\n" "bge 65f\n" @@ -905,13 +904,13 @@ void a64_hybrid_fp32_mla_6x16 ( "81:" // 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, 82f\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, 83f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -920,8 +919,8 @@ void a64_hybrid_fp32_mla_6x16 ( "b 83f\n" "82:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" "83:" // Height 3: input setup done "cmp x27, #0x4\n" "blt 86f\n" @@ -938,75 +937,75 @@ void a64_hybrid_fp32_mla_6x16 ( "sub x27, x27, #0x4\n" "add x26, x26, #0x10\n" "fmla v16.4s, v6.4s, v2.s[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q21, [x10, #0x20]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "add x25, x25, #0x10\n" "fmla v13.4s, v7.4s, v1.s[0]\n" "fmla v17.4s, v7.4s, v2.s[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q20, [x10, #0x30]\n" "add x24, x24, #0x10\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" + "fmla v10.4s, v21.4s, v0.s[0]\n" + "fmla v14.4s, v21.4s, v1.s[0]\n" "cmp x27, #0x8\n" "prfm pldl1keep, [x26, #0x80]\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" + "fmla v18.4s, v21.4s, v2.s[0]\n" + "ldr q21, [x10, #0x40]\n" + "fmla v11.4s, v20.4s, v0.s[0]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "fmla v19.4s, v7.4s, v2.s[0]\n" - "ldr q7, [x10, #0x50]\n" + "fmla v15.4s, v20.4s, v1.s[0]\n" + "fmla v19.4s, v20.4s, v2.s[0]\n" + "ldr q20, [x10, #0x50]\n" "prfm pldl1keep, [x24, #0x80]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v12.4s, v6.4s, v1.s[1]\n" - "fmla v16.4s, v6.4s, v2.s[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v13.4s, v7.4s, v1.s[1]\n" - "fmla v17.4s, v7.4s, v2.s[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "fmla v18.4s, v6.4s, v2.s[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "fmla v19.4s, v7.4s, v2.s[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "fmla v12.4s, v6.4s, v1.s[2]\n" - "fmla v16.4s, v6.4s, v2.s[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "fmla v13.4s, v7.4s, v1.s[2]\n" - "fmla v17.4s, v7.4s, v2.s[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v14.4s, v6.4s, v1.s[2]\n" - "fmla v18.4s, v6.4s, v2.s[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "fmla v19.4s, v7.4s, v2.s[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v12.4s, v6.4s, v1.s[3]\n" - "fmla v16.4s, v6.4s, v2.s[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "fmla v17.4s, v7.4s, v2.s[3]\n" - "ldr q7, [x10, #0xf0]\n" + "fmla v8.4s, v21.4s, v0.s[1]\n" + "fmla v12.4s, v21.4s, v1.s[1]\n" + "fmla v16.4s, v21.4s, v2.s[1]\n" + "ldr q21, [x10, #0x60]\n" + "fmla v9.4s, v20.4s, v0.s[1]\n" + "fmla v13.4s, v20.4s, v1.s[1]\n" + "fmla v17.4s, v20.4s, v2.s[1]\n" + "ldr q20, [x10, #0x70]\n" + "fmla v10.4s, v21.4s, v0.s[1]\n" + "fmla v14.4s, v21.4s, v1.s[1]\n" + "fmla v18.4s, v21.4s, v2.s[1]\n" + "ldr q21, [x10, #0x80]\n" + "fmla v11.4s, v20.4s, v0.s[1]\n" + "fmla v15.4s, v20.4s, v1.s[1]\n" + "fmla v19.4s, v20.4s, v2.s[1]\n" + "ldr q20, [x10, #0x90]\n" + "fmla v8.4s, v21.4s, v0.s[2]\n" + "fmla v12.4s, v21.4s, v1.s[2]\n" + "fmla v16.4s, v21.4s, v2.s[2]\n" + "ldr q21, [x10, #0xa0]\n" + "fmla v9.4s, v20.4s, v0.s[2]\n" + "fmla v13.4s, v20.4s, v1.s[2]\n" + "fmla v17.4s, v20.4s, v2.s[2]\n" + "ldr q20, [x10, #0xb0]\n" + "fmla v10.4s, v21.4s, v0.s[2]\n" + "fmla v14.4s, v21.4s, v1.s[2]\n" + "fmla v18.4s, v21.4s, v2.s[2]\n" + "ldr q21, [x10, #0xc0]\n" + "fmla v11.4s, v20.4s, v0.s[2]\n" + "fmla v15.4s, v20.4s, v1.s[2]\n" + "fmla v19.4s, v20.4s, v2.s[2]\n" + "ldr q20, [x10, #0xd0]\n" + "fmla v8.4s, v21.4s, v0.s[3]\n" + "fmla v12.4s, v21.4s, v1.s[3]\n" + "fmla v16.4s, v21.4s, v2.s[3]\n" + "ldr q21, [x10, #0xe0]\n" + "fmla v9.4s, v20.4s, v0.s[3]\n" + "fmla v13.4s, v20.4s, v1.s[3]\n" + "fmla v17.4s, v20.4s, v2.s[3]\n" + "ldr q20, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - "fmla v10.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" - "fmla v18.4s, v6.4s, v2.s[3]\n" + "fmla v10.4s, v21.4s, v0.s[3]\n" + "fmla v14.4s, v21.4s, v1.s[3]\n" + "fmla v18.4s, v21.4s, v2.s[3]\n" "ldr q6, [x10, #0x0]\n" - "fmla v11.4s, v7.4s, v0.s[3]\n" + "fmla v11.4s, v20.4s, v0.s[3]\n" "ldr q0, [x26, #0x0]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" + "fmla v15.4s, v20.4s, v1.s[3]\n" "ldr q1, [x25, #0x0]\n" - "fmla v19.4s, v7.4s, v2.s[3]\n" + "fmla v19.4s, v20.4s, v2.s[3]\n" "ldr q2, [x24, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 84b\n" @@ -1016,95 +1015,95 @@ void a64_hybrid_fp32_mla_6x16 ( "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" "fmla v16.4s, v6.4s, v2.s[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q21, [x10, #0x20]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "add x24, x24, #0x10\n" "fmla v13.4s, v7.4s, v1.s[0]\n" "fmla v17.4s, v7.4s, v2.s[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q20, [x10, #0x30]\n" "sub x27, x27, #0x4\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" + "fmla v10.4s, v21.4s, v0.s[0]\n" + "fmla v14.4s, v21.4s, v1.s[0]\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" + "fmla v18.4s, v21.4s, v2.s[0]\n" + "ldr q21, [x10, #0x40]\n" + "fmla v11.4s, v20.4s, v0.s[0]\n" "prfm pldl1keep, [x24, #0x80]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "fmla v19.4s, v7.4s, v2.s[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v12.4s, v6.4s, v1.s[1]\n" - "fmla v16.4s, v6.4s, v2.s[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v13.4s, v7.4s, v1.s[1]\n" - "fmla v17.4s, v7.4s, v2.s[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "fmla v18.4s, v6.4s, v2.s[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "fmla v19.4s, v7.4s, v2.s[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "fmla v12.4s, v6.4s, v1.s[2]\n" - "fmla v16.4s, v6.4s, v2.s[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "fmla v13.4s, v7.4s, v1.s[2]\n" - "fmla v17.4s, v7.4s, v2.s[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v14.4s, v6.4s, v1.s[2]\n" - "fmla v18.4s, v6.4s, v2.s[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "fmla v19.4s, v7.4s, v2.s[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v12.4s, v6.4s, v1.s[3]\n" - "fmla v16.4s, v6.4s, v2.s[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "fmla v17.4s, v7.4s, v2.s[3]\n" - "ldr q7, [x10, #0xf0]\n" + "fmla v15.4s, v20.4s, v1.s[0]\n" + "fmla v19.4s, v20.4s, v2.s[0]\n" + "ldr q20, [x10, #0x50]\n" + "fmla v8.4s, v21.4s, v0.s[1]\n" + "fmla v12.4s, v21.4s, v1.s[1]\n" + "fmla v16.4s, v21.4s, v2.s[1]\n" + "ldr q21, [x10, #0x60]\n" + "fmla v9.4s, v20.4s, v0.s[1]\n" + "fmla v13.4s, v20.4s, v1.s[1]\n" + "fmla v17.4s, v20.4s, v2.s[1]\n" + "ldr q20, [x10, #0x70]\n" + "fmla v10.4s, v21.4s, v0.s[1]\n" + "fmla v14.4s, v21.4s, v1.s[1]\n" + "fmla v18.4s, v21.4s, v2.s[1]\n" + "ldr q21, [x10, #0x80]\n" + "fmla v11.4s, v20.4s, v0.s[1]\n" + "fmla v15.4s, v20.4s, v1.s[1]\n" + "fmla v19.4s, v20.4s, v2.s[1]\n" + "ldr q20, [x10, #0x90]\n" + "fmla v8.4s, v21.4s, v0.s[2]\n" + "fmla v12.4s, v21.4s, v1.s[2]\n" + "fmla v16.4s, v21.4s, v2.s[2]\n" + "ldr q21, [x10, #0xa0]\n" + "fmla v9.4s, v20.4s, v0.s[2]\n" + "fmla v13.4s, v20.4s, v1.s[2]\n" + "fmla v17.4s, v20.4s, v2.s[2]\n" + "ldr q20, [x10, #0xb0]\n" + "fmla v10.4s, v21.4s, v0.s[2]\n" + "fmla v14.4s, v21.4s, v1.s[2]\n" + "fmla v18.4s, v21.4s, v2.s[2]\n" + "ldr q21, [x10, #0xc0]\n" + "fmla v11.4s, v20.4s, v0.s[2]\n" + "fmla v15.4s, v20.4s, v1.s[2]\n" + "fmla v19.4s, v20.4s, v2.s[2]\n" + "ldr q20, [x10, #0xd0]\n" + "fmla v8.4s, v21.4s, v0.s[3]\n" + "fmla v12.4s, v21.4s, v1.s[3]\n" + "fmla v16.4s, v21.4s, v2.s[3]\n" + "ldr q21, [x10, #0xe0]\n" + "fmla v9.4s, v20.4s, v0.s[3]\n" + "fmla v13.4s, v20.4s, v1.s[3]\n" + "fmla v17.4s, v20.4s, v2.s[3]\n" + "ldr q20, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - "fmla v10.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" - "fmla v18.4s, v6.4s, v2.s[3]\n" - "fmla v11.4s, v7.4s, v0.s[3]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" - "fmla v19.4s, v7.4s, v2.s[3]\n" + "fmla v10.4s, v21.4s, v0.s[3]\n" + "fmla v14.4s, v21.4s, v1.s[3]\n" + "fmla v18.4s, v21.4s, v2.s[3]\n" + "fmla v11.4s, v20.4s, v0.s[3]\n" + "fmla v15.4s, v20.4s, v1.s[3]\n" + "fmla v19.4s, v20.4s, v2.s[3]\n" "86:" // Height 3: Multiply loop: Main loop skip "cbz x27, 88f\n" "87:" // 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, #0x1\n" - "ldr s2, [x24], #0x4\n" - "ldr q6, [x10, #0x0]\n" - "fmla v8.4s, v6.4s, v0.s[0]\n" - "fmla v12.4s, v6.4s, v1.s[0]\n" - "ldr q7, [x10, #0x10]\n" - "fmla v16.4s, v6.4s, v2.s[0]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.4s, v7.4s, v0.s[0]\n" - "fmla v13.4s, v7.4s, v1.s[0]\n" - "fmla v17.4s, v7.4s, v2.s[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr s22, [x24], #0x4\n" + "ldr q21, [x10, #0x0]\n" + "fmla v8.4s, v21.4s, v24.s[0]\n" + "fmla v12.4s, v21.4s, v23.s[0]\n" + "ldr q20, [x10, #0x10]\n" + "fmla v16.4s, v21.4s, v22.s[0]\n" + "ldr q21, [x10, #0x20]\n" + "fmla v9.4s, v20.4s, v24.s[0]\n" + "fmla v13.4s, v20.4s, v23.s[0]\n" + "fmla v17.4s, v20.4s, v22.s[0]\n" + "ldr q20, [x10, #0x30]\n" "add x10, x10, #0x40\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "fmla v19.4s, v7.4s, v2.s[0]\n" + "fmla v10.4s, v21.4s, v24.s[0]\n" + "fmla v14.4s, v21.4s, v23.s[0]\n" + "fmla v18.4s, v21.4s, v22.s[0]\n" + "fmla v11.4s, v20.4s, v24.s[0]\n" + "fmla v15.4s, v20.4s, v23.s[0]\n" + "fmla v19.4s, v20.4s, v22.s[0]\n" "cbnz x27, 87b\n" "88:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1119,33 +1118,33 @@ void a64_hybrid_fp32_mla_6x16 ( "prfm pstl1keep, [x24, #0x0]\n" "tbz %x[flags], #1, 89f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v21.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmin v16.4s, v16.4s, v1.4s\n" - "fmin v17.4s, v17.4s, v1.4s\n" - "fmin v18.4s, v18.4s, v1.4s\n" - "fmin v19.4s, v19.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" - "fmax v16.4s, v16.4s, v0.4s\n" - "fmax v17.4s, v17.4s, v0.4s\n" - "fmax v18.4s, v18.4s, v0.4s\n" - "fmax v19.4s, v19.4s, v0.4s\n" + "ld1r { v20.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v21.4s\n" + "fmin v9.4s, v9.4s, v21.4s\n" + "fmin v10.4s, v10.4s, v21.4s\n" + "fmin v11.4s, v11.4s, v21.4s\n" + "fmin v12.4s, v12.4s, v21.4s\n" + "fmin v13.4s, v13.4s, v21.4s\n" + "fmin v14.4s, v14.4s, v21.4s\n" + "fmin v15.4s, v15.4s, v21.4s\n" + "fmin v16.4s, v16.4s, v21.4s\n" + "fmin v17.4s, v17.4s, v21.4s\n" + "fmin v18.4s, v18.4s, v21.4s\n" + "fmin v19.4s, v19.4s, v21.4s\n" + "fmax v8.4s, v8.4s, v20.4s\n" + "fmax v9.4s, v9.4s, v20.4s\n" + "fmax v10.4s, v10.4s, v20.4s\n" + "fmax v11.4s, v11.4s, v20.4s\n" + "fmax v12.4s, v12.4s, v20.4s\n" + "fmax v13.4s, v13.4s, v20.4s\n" + "fmax v14.4s, v14.4s, v20.4s\n" + "fmax v15.4s, v15.4s, v20.4s\n" + "fmax v16.4s, v16.4s, v20.4s\n" + "fmax v17.4s, v17.4s, v20.4s\n" + "fmax v18.4s, v18.4s, v20.4s\n" + "fmax v19.4s, v19.4s, v20.4s\n" "89:" // Height 3: No activation "cmp x11, #0x10\n" "bge 98f\n" @@ -1421,14 +1420,14 @@ void a64_hybrid_fp32_mla_6x16 ( "114:" // 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, 115f\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, 116f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1438,9 +1437,9 @@ void a64_hybrid_fp32_mla_6x16 ( "b 116f\n" "115:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" "116:" // Height 4: input setup done "cmp x27, #0x4\n" "blt 119f\n" @@ -1459,7 +1458,7 @@ void a64_hybrid_fp32_mla_6x16 ( "add x26, x26, #0x10\n" "fmla v16.4s, v6.4s, v2.s[0]\n" "fmla v20.4s, v6.4s, v3.s[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q25, [x10, #0x20]\n" "add x25, x25, #0x10\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "fmla v13.4s, v7.4s, v1.s[0]\n" @@ -1467,85 +1466,85 @@ void a64_hybrid_fp32_mla_6x16 ( "add x23, x23, #0x10\n" "fmla v17.4s, v7.4s, v2.s[0]\n" "fmla v21.4s, v7.4s, v3.s[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q24, [x10, #0x30]\n" "cmp x27, #0x8\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" + "fmla v10.4s, v25.4s, v0.s[0]\n" + "fmla v14.4s, v25.4s, v1.s[0]\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "fmla v22.4s, v6.4s, v3.s[0]\n" - "ldr q6, [x10, #0x40]\n" + "fmla v18.4s, v25.4s, v2.s[0]\n" + "fmla v22.4s, v25.4s, v3.s[0]\n" + "ldr q25, [x10, #0x40]\n" "prfm pldl1keep, [x24, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" + "fmla v11.4s, v24.4s, v0.s[0]\n" + "fmla v15.4s, v24.4s, v1.s[0]\n" "prfm pldl1keep, [x23, #0x80]\n" - "fmla v19.4s, v7.4s, v2.s[0]\n" - "fmla v23.4s, v7.4s, v3.s[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v12.4s, v6.4s, v1.s[1]\n" - "fmla v16.4s, v6.4s, v2.s[1]\n" - "fmla v20.4s, v6.4s, v3.s[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v13.4s, v7.4s, v1.s[1]\n" - "fmla v17.4s, v7.4s, v2.s[1]\n" - "fmla v21.4s, v7.4s, v3.s[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "fmla v18.4s, v6.4s, v2.s[1]\n" - "fmla v22.4s, v6.4s, v3.s[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "fmla v19.4s, v7.4s, v2.s[1]\n" - "fmla v23.4s, v7.4s, v3.s[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "fmla v12.4s, v6.4s, v1.s[2]\n" - "fmla v16.4s, v6.4s, v2.s[2]\n" - "fmla v20.4s, v6.4s, v3.s[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "fmla v13.4s, v7.4s, v1.s[2]\n" - "fmla v17.4s, v7.4s, v2.s[2]\n" - "fmla v21.4s, v7.4s, v3.s[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v14.4s, v6.4s, v1.s[2]\n" - "fmla v18.4s, v6.4s, v2.s[2]\n" - "fmla v22.4s, v6.4s, v3.s[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "fmla v19.4s, v7.4s, v2.s[2]\n" - "fmla v23.4s, v7.4s, v3.s[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v12.4s, v6.4s, v1.s[3]\n" - "fmla v16.4s, v6.4s, v2.s[3]\n" - "fmla v20.4s, v6.4s, v3.s[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "fmla v17.4s, v7.4s, v2.s[3]\n" - "fmla v21.4s, v7.4s, v3.s[3]\n" - "ldr q7, [x10, #0xf0]\n" + "fmla v19.4s, v24.4s, v2.s[0]\n" + "fmla v23.4s, v24.4s, v3.s[0]\n" + "ldr q24, [x10, #0x50]\n" + "fmla v8.4s, v25.4s, v0.s[1]\n" + "fmla v12.4s, v25.4s, v1.s[1]\n" + "fmla v16.4s, v25.4s, v2.s[1]\n" + "fmla v20.4s, v25.4s, v3.s[1]\n" + "ldr q25, [x10, #0x60]\n" + "fmla v9.4s, v24.4s, v0.s[1]\n" + "fmla v13.4s, v24.4s, v1.s[1]\n" + "fmla v17.4s, v24.4s, v2.s[1]\n" + "fmla v21.4s, v24.4s, v3.s[1]\n" + "ldr q24, [x10, #0x70]\n" + "fmla v10.4s, v25.4s, v0.s[1]\n" + "fmla v14.4s, v25.4s, v1.s[1]\n" + "fmla v18.4s, v25.4s, v2.s[1]\n" + "fmla v22.4s, v25.4s, v3.s[1]\n" + "ldr q25, [x10, #0x80]\n" + "fmla v11.4s, v24.4s, v0.s[1]\n" + "fmla v15.4s, v24.4s, v1.s[1]\n" + "fmla v19.4s, v24.4s, v2.s[1]\n" + "fmla v23.4s, v24.4s, v3.s[1]\n" + "ldr q24, [x10, #0x90]\n" + "fmla v8.4s, v25.4s, v0.s[2]\n" + "fmla v12.4s, v25.4s, v1.s[2]\n" + "fmla v16.4s, v25.4s, v2.s[2]\n" + "fmla v20.4s, v25.4s, v3.s[2]\n" + "ldr q25, [x10, #0xa0]\n" + "fmla v9.4s, v24.4s, v0.s[2]\n" + "fmla v13.4s, v24.4s, v1.s[2]\n" + "fmla v17.4s, v24.4s, v2.s[2]\n" + "fmla v21.4s, v24.4s, v3.s[2]\n" + "ldr q24, [x10, #0xb0]\n" + "fmla v10.4s, v25.4s, v0.s[2]\n" + "fmla v14.4s, v25.4s, v1.s[2]\n" + "fmla v18.4s, v25.4s, v2.s[2]\n" + "fmla v22.4s, v25.4s, v3.s[2]\n" + "ldr q25, [x10, #0xc0]\n" + "fmla v11.4s, v24.4s, v0.s[2]\n" + "fmla v15.4s, v24.4s, v1.s[2]\n" + "fmla v19.4s, v24.4s, v2.s[2]\n" + "fmla v23.4s, v24.4s, v3.s[2]\n" + "ldr q24, [x10, #0xd0]\n" + "fmla v8.4s, v25.4s, v0.s[3]\n" + "fmla v12.4s, v25.4s, v1.s[3]\n" + "fmla v16.4s, v25.4s, v2.s[3]\n" + "fmla v20.4s, v25.4s, v3.s[3]\n" + "ldr q25, [x10, #0xe0]\n" + "fmla v9.4s, v24.4s, v0.s[3]\n" + "fmla v13.4s, v24.4s, v1.s[3]\n" + "fmla v17.4s, v24.4s, v2.s[3]\n" + "fmla v21.4s, v24.4s, v3.s[3]\n" + "ldr q24, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - "fmla v10.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" - "fmla v18.4s, v6.4s, v2.s[3]\n" - "fmla v22.4s, v6.4s, v3.s[3]\n" + "fmla v10.4s, v25.4s, v0.s[3]\n" + "fmla v14.4s, v25.4s, v1.s[3]\n" + "fmla v18.4s, v25.4s, v2.s[3]\n" + "fmla v22.4s, v25.4s, v3.s[3]\n" "ldr q6, [x10, #0x0]\n" - "fmla v11.4s, v7.4s, v0.s[3]\n" + "fmla v11.4s, v24.4s, v0.s[3]\n" "ldr q0, [x26, #0x0]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" + "fmla v15.4s, v24.4s, v1.s[3]\n" "ldr q1, [x25, #0x0]\n" - "fmla v19.4s, v7.4s, v2.s[3]\n" + "fmla v19.4s, v24.4s, v2.s[3]\n" "ldr q2, [x24, #0x0]\n" - "fmla v23.4s, v7.4s, v3.s[3]\n" + "fmla v23.4s, v24.4s, v3.s[3]\n" "ldr q3, [x23, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 117b\n" @@ -1556,7 +1555,7 @@ void a64_hybrid_fp32_mla_6x16 ( "add x25, x25, #0x10\n" "fmla v16.4s, v6.4s, v2.s[0]\n" "fmla v20.4s, v6.4s, v3.s[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q25, [x10, #0x20]\n" "add x24, x24, #0x10\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "fmla v13.4s, v7.4s, v1.s[0]\n" @@ -1564,109 +1563,109 @@ void a64_hybrid_fp32_mla_6x16 ( "sub x27, x27, #0x4\n" "fmla v17.4s, v7.4s, v2.s[0]\n" "fmla v21.4s, v7.4s, v3.s[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q24, [x10, #0x30]\n" "prfm pldl1keep, [x26, #0x80]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" + "fmla v10.4s, v25.4s, v0.s[0]\n" + "fmla v14.4s, v25.4s, v1.s[0]\n" "prfm pldl1keep, [x25, #0x80]\n" "prfm pldl1keep, [x24, #0x80]\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "fmla v22.4s, v6.4s, v3.s[0]\n" - "ldr q6, [x10, #0x40]\n" + "fmla v18.4s, v25.4s, v2.s[0]\n" + "fmla v22.4s, v25.4s, v3.s[0]\n" + "ldr q25, [x10, #0x40]\n" "prfm pldl1keep, [x23, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "fmla v19.4s, v7.4s, v2.s[0]\n" - "fmla v23.4s, v7.4s, v3.s[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v12.4s, v6.4s, v1.s[1]\n" - "fmla v16.4s, v6.4s, v2.s[1]\n" - "fmla v20.4s, v6.4s, v3.s[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v13.4s, v7.4s, v1.s[1]\n" - "fmla v17.4s, v7.4s, v2.s[1]\n" - "fmla v21.4s, v7.4s, v3.s[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "fmla v18.4s, v6.4s, v2.s[1]\n" - "fmla v22.4s, v6.4s, v3.s[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "fmla v19.4s, v7.4s, v2.s[1]\n" - "fmla v23.4s, v7.4s, v3.s[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "fmla v12.4s, v6.4s, v1.s[2]\n" - "fmla v16.4s, v6.4s, v2.s[2]\n" - "fmla v20.4s, v6.4s, v3.s[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "fmla v13.4s, v7.4s, v1.s[2]\n" - "fmla v17.4s, v7.4s, v2.s[2]\n" - "fmla v21.4s, v7.4s, v3.s[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v14.4s, v6.4s, v1.s[2]\n" - "fmla v18.4s, v6.4s, v2.s[2]\n" - "fmla v22.4s, v6.4s, v3.s[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "fmla v19.4s, v7.4s, v2.s[2]\n" - "fmla v23.4s, v7.4s, v3.s[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v12.4s, v6.4s, v1.s[3]\n" - "fmla v16.4s, v6.4s, v2.s[3]\n" - "fmla v20.4s, v6.4s, v3.s[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "fmla v17.4s, v7.4s, v2.s[3]\n" - "fmla v21.4s, v7.4s, v3.s[3]\n" - "ldr q7, [x10, #0xf0]\n" + "fmla v11.4s, v24.4s, v0.s[0]\n" + "fmla v15.4s, v24.4s, v1.s[0]\n" + "fmla v19.4s, v24.4s, v2.s[0]\n" + "fmla v23.4s, v24.4s, v3.s[0]\n" + "ldr q24, [x10, #0x50]\n" + "fmla v8.4s, v25.4s, v0.s[1]\n" + "fmla v12.4s, v25.4s, v1.s[1]\n" + "fmla v16.4s, v25.4s, v2.s[1]\n" + "fmla v20.4s, v25.4s, v3.s[1]\n" + "ldr q25, [x10, #0x60]\n" + "fmla v9.4s, v24.4s, v0.s[1]\n" + "fmla v13.4s, v24.4s, v1.s[1]\n" + "fmla v17.4s, v24.4s, v2.s[1]\n" + "fmla v21.4s, v24.4s, v3.s[1]\n" + "ldr q24, [x10, #0x70]\n" + "fmla v10.4s, v25.4s, v0.s[1]\n" + "fmla v14.4s, v25.4s, v1.s[1]\n" + "fmla v18.4s, v25.4s, v2.s[1]\n" + "fmla v22.4s, v25.4s, v3.s[1]\n" + "ldr q25, [x10, #0x80]\n" + "fmla v11.4s, v24.4s, v0.s[1]\n" + "fmla v15.4s, v24.4s, v1.s[1]\n" + "fmla v19.4s, v24.4s, v2.s[1]\n" + "fmla v23.4s, v24.4s, v3.s[1]\n" + "ldr q24, [x10, #0x90]\n" + "fmla v8.4s, v25.4s, v0.s[2]\n" + "fmla v12.4s, v25.4s, v1.s[2]\n" + "fmla v16.4s, v25.4s, v2.s[2]\n" + "fmla v20.4s, v25.4s, v3.s[2]\n" + "ldr q25, [x10, #0xa0]\n" + "fmla v9.4s, v24.4s, v0.s[2]\n" + "fmla v13.4s, v24.4s, v1.s[2]\n" + "fmla v17.4s, v24.4s, v2.s[2]\n" + "fmla v21.4s, v24.4s, v3.s[2]\n" + "ldr q24, [x10, #0xb0]\n" + "fmla v10.4s, v25.4s, v0.s[2]\n" + "fmla v14.4s, v25.4s, v1.s[2]\n" + "fmla v18.4s, v25.4s, v2.s[2]\n" + "fmla v22.4s, v25.4s, v3.s[2]\n" + "ldr q25, [x10, #0xc0]\n" + "fmla v11.4s, v24.4s, v0.s[2]\n" + "fmla v15.4s, v24.4s, v1.s[2]\n" + "fmla v19.4s, v24.4s, v2.s[2]\n" + "fmla v23.4s, v24.4s, v3.s[2]\n" + "ldr q24, [x10, #0xd0]\n" + "fmla v8.4s, v25.4s, v0.s[3]\n" + "fmla v12.4s, v25.4s, v1.s[3]\n" + "fmla v16.4s, v25.4s, v2.s[3]\n" + "fmla v20.4s, v25.4s, v3.s[3]\n" + "ldr q25, [x10, #0xe0]\n" + "fmla v9.4s, v24.4s, v0.s[3]\n" + "fmla v13.4s, v24.4s, v1.s[3]\n" + "fmla v17.4s, v24.4s, v2.s[3]\n" + "fmla v21.4s, v24.4s, v3.s[3]\n" + "ldr q24, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - "fmla v10.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" - "fmla v18.4s, v6.4s, v2.s[3]\n" - "fmla v22.4s, v6.4s, v3.s[3]\n" - "fmla v11.4s, v7.4s, v0.s[3]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" - "fmla v19.4s, v7.4s, v2.s[3]\n" - "fmla v23.4s, v7.4s, v3.s[3]\n" + "fmla v10.4s, v25.4s, v0.s[3]\n" + "fmla v14.4s, v25.4s, v1.s[3]\n" + "fmla v18.4s, v25.4s, v2.s[3]\n" + "fmla v22.4s, v25.4s, v3.s[3]\n" + "fmla v11.4s, v24.4s, v0.s[3]\n" + "fmla v15.4s, v24.4s, v1.s[3]\n" + "fmla v19.4s, v24.4s, v2.s[3]\n" + "fmla v23.4s, v24.4s, v3.s[3]\n" "119:" // Height 4: Multiply loop: Main loop skip "cbz x27, 121f\n" "120:" // 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, #0x1\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - "fmla v8.4s, v6.4s, v0.s[0]\n" - "fmla v12.4s, v6.4s, v1.s[0]\n" - "fmla v16.4s, v6.4s, v2.s[0]\n" - "fmla v20.4s, v6.4s, v3.s[0]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.4s, v7.4s, v0.s[0]\n" - "fmla v13.4s, v7.4s, v1.s[0]\n" - "fmla v17.4s, v7.4s, v2.s[0]\n" - "fmla v21.4s, v7.4s, v3.s[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" + "fmla v8.4s, v25.4s, v29.s[0]\n" + "fmla v12.4s, v25.4s, v28.s[0]\n" + "fmla v16.4s, v25.4s, v27.s[0]\n" + "fmla v20.4s, v25.4s, v26.s[0]\n" + "ldr q25, [x10, #0x20]\n" + "fmla v9.4s, v24.4s, v29.s[0]\n" + "fmla v13.4s, v24.4s, v28.s[0]\n" + "fmla v17.4s, v24.4s, v27.s[0]\n" + "fmla v21.4s, v24.4s, v26.s[0]\n" + "ldr q24, [x10, #0x30]\n" "add x10, x10, #0x40\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "fmla v22.4s, v6.4s, v3.s[0]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "fmla v19.4s, v7.4s, v2.s[0]\n" - "fmla v23.4s, v7.4s, v3.s[0]\n" + "fmla v10.4s, v25.4s, v29.s[0]\n" + "fmla v14.4s, v25.4s, v28.s[0]\n" + "fmla v18.4s, v25.4s, v27.s[0]\n" + "fmla v22.4s, v25.4s, v26.s[0]\n" + "fmla v11.4s, v24.4s, v29.s[0]\n" + "fmla v15.4s, v24.4s, v28.s[0]\n" + "fmla v19.4s, v24.4s, v27.s[0]\n" + "fmla v23.4s, v24.4s, v26.s[0]\n" "cbnz x27, 120b\n" "121:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1683,41 +1682,41 @@ void a64_hybrid_fp32_mla_6x16 ( "prfm pstl1keep, [x23, #0x0]\n" "tbz %x[flags], #1, 122f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v25.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmin v16.4s, v16.4s, v1.4s\n" - "fmin v17.4s, v17.4s, v1.4s\n" - "fmin v18.4s, v18.4s, v1.4s\n" - "fmin v19.4s, v19.4s, v1.4s\n" - "fmin v20.4s, v20.4s, v1.4s\n" - "fmin v21.4s, v21.4s, v1.4s\n" - "fmin v22.4s, v22.4s, v1.4s\n" - "fmin v23.4s, v23.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" - "fmax v16.4s, v16.4s, v0.4s\n" - "fmax v17.4s, v17.4s, v0.4s\n" - "fmax v18.4s, v18.4s, v0.4s\n" - "fmax v19.4s, v19.4s, v0.4s\n" - "fmax v20.4s, v20.4s, v0.4s\n" - "fmax v21.4s, v21.4s, v0.4s\n" - "fmax v22.4s, v22.4s, v0.4s\n" - "fmax v23.4s, v23.4s, v0.4s\n" + "ld1r { v24.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v25.4s\n" + "fmin v9.4s, v9.4s, v25.4s\n" + "fmin v10.4s, v10.4s, v25.4s\n" + "fmin v11.4s, v11.4s, v25.4s\n" + "fmin v12.4s, v12.4s, v25.4s\n" + "fmin v13.4s, v13.4s, v25.4s\n" + "fmin v14.4s, v14.4s, v25.4s\n" + "fmin v15.4s, v15.4s, v25.4s\n" + "fmin v16.4s, v16.4s, v25.4s\n" + "fmin v17.4s, v17.4s, v25.4s\n" + "fmin v18.4s, v18.4s, v25.4s\n" + "fmin v19.4s, v19.4s, v25.4s\n" + "fmin v20.4s, v20.4s, v25.4s\n" + "fmin v21.4s, v21.4s, v25.4s\n" + "fmin v22.4s, v22.4s, v25.4s\n" + "fmin v23.4s, v23.4s, v25.4s\n" + "fmax v8.4s, v8.4s, v24.4s\n" + "fmax v9.4s, v9.4s, v24.4s\n" + "fmax v10.4s, v10.4s, v24.4s\n" + "fmax v11.4s, v11.4s, v24.4s\n" + "fmax v12.4s, v12.4s, v24.4s\n" + "fmax v13.4s, v13.4s, v24.4s\n" + "fmax v14.4s, v14.4s, v24.4s\n" + "fmax v15.4s, v15.4s, v24.4s\n" + "fmax v16.4s, v16.4s, v24.4s\n" + "fmax v17.4s, v17.4s, v24.4s\n" + "fmax v18.4s, v18.4s, v24.4s\n" + "fmax v19.4s, v19.4s, v24.4s\n" + "fmax v20.4s, v20.4s, v24.4s\n" + "fmax v21.4s, v21.4s, v24.4s\n" + "fmax v22.4s, v22.4s, v24.4s\n" + "fmax v23.4s, v23.4s, v24.4s\n" "122:" // Height 4: No activation "cmp x11, #0x10\n" "bge 131f\n" @@ -2042,15 +2041,15 @@ void a64_hybrid_fp32_mla_6x16 ( "147:" // Height 5: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 148f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" + "ldr x24, [x20, #0x10]\n" + "ldr x23, [x20, #0x18]\n" + "ldr x22, [x20, #0x20]\n" "cbnz x28, 149f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -2061,10 +2060,10 @@ void a64_hybrid_fp32_mla_6x16 ( "b 149f\n" "148:" // Height 5: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" "149:" // Height 5: input setup done "cmp x27, #0x4\n" "blt 152f\n" @@ -2087,7 +2086,7 @@ void a64_hybrid_fp32_mla_6x16 ( "add x25, x25, #0x10\n" "add x24, x24, #0x10\n" "fmla v24.4s, v6.4s, v4.s[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q29, [x10, #0x20]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "add x23, x23, #0x10\n" "fmla v13.4s, v7.4s, v1.s[0]\n" @@ -2096,100 +2095,100 @@ void a64_hybrid_fp32_mla_6x16 ( "cmp x27, #0x8\n" "fmla v21.4s, v7.4s, v3.s[0]\n" "fmla v25.4s, v7.4s, v4.s[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q28, [x10, #0x30]\n" "prfm pldl1keep, [x26, #0x80]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" + "fmla v10.4s, v29.4s, v0.s[0]\n" + "fmla v14.4s, v29.4s, v1.s[0]\n" "prfm pldl1keep, [x25, #0x80]\n" "prfm pldl1keep, [x24, #0x80]\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "fmla v22.4s, v6.4s, v3.s[0]\n" + "fmla v18.4s, v29.4s, v2.s[0]\n" + "fmla v22.4s, v29.4s, v3.s[0]\n" "prfm pldl1keep, [x23, #0x80]\n" "prfm pldl1keep, [x22, #0x80]\n" - "fmla v26.4s, v6.4s, v4.s[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "fmla v19.4s, v7.4s, v2.s[0]\n" - "fmla v23.4s, v7.4s, v3.s[0]\n" - "fmla v27.4s, v7.4s, v4.s[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v12.4s, v6.4s, v1.s[1]\n" - "fmla v16.4s, v6.4s, v2.s[1]\n" - "fmla v20.4s, v6.4s, v3.s[1]\n" - "fmla v24.4s, v6.4s, v4.s[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v13.4s, v7.4s, v1.s[1]\n" - "fmla v17.4s, v7.4s, v2.s[1]\n" - "fmla v21.4s, v7.4s, v3.s[1]\n" - "fmla v25.4s, v7.4s, v4.s[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "fmla v18.4s, v6.4s, v2.s[1]\n" - "fmla v22.4s, v6.4s, v3.s[1]\n" - "fmla v26.4s, v6.4s, v4.s[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "fmla v19.4s, v7.4s, v2.s[1]\n" - "fmla v23.4s, v7.4s, v3.s[1]\n" - "fmla v27.4s, v7.4s, v4.s[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "fmla v12.4s, v6.4s, v1.s[2]\n" - "fmla v16.4s, v6.4s, v2.s[2]\n" - "fmla v20.4s, v6.4s, v3.s[2]\n" - "fmla v24.4s, v6.4s, v4.s[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "fmla v13.4s, v7.4s, v1.s[2]\n" - "fmla v17.4s, v7.4s, v2.s[2]\n" - "fmla v21.4s, v7.4s, v3.s[2]\n" - "fmla v25.4s, v7.4s, v4.s[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v14.4s, v6.4s, v1.s[2]\n" - "fmla v18.4s, v6.4s, v2.s[2]\n" - "fmla v22.4s, v6.4s, v3.s[2]\n" - "fmla v26.4s, v6.4s, v4.s[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "fmla v19.4s, v7.4s, v2.s[2]\n" - "fmla v23.4s, v7.4s, v3.s[2]\n" - "fmla v27.4s, v7.4s, v4.s[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v12.4s, v6.4s, v1.s[3]\n" - "fmla v16.4s, v6.4s, v2.s[3]\n" - "fmla v20.4s, v6.4s, v3.s[3]\n" - "fmla v24.4s, v6.4s, v4.s[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "fmla v17.4s, v7.4s, v2.s[3]\n" - "fmla v21.4s, v7.4s, v3.s[3]\n" - "fmla v25.4s, v7.4s, v4.s[3]\n" - "ldr q7, [x10, #0xf0]\n" + "fmla v26.4s, v29.4s, v4.s[0]\n" + "ldr q29, [x10, #0x40]\n" + "fmla v11.4s, v28.4s, v0.s[0]\n" + "fmla v15.4s, v28.4s, v1.s[0]\n" + "fmla v19.4s, v28.4s, v2.s[0]\n" + "fmla v23.4s, v28.4s, v3.s[0]\n" + "fmla v27.4s, v28.4s, v4.s[0]\n" + "ldr q28, [x10, #0x50]\n" + "fmla v8.4s, v29.4s, v0.s[1]\n" + "fmla v12.4s, v29.4s, v1.s[1]\n" + "fmla v16.4s, v29.4s, v2.s[1]\n" + "fmla v20.4s, v29.4s, v3.s[1]\n" + "fmla v24.4s, v29.4s, v4.s[1]\n" + "ldr q29, [x10, #0x60]\n" + "fmla v9.4s, v28.4s, v0.s[1]\n" + "fmla v13.4s, v28.4s, v1.s[1]\n" + "fmla v17.4s, v28.4s, v2.s[1]\n" + "fmla v21.4s, v28.4s, v3.s[1]\n" + "fmla v25.4s, v28.4s, v4.s[1]\n" + "ldr q28, [x10, #0x70]\n" + "fmla v10.4s, v29.4s, v0.s[1]\n" + "fmla v14.4s, v29.4s, v1.s[1]\n" + "fmla v18.4s, v29.4s, v2.s[1]\n" + "fmla v22.4s, v29.4s, v3.s[1]\n" + "fmla v26.4s, v29.4s, v4.s[1]\n" + "ldr q29, [x10, #0x80]\n" + "fmla v11.4s, v28.4s, v0.s[1]\n" + "fmla v15.4s, v28.4s, v1.s[1]\n" + "fmla v19.4s, v28.4s, v2.s[1]\n" + "fmla v23.4s, v28.4s, v3.s[1]\n" + "fmla v27.4s, v28.4s, v4.s[1]\n" + "ldr q28, [x10, #0x90]\n" + "fmla v8.4s, v29.4s, v0.s[2]\n" + "fmla v12.4s, v29.4s, v1.s[2]\n" + "fmla v16.4s, v29.4s, v2.s[2]\n" + "fmla v20.4s, v29.4s, v3.s[2]\n" + "fmla v24.4s, v29.4s, v4.s[2]\n" + "ldr q29, [x10, #0xa0]\n" + "fmla v9.4s, v28.4s, v0.s[2]\n" + "fmla v13.4s, v28.4s, v1.s[2]\n" + "fmla v17.4s, v28.4s, v2.s[2]\n" + "fmla v21.4s, v28.4s, v3.s[2]\n" + "fmla v25.4s, v28.4s, v4.s[2]\n" + "ldr q28, [x10, #0xb0]\n" + "fmla v10.4s, v29.4s, v0.s[2]\n" + "fmla v14.4s, v29.4s, v1.s[2]\n" + "fmla v18.4s, v29.4s, v2.s[2]\n" + "fmla v22.4s, v29.4s, v3.s[2]\n" + "fmla v26.4s, v29.4s, v4.s[2]\n" + "ldr q29, [x10, #0xc0]\n" + "fmla v11.4s, v28.4s, v0.s[2]\n" + "fmla v15.4s, v28.4s, v1.s[2]\n" + "fmla v19.4s, v28.4s, v2.s[2]\n" + "fmla v23.4s, v28.4s, v3.s[2]\n" + "fmla v27.4s, v28.4s, v4.s[2]\n" + "ldr q28, [x10, #0xd0]\n" + "fmla v8.4s, v29.4s, v0.s[3]\n" + "fmla v12.4s, v29.4s, v1.s[3]\n" + "fmla v16.4s, v29.4s, v2.s[3]\n" + "fmla v20.4s, v29.4s, v3.s[3]\n" + "fmla v24.4s, v29.4s, v4.s[3]\n" + "ldr q29, [x10, #0xe0]\n" + "fmla v9.4s, v28.4s, v0.s[3]\n" + "fmla v13.4s, v28.4s, v1.s[3]\n" + "fmla v17.4s, v28.4s, v2.s[3]\n" + "fmla v21.4s, v28.4s, v3.s[3]\n" + "fmla v25.4s, v28.4s, v4.s[3]\n" + "ldr q28, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - "fmla v10.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" - "fmla v18.4s, v6.4s, v2.s[3]\n" - "fmla v22.4s, v6.4s, v3.s[3]\n" - "fmla v26.4s, v6.4s, v4.s[3]\n" + "fmla v10.4s, v29.4s, v0.s[3]\n" + "fmla v14.4s, v29.4s, v1.s[3]\n" + "fmla v18.4s, v29.4s, v2.s[3]\n" + "fmla v22.4s, v29.4s, v3.s[3]\n" + "fmla v26.4s, v29.4s, v4.s[3]\n" "ldr q6, [x10, #0x0]\n" - "fmla v11.4s, v7.4s, v0.s[3]\n" + "fmla v11.4s, v28.4s, v0.s[3]\n" "ldr q0, [x26, #0x0]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" + "fmla v15.4s, v28.4s, v1.s[3]\n" "ldr q1, [x25, #0x0]\n" - "fmla v19.4s, v7.4s, v2.s[3]\n" + "fmla v19.4s, v28.4s, v2.s[3]\n" "ldr q2, [x24, #0x0]\n" - "fmla v23.4s, v7.4s, v3.s[3]\n" + "fmla v23.4s, v28.4s, v3.s[3]\n" "ldr q3, [x23, #0x0]\n" - "fmla v27.4s, v7.4s, v4.s[3]\n" + "fmla v27.4s, v28.4s, v4.s[3]\n" "ldr q4, [x22, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 150b\n" @@ -2203,7 +2202,7 @@ void a64_hybrid_fp32_mla_6x16 ( "add x24, x24, #0x10\n" "add x23, x23, #0x10\n" "fmla v24.4s, v6.4s, v4.s[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q29, [x10, #0x20]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "add x22, x22, #0x10\n" "fmla v13.4s, v7.4s, v1.s[0]\n" @@ -2212,128 +2211,128 @@ void a64_hybrid_fp32_mla_6x16 ( "prfm pldl1keep, [x26, #0x80]\n" "fmla v21.4s, v7.4s, v3.s[0]\n" "fmla v25.4s, v7.4s, v4.s[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q28, [x10, #0x30]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" + "fmla v10.4s, v29.4s, v0.s[0]\n" + "fmla v14.4s, v29.4s, v1.s[0]\n" "prfm pldl1keep, [x24, #0x80]\n" "prfm pldl1keep, [x23, #0x80]\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "fmla v22.4s, v6.4s, v3.s[0]\n" + "fmla v18.4s, v29.4s, v2.s[0]\n" + "fmla v22.4s, v29.4s, v3.s[0]\n" "prfm pldl1keep, [x22, #0x80]\n" - "fmla v26.4s, v6.4s, v4.s[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "fmla v19.4s, v7.4s, v2.s[0]\n" - "fmla v23.4s, v7.4s, v3.s[0]\n" - "fmla v27.4s, v7.4s, v4.s[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v12.4s, v6.4s, v1.s[1]\n" - "fmla v16.4s, v6.4s, v2.s[1]\n" - "fmla v20.4s, v6.4s, v3.s[1]\n" - "fmla v24.4s, v6.4s, v4.s[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v13.4s, v7.4s, v1.s[1]\n" - "fmla v17.4s, v7.4s, v2.s[1]\n" - "fmla v21.4s, v7.4s, v3.s[1]\n" - "fmla v25.4s, v7.4s, v4.s[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "fmla v18.4s, v6.4s, v2.s[1]\n" - "fmla v22.4s, v6.4s, v3.s[1]\n" - "fmla v26.4s, v6.4s, v4.s[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "fmla v19.4s, v7.4s, v2.s[1]\n" - "fmla v23.4s, v7.4s, v3.s[1]\n" - "fmla v27.4s, v7.4s, v4.s[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "fmla v12.4s, v6.4s, v1.s[2]\n" - "fmla v16.4s, v6.4s, v2.s[2]\n" - "fmla v20.4s, v6.4s, v3.s[2]\n" - "fmla v24.4s, v6.4s, v4.s[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "fmla v13.4s, v7.4s, v1.s[2]\n" - "fmla v17.4s, v7.4s, v2.s[2]\n" - "fmla v21.4s, v7.4s, v3.s[2]\n" - "fmla v25.4s, v7.4s, v4.s[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v14.4s, v6.4s, v1.s[2]\n" - "fmla v18.4s, v6.4s, v2.s[2]\n" - "fmla v22.4s, v6.4s, v3.s[2]\n" - "fmla v26.4s, v6.4s, v4.s[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "fmla v19.4s, v7.4s, v2.s[2]\n" - "fmla v23.4s, v7.4s, v3.s[2]\n" - "fmla v27.4s, v7.4s, v4.s[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v12.4s, v6.4s, v1.s[3]\n" - "fmla v16.4s, v6.4s, v2.s[3]\n" - "fmla v20.4s, v6.4s, v3.s[3]\n" - "fmla v24.4s, v6.4s, v4.s[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "fmla v17.4s, v7.4s, v2.s[3]\n" - "fmla v21.4s, v7.4s, v3.s[3]\n" - "fmla v25.4s, v7.4s, v4.s[3]\n" - "ldr q7, [x10, #0xf0]\n" + "fmla v26.4s, v29.4s, v4.s[0]\n" + "ldr q29, [x10, #0x40]\n" + "fmla v11.4s, v28.4s, v0.s[0]\n" + "fmla v15.4s, v28.4s, v1.s[0]\n" + "fmla v19.4s, v28.4s, v2.s[0]\n" + "fmla v23.4s, v28.4s, v3.s[0]\n" + "fmla v27.4s, v28.4s, v4.s[0]\n" + "ldr q28, [x10, #0x50]\n" + "fmla v8.4s, v29.4s, v0.s[1]\n" + "fmla v12.4s, v29.4s, v1.s[1]\n" + "fmla v16.4s, v29.4s, v2.s[1]\n" + "fmla v20.4s, v29.4s, v3.s[1]\n" + "fmla v24.4s, v29.4s, v4.s[1]\n" + "ldr q29, [x10, #0x60]\n" + "fmla v9.4s, v28.4s, v0.s[1]\n" + "fmla v13.4s, v28.4s, v1.s[1]\n" + "fmla v17.4s, v28.4s, v2.s[1]\n" + "fmla v21.4s, v28.4s, v3.s[1]\n" + "fmla v25.4s, v28.4s, v4.s[1]\n" + "ldr q28, [x10, #0x70]\n" + "fmla v10.4s, v29.4s, v0.s[1]\n" + "fmla v14.4s, v29.4s, v1.s[1]\n" + "fmla v18.4s, v29.4s, v2.s[1]\n" + "fmla v22.4s, v29.4s, v3.s[1]\n" + "fmla v26.4s, v29.4s, v4.s[1]\n" + "ldr q29, [x10, #0x80]\n" + "fmla v11.4s, v28.4s, v0.s[1]\n" + "fmla v15.4s, v28.4s, v1.s[1]\n" + "fmla v19.4s, v28.4s, v2.s[1]\n" + "fmla v23.4s, v28.4s, v3.s[1]\n" + "fmla v27.4s, v28.4s, v4.s[1]\n" + "ldr q28, [x10, #0x90]\n" + "fmla v8.4s, v29.4s, v0.s[2]\n" + "fmla v12.4s, v29.4s, v1.s[2]\n" + "fmla v16.4s, v29.4s, v2.s[2]\n" + "fmla v20.4s, v29.4s, v3.s[2]\n" + "fmla v24.4s, v29.4s, v4.s[2]\n" + "ldr q29, [x10, #0xa0]\n" + "fmla v9.4s, v28.4s, v0.s[2]\n" + "fmla v13.4s, v28.4s, v1.s[2]\n" + "fmla v17.4s, v28.4s, v2.s[2]\n" + "fmla v21.4s, v28.4s, v3.s[2]\n" + "fmla v25.4s, v28.4s, v4.s[2]\n" + "ldr q28, [x10, #0xb0]\n" + "fmla v10.4s, v29.4s, v0.s[2]\n" + "fmla v14.4s, v29.4s, v1.s[2]\n" + "fmla v18.4s, v29.4s, v2.s[2]\n" + "fmla v22.4s, v29.4s, v3.s[2]\n" + "fmla v26.4s, v29.4s, v4.s[2]\n" + "ldr q29, [x10, #0xc0]\n" + "fmla v11.4s, v28.4s, v0.s[2]\n" + "fmla v15.4s, v28.4s, v1.s[2]\n" + "fmla v19.4s, v28.4s, v2.s[2]\n" + "fmla v23.4s, v28.4s, v3.s[2]\n" + "fmla v27.4s, v28.4s, v4.s[2]\n" + "ldr q28, [x10, #0xd0]\n" + "fmla v8.4s, v29.4s, v0.s[3]\n" + "fmla v12.4s, v29.4s, v1.s[3]\n" + "fmla v16.4s, v29.4s, v2.s[3]\n" + "fmla v20.4s, v29.4s, v3.s[3]\n" + "fmla v24.4s, v29.4s, v4.s[3]\n" + "ldr q29, [x10, #0xe0]\n" + "fmla v9.4s, v28.4s, v0.s[3]\n" + "fmla v13.4s, v28.4s, v1.s[3]\n" + "fmla v17.4s, v28.4s, v2.s[3]\n" + "fmla v21.4s, v28.4s, v3.s[3]\n" + "fmla v25.4s, v28.4s, v4.s[3]\n" + "ldr q28, [x10, #0xf0]\n" "add x10, x10, #0x100\n" - "fmla v10.4s, v6.4s, v0.s[3]\n" - "fmla v14.4s, v6.4s, v1.s[3]\n" - "fmla v18.4s, v6.4s, v2.s[3]\n" - "fmla v22.4s, v6.4s, v3.s[3]\n" - "fmla v26.4s, v6.4s, v4.s[3]\n" - "fmla v11.4s, v7.4s, v0.s[3]\n" - "fmla v15.4s, v7.4s, v1.s[3]\n" - "fmla v19.4s, v7.4s, v2.s[3]\n" - "fmla v23.4s, v7.4s, v3.s[3]\n" - "fmla v27.4s, v7.4s, v4.s[3]\n" + "fmla v10.4s, v29.4s, v0.s[3]\n" + "fmla v14.4s, v29.4s, v1.s[3]\n" + "fmla v18.4s, v29.4s, v2.s[3]\n" + "fmla v22.4s, v29.4s, v3.s[3]\n" + "fmla v26.4s, v29.4s, v4.s[3]\n" + "fmla v11.4s, v28.4s, v0.s[3]\n" + "fmla v15.4s, v28.4s, v1.s[3]\n" + "fmla v19.4s, v28.4s, v2.s[3]\n" + "fmla v23.4s, v28.4s, v3.s[3]\n" + "fmla v27.4s, v28.4s, v4.s[3]\n" "152:" // Height 5: Multiply loop: Main loop skip "cbz x27, 154f\n" "153:" // 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, #0x1\n" - "ldr s2, [x24], #0x4\n" - "ldr s3, [x23], #0x4\n" - "ldr s4, [x22], #0x4\n" - "ldr q6, [x10, #0x0]\n" - "fmla v8.4s, v6.4s, v0.s[0]\n" - "fmla v12.4s, v6.4s, v1.s[0]\n" - "ldr q7, [x10, #0x10]\n" - "fmla v16.4s, v6.4s, v2.s[0]\n" - "fmla v20.4s, v6.4s, v3.s[0]\n" - "fmla v24.4s, v6.4s, v4.s[0]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.4s, v7.4s, v0.s[0]\n" - "fmla v13.4s, v7.4s, v1.s[0]\n" - "fmla v17.4s, v7.4s, v2.s[0]\n" - "fmla v21.4s, v7.4s, v3.s[0]\n" - "fmla v25.4s, v7.4s, v4.s[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" + "fmla v8.4s, v29.4s, v2.s[0]\n" + "fmla v12.4s, v29.4s, v1.s[0]\n" + "ldr q28, [x10, #0x10]\n" + "fmla v16.4s, v29.4s, v0.s[0]\n" + "fmla v20.4s, v29.4s, v31.s[0]\n" + "fmla v24.4s, v29.4s, v30.s[0]\n" + "ldr q29, [x10, #0x20]\n" + "fmla v9.4s, v28.4s, v2.s[0]\n" + "fmla v13.4s, v28.4s, v1.s[0]\n" + "fmla v17.4s, v28.4s, v0.s[0]\n" + "fmla v21.4s, v28.4s, v31.s[0]\n" + "fmla v25.4s, v28.4s, v30.s[0]\n" + "ldr q28, [x10, #0x30]\n" "add x10, x10, #0x40\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "fmla v22.4s, v6.4s, v3.s[0]\n" - "fmla v26.4s, v6.4s, v4.s[0]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "fmla v19.4s, v7.4s, v2.s[0]\n" - "fmla v23.4s, v7.4s, v3.s[0]\n" - "fmla v27.4s, v7.4s, v4.s[0]\n" + "fmla v10.4s, v29.4s, v2.s[0]\n" + "fmla v14.4s, v29.4s, v1.s[0]\n" + "fmla v18.4s, v29.4s, v0.s[0]\n" + "fmla v22.4s, v29.4s, v31.s[0]\n" + "fmla v26.4s, v29.4s, v30.s[0]\n" + "fmla v11.4s, v28.4s, v2.s[0]\n" + "fmla v15.4s, v28.4s, v1.s[0]\n" + "fmla v19.4s, v28.4s, v0.s[0]\n" + "fmla v23.4s, v28.4s, v31.s[0]\n" + "fmla v27.4s, v28.4s, v30.s[0]\n" "cbnz x27, 153b\n" "154:" // Height 5: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -2352,49 +2351,49 @@ void a64_hybrid_fp32_mla_6x16 ( "prfm pstl1keep, [x22, #0x0]\n" "tbz %x[flags], #1, 155f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.4s }, [x20]\n" + "ld1r { v29.4s }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.4s }, [x20]\n" - "fmin v8.4s, v8.4s, v1.4s\n" - "fmin v9.4s, v9.4s, v1.4s\n" - "fmin v10.4s, v10.4s, v1.4s\n" - "fmin v11.4s, v11.4s, v1.4s\n" - "fmin v12.4s, v12.4s, v1.4s\n" - "fmin v13.4s, v13.4s, v1.4s\n" - "fmin v14.4s, v14.4s, v1.4s\n" - "fmin v15.4s, v15.4s, v1.4s\n" - "fmin v16.4s, v16.4s, v1.4s\n" - "fmin v17.4s, v17.4s, v1.4s\n" - "fmin v18.4s, v18.4s, v1.4s\n" - "fmin v19.4s, v19.4s, v1.4s\n" - "fmin v20.4s, v20.4s, v1.4s\n" - "fmin v21.4s, v21.4s, v1.4s\n" - "fmin v22.4s, v22.4s, v1.4s\n" - "fmin v23.4s, v23.4s, v1.4s\n" - "fmin v24.4s, v24.4s, v1.4s\n" - "fmin v25.4s, v25.4s, v1.4s\n" - "fmin v26.4s, v26.4s, v1.4s\n" - "fmin v27.4s, v27.4s, v1.4s\n" - "fmax v8.4s, v8.4s, v0.4s\n" - "fmax v9.4s, v9.4s, v0.4s\n" - "fmax v10.4s, v10.4s, v0.4s\n" - "fmax v11.4s, v11.4s, v0.4s\n" - "fmax v12.4s, v12.4s, v0.4s\n" - "fmax v13.4s, v13.4s, v0.4s\n" - "fmax v14.4s, v14.4s, v0.4s\n" - "fmax v15.4s, v15.4s, v0.4s\n" - "fmax v16.4s, v16.4s, v0.4s\n" - "fmax v17.4s, v17.4s, v0.4s\n" - "fmax v18.4s, v18.4s, v0.4s\n" - "fmax v19.4s, v19.4s, v0.4s\n" - "fmax v20.4s, v20.4s, v0.4s\n" - "fmax v21.4s, v21.4s, v0.4s\n" - "fmax v22.4s, v22.4s, v0.4s\n" - "fmax v23.4s, v23.4s, v0.4s\n" - "fmax v24.4s, v24.4s, v0.4s\n" - "fmax v25.4s, v25.4s, v0.4s\n" - "fmax v26.4s, v26.4s, v0.4s\n" - "fmax v27.4s, v27.4s, v0.4s\n" + "ld1r { v28.4s }, [x20]\n" + "fmin v8.4s, v8.4s, v29.4s\n" + "fmin v9.4s, v9.4s, v29.4s\n" + "fmin v10.4s, v10.4s, v29.4s\n" + "fmin v11.4s, v11.4s, v29.4s\n" + "fmin v12.4s, v12.4s, v29.4s\n" + "fmin v13.4s, v13.4s, v29.4s\n" + "fmin v14.4s, v14.4s, v29.4s\n" + "fmin v15.4s, v15.4s, v29.4s\n" + "fmin v16.4s, v16.4s, v29.4s\n" + "fmin v17.4s, v17.4s, v29.4s\n" + "fmin v18.4s, v18.4s, v29.4s\n" + "fmin v19.4s, v19.4s, v29.4s\n" + "fmin v20.4s, v20.4s, v29.4s\n" + "fmin v21.4s, v21.4s, v29.4s\n" + "fmin v22.4s, v22.4s, v29.4s\n" + "fmin v23.4s, v23.4s, v29.4s\n" + "fmin v24.4s, v24.4s, v29.4s\n" + "fmin v25.4s, v25.4s, v29.4s\n" + "fmin v26.4s, v26.4s, v29.4s\n" + "fmin v27.4s, v27.4s, v29.4s\n" + "fmax v8.4s, v8.4s, v28.4s\n" + "fmax v9.4s, v9.4s, v28.4s\n" + "fmax v10.4s, v10.4s, v28.4s\n" + "fmax v11.4s, v11.4s, v28.4s\n" + "fmax v12.4s, v12.4s, v28.4s\n" + "fmax v13.4s, v13.4s, v28.4s\n" + "fmax v14.4s, v14.4s, v28.4s\n" + "fmax v15.4s, v15.4s, v28.4s\n" + "fmax v16.4s, v16.4s, v28.4s\n" + "fmax v17.4s, v17.4s, v28.4s\n" + "fmax v18.4s, v18.4s, v28.4s\n" + "fmax v19.4s, v19.4s, v28.4s\n" + "fmax v20.4s, v20.4s, v28.4s\n" + "fmax v21.4s, v21.4s, v28.4s\n" + "fmax v22.4s, v22.4s, v28.4s\n" + "fmax v23.4s, v23.4s, v28.4s\n" + "fmax v24.4s, v24.4s, v28.4s\n" + "fmax v25.4s, v25.4s, v28.4s\n" + "fmax v26.4s, v26.4s, v28.4s\n" + "fmax v27.4s, v27.4s, v28.4s\n" "155:" // Height 5: No activation "cmp x11, #0x10\n" "bge 164f\n" @@ -2771,16 +2770,16 @@ void a64_hybrid_fp32_mla_6x16 ( "180:" // 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, 181f\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, 182f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -2792,11 +2791,11 @@ void a64_hybrid_fp32_mla_6x16 ( "b 182f\n" "181:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #2\n" - "add x24, x25, x20, LSL #2\n" - "add x23, x24, x20, LSL #2\n" - "add x22, x23, x20, LSL #2\n" - "add x21, x22, x20, LSL #2\n" + "add x25, x26, x21, LSL #2\n" + "add x24, x25, x21, LSL #2\n" + "add x23, x24, x21, LSL #2\n" + "add x22, x23, x21, LSL #2\n" + "add x21, x22, x21, LSL #2\n" "182:" // Height 6: input setup done "cmp x27, #0x4\n" "blt 185f\n" @@ -3073,42 +3072,42 @@ void a64_hybrid_fp32_mla_6x16 ( "185:" // Height 6: Multiply loop: Main loop skip "cbz x27, 187f\n" "186:" // 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, #0x1\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" - "fmla v8.4s, v6.4s, v0.s[0]\n" - "fmla v12.4s, v6.4s, v1.s[0]\n" - "fmla v16.4s, v6.4s, v2.s[0]\n" - "fmla v20.4s, v6.4s, v3.s[0]\n" - "fmla v24.4s, v6.4s, v4.s[0]\n" - "fmla v28.4s, v6.4s, v5.s[0]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.4s, v7.4s, v0.s[0]\n" - "fmla v13.4s, v7.4s, v1.s[0]\n" - "fmla v17.4s, v7.4s, v2.s[0]\n" - "fmla v21.4s, v7.4s, v3.s[0]\n" - "fmla v25.4s, v7.4s, v4.s[0]\n" - "fmla v29.4s, v7.4s, v5.s[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" + "fmla v8.4s, v1.4s, v7.s[0]\n" + "fmla v12.4s, v1.4s, v6.s[0]\n" + "fmla v16.4s, v1.4s, v5.s[0]\n" + "fmla v20.4s, v1.4s, v4.s[0]\n" + "fmla v24.4s, v1.4s, v3.s[0]\n" + "fmla v28.4s, v1.4s, v2.s[0]\n" + "ldr q1, [x10, #0x20]\n" + "fmla v9.4s, v0.4s, v7.s[0]\n" + "fmla v13.4s, v0.4s, v6.s[0]\n" + "fmla v17.4s, v0.4s, v5.s[0]\n" + "fmla v21.4s, v0.4s, v4.s[0]\n" + "fmla v25.4s, v0.4s, v3.s[0]\n" + "fmla v29.4s, v0.4s, v2.s[0]\n" + "ldr q0, [x10, #0x30]\n" "add x10, x10, #0x40\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "fmla v22.4s, v6.4s, v3.s[0]\n" - "fmla v26.4s, v6.4s, v4.s[0]\n" - "fmla v30.4s, v6.4s, v5.s[0]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "fmla v19.4s, v7.4s, v2.s[0]\n" - "fmla v23.4s, v7.4s, v3.s[0]\n" - "fmla v27.4s, v7.4s, v4.s[0]\n" - "fmla v31.4s, v7.4s, v5.s[0]\n" + "fmla v10.4s, v1.4s, v7.s[0]\n" + "fmla v14.4s, v1.4s, v6.s[0]\n" + "fmla v18.4s, v1.4s, v5.s[0]\n" + "fmla v22.4s, v1.4s, v4.s[0]\n" + "fmla v26.4s, v1.4s, v3.s[0]\n" + "fmla v30.4s, v1.4s, v2.s[0]\n" + "fmla v11.4s, v0.4s, v7.s[0]\n" + "fmla v15.4s, v0.4s, v6.s[0]\n" + "fmla v19.4s, v0.4s, v5.s[0]\n" + "fmla v23.4s, v0.4s, v4.s[0]\n" + "fmla v27.4s, v0.4s, v3.s[0]\n" + "fmla v31.4s, v0.4s, v2.s[0]\n" "cbnz x27, 186b\n" "187:" // Height 6: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -3350,7 +3349,6 @@ void a64_hybrid_fp32_mla_6x16 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "200:" // Exit - : [M] "+&r" (M), [input_ptr] "+&r" (input_ptr), [output_ptr] "+&r" (output_ptr) : [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)) : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" |