diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp | 1744 |
1 files changed, 872 insertions, 872 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp index e0fbe17bad..b1cd6dc970 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_ffhybrid_fp32_mla_6x16/generic.cpp @@ -209,11 +209,11 @@ void a64_ffhybrid_fp32_mla_6x16 ( "16:" // 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, 17f\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, 18f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -230,37 +230,37 @@ void a64_ffhybrid_fp32_mla_6x16 ( "blt 20f\n" "19:" // Height 1: Multiply loop: Main loop head "fmla v8.4s, v6.4s, v0.s[0]\n" - "ldr q6, [x10, #0x0]\n" + "ldr q17, [x10, #0x0]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" - "ldr q7, [x9, #0x0]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "ldr q6, [x12, #0x10]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "ldr q7, [x11, #0x10]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x10, #0x10]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x9, #0x10]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x12, #0x20]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x11, #0x20]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "ldr q7, [x9, #0x20]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "ldr q6, [x12, #0x30]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "ldr q7, [x11, #0x30]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "ldr q6, [x10, #0x30]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "ldr q7, [x9, #0x30]\n" + "ldr q16, [x9, #0x0]\n" + "fmla v10.4s, v17.4s, v0.s[0]\n" + "ldr q17, [x12, #0x10]\n" + "fmla v11.4s, v16.4s, v0.s[0]\n" + "ldr q16, [x11, #0x10]\n" + "fmla v8.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x10, #0x10]\n" + "fmla v9.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x9, #0x10]\n" + "fmla v10.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x12, #0x20]\n" + "fmla v11.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x11, #0x20]\n" + "fmla v8.4s, v17.4s, v0.s[2]\n" + "ldr q17, [x10, #0x20]\n" + "fmla v9.4s, v16.4s, v0.s[2]\n" + "ldr q16, [x9, #0x20]\n" + "fmla v10.4s, v17.4s, v0.s[2]\n" + "ldr q17, [x12, #0x30]\n" + "fmla v11.4s, v16.4s, v0.s[2]\n" + "ldr q16, [x11, #0x30]\n" + "fmla v8.4s, v17.4s, v0.s[3]\n" + "ldr q17, [x10, #0x30]\n" + "fmla v9.4s, v16.4s, v0.s[3]\n" + "ldr q16, [x9, #0x30]\n" "sub x27, x27, #0x4\n" "cmp x27, #0x8\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" "add x26, x26, #0x10\n" "ldr q0, [x26, #0x0]\n" "add x12, x12, #0x40\n" @@ -272,36 +272,36 @@ void a64_ffhybrid_fp32_mla_6x16 ( "bge 19b\n" "20:" // Height 1: Multiply loop: Single iteration only "fmla v8.4s, v6.4s, v0.s[0]\n" - "ldr q6, [x10, #0x0]\n" + "ldr q17, [x10, #0x0]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" - "ldr q7, [x9, #0x0]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "ldr q6, [x12, #0x10]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "ldr q7, [x11, #0x10]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x10, #0x10]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x9, #0x10]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "ldr q6, [x12, #0x20]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "ldr q7, [x11, #0x20]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "ldr q7, [x9, #0x20]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "ldr q6, [x12, #0x30]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "ldr q7, [x11, #0x30]\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "ldr q6, [x10, #0x30]\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "ldr q7, [x9, #0x30]\n" + "ldr q16, [x9, #0x0]\n" + "fmla v10.4s, v17.4s, v0.s[0]\n" + "ldr q17, [x12, #0x10]\n" + "fmla v11.4s, v16.4s, v0.s[0]\n" + "ldr q16, [x11, #0x10]\n" + "fmla v8.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x10, #0x10]\n" + "fmla v9.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x9, #0x10]\n" + "fmla v10.4s, v17.4s, v0.s[1]\n" + "ldr q17, [x12, #0x20]\n" + "fmla v11.4s, v16.4s, v0.s[1]\n" + "ldr q16, [x11, #0x20]\n" + "fmla v8.4s, v17.4s, v0.s[2]\n" + "ldr q17, [x10, #0x20]\n" + "fmla v9.4s, v16.4s, v0.s[2]\n" + "ldr q16, [x9, #0x20]\n" + "fmla v10.4s, v17.4s, v0.s[2]\n" + "ldr q17, [x12, #0x30]\n" + "fmla v11.4s, v16.4s, v0.s[2]\n" + "ldr q16, [x11, #0x30]\n" + "fmla v8.4s, v17.4s, v0.s[3]\n" + "ldr q17, [x10, #0x30]\n" + "fmla v9.4s, v16.4s, v0.s[3]\n" + "ldr q16, [x9, #0x30]\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" "add x26, x26, #0x10\n" "add x12, x12, #0x40\n" "add x11, x11, #0x40\n" @@ -310,16 +310,16 @@ void a64_ffhybrid_fp32_mla_6x16 ( "21:" // Height 1: Multiply loop: Main loop skip "cbz x27, 23f\n" "22:" // Height 1: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr q6, [x12, #0x0]\n" - "fmla v8.4s, v6.4s, v0.s[0]\n" + "ldr s18, [x26], #0x4\n" + "ldr q16, [x12, #0x0]\n" + "fmla v8.4s, v16.4s, v18.s[0]\n" "sub x27, x27, #0x1\n" - "ldr q7, [x11, #0x0]\n" - "ldr q6, [x10, #0x0]\n" - "fmla v9.4s, v7.4s, v0.s[0]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "ldr q7, [x9, #0x0]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" + "ldr q17, [x11, #0x0]\n" + "ldr q16, [x10, #0x0]\n" + "fmla v9.4s, v17.4s, v18.s[0]\n" + "fmla v10.4s, v16.4s, v18.s[0]\n" + "ldr q16, [x9, #0x0]\n" + "fmla v11.4s, v16.4s, v18.s[0]\n" "add x12, x12, #0x10\n" "add x11, x11, #0x10\n" "add x10, x10, #0x10\n" @@ -332,17 +332,17 @@ void a64_ffhybrid_fp32_mla_6x16 ( "bne 16b\n" "tbz %x[flags], #1, 24f\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" "24:" // Height 1: No activation "cmp x14, #0x10\n" "bge 33f\n" @@ -538,12 +538,12 @@ void a64_ffhybrid_fp32_mla_6x16 ( "50:" // Height 2: String loop "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" "tbz %x[flags], #3, 51f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" + "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n" + "add x20, x20, x21, LSL #3\n" + "ldr x26, [x20, #0x0]\n" + "ldr x25, [x20, #0x8]\n" "cbnz x28, 52f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -551,7 +551,7 @@ void a64_ffhybrid_fp32_mla_6x16 ( "b 52f\n" "51:" // 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" "52:" // Height 2: input setup done "cmp x27, #0x4\n" "blt 55f\n" @@ -564,137 +564,137 @@ void a64_ffhybrid_fp32_mla_6x16 ( "53:" // 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, #0x0]\n" + "ldr q17, [x10, #0x0]\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, [x9, #0x0]\n" + "ldr q16, [x9, #0x0]\n" "cmp x27, #0x8\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" - "ldr q6, [x12, #0x10]\n" + "fmla v10.4s, v17.4s, v0.s[0]\n" + "fmla v14.4s, v17.4s, v1.s[0]\n" + "ldr q17, [x12, #0x10]\n" "add x26, x26, #0x10\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "ldr q7, [x11, #0x10]\n" + "fmla v11.4s, v16.4s, v0.s[0]\n" + "fmla v15.4s, v16.4s, v1.s[0]\n" + "ldr q16, [x11, #0x10]\n" "add x25, x25, #0x10\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v12.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x10, #0x10]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v13.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x9, #0x10]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x12, #0x20]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x11, #0x20]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "fmla v12.4s, v6.4s, v1.s[2]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "fmla v13.4s, v7.4s, v1.s[2]\n" - "ldr q7, [x9, #0x20]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v14.4s, v6.4s, v1.s[2]\n" - "ldr q6, [x12, #0x30]\n" + "fmla v8.4s, v17.4s, v0.s[1]\n" + "fmla v12.4s, v17.4s, v1.s[1]\n" + "ldr q17, [x10, #0x10]\n" + "fmla v9.4s, v16.4s, v0.s[1]\n" + "fmla v13.4s, v16.4s, v1.s[1]\n" + "ldr q16, [x9, #0x10]\n" + "fmla v10.4s, v17.4s, v0.s[1]\n" + "fmla v14.4s, v17.4s, v1.s[1]\n" + "ldr q17, [x12, #0x20]\n" + "fmla v11.4s, v16.4s, v0.s[1]\n" + "fmla v15.4s, v16.4s, v1.s[1]\n" + "ldr q16, [x11, #0x20]\n" + "fmla v8.4s, v17.4s, v0.s[2]\n" + "fmla v12.4s, v17.4s, v1.s[2]\n" + "ldr q17, [x10, #0x20]\n" + "fmla v9.4s, v16.4s, v0.s[2]\n" + "fmla v13.4s, v16.4s, v1.s[2]\n" + "ldr q16, [x9, #0x20]\n" + "fmla v10.4s, v17.4s, v0.s[2]\n" + "fmla v14.4s, v17.4s, v1.s[2]\n" + "ldr q17, [x12, #0x30]\n" "add x12, x12, #0x40\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "ldr q7, [x11, #0x30]\n" + "fmla v11.4s, v16.4s, v0.s[2]\n" + "fmla v15.4s, v16.4s, v1.s[2]\n" + "ldr q16, [x11, #0x30]\n" "add x11, x11, #0x40\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v12.4s, v6.4s, v1.s[3]\n" - "ldr q6, [x10, #0x30]\n" + "fmla v8.4s, v17.4s, v0.s[3]\n" + "fmla v12.4s, v17.4s, v1.s[3]\n" + "ldr q17, [x10, #0x30]\n" "add x10, x10, #0x40\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "ldr q7, [x9, #0x30]\n" + "fmla v9.4s, v16.4s, v0.s[3]\n" + "fmla v13.4s, v16.4s, v1.s[3]\n" + "ldr q16, [x9, #0x30]\n" "add x9, x9, #0x40\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, [x12, #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, [x11, #0x0]\n" "bge 53b\n" "54:" // 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, #0x0]\n" + "ldr q17, [x10, #0x0]\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, [x9, #0x0]\n" + "ldr q16, [x9, #0x0]\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, [x12, #0x10]\n" + "fmla v10.4s, v17.4s, v0.s[0]\n" + "fmla v14.4s, v17.4s, v1.s[0]\n" + "ldr q17, [x12, #0x10]\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, [x11, #0x10]\n" - "fmla v8.4s, v6.4s, v0.s[1]\n" - "fmla v12.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x10, #0x10]\n" - "fmla v9.4s, v7.4s, v0.s[1]\n" - "fmla v13.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x9, #0x10]\n" - "fmla v10.4s, v6.4s, v0.s[1]\n" - "fmla v14.4s, v6.4s, v1.s[1]\n" - "ldr q6, [x12, #0x20]\n" - "fmla v11.4s, v7.4s, v0.s[1]\n" - "fmla v15.4s, v7.4s, v1.s[1]\n" - "ldr q7, [x11, #0x20]\n" - "fmla v8.4s, v6.4s, v0.s[2]\n" - "fmla v12.4s, v6.4s, v1.s[2]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.4s, v7.4s, v0.s[2]\n" - "fmla v13.4s, v7.4s, v1.s[2]\n" - "ldr q7, [x9, #0x20]\n" - "fmla v10.4s, v6.4s, v0.s[2]\n" - "fmla v14.4s, v6.4s, v1.s[2]\n" - "ldr q6, [x12, #0x30]\n" + "fmla v11.4s, v16.4s, v0.s[0]\n" + "fmla v15.4s, v16.4s, v1.s[0]\n" + "ldr q16, [x11, #0x10]\n" + "fmla v8.4s, v17.4s, v0.s[1]\n" + "fmla v12.4s, v17.4s, v1.s[1]\n" + "ldr q17, [x10, #0x10]\n" + "fmla v9.4s, v16.4s, v0.s[1]\n" + "fmla v13.4s, v16.4s, v1.s[1]\n" + "ldr q16, [x9, #0x10]\n" + "fmla v10.4s, v17.4s, v0.s[1]\n" + "fmla v14.4s, v17.4s, v1.s[1]\n" + "ldr q17, [x12, #0x20]\n" + "fmla v11.4s, v16.4s, v0.s[1]\n" + "fmla v15.4s, v16.4s, v1.s[1]\n" + "ldr q16, [x11, #0x20]\n" + "fmla v8.4s, v17.4s, v0.s[2]\n" + "fmla v12.4s, v17.4s, v1.s[2]\n" + "ldr q17, [x10, #0x20]\n" + "fmla v9.4s, v16.4s, v0.s[2]\n" + "fmla v13.4s, v16.4s, v1.s[2]\n" + "ldr q16, [x9, #0x20]\n" + "fmla v10.4s, v17.4s, v0.s[2]\n" + "fmla v14.4s, v17.4s, v1.s[2]\n" + "ldr q17, [x12, #0x30]\n" "add x12, x12, #0x40\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "ldr q7, [x11, #0x30]\n" + "fmla v11.4s, v16.4s, v0.s[2]\n" + "fmla v15.4s, v16.4s, v1.s[2]\n" + "ldr q16, [x11, #0x30]\n" "add x11, x11, #0x40\n" - "fmla v8.4s, v6.4s, v0.s[3]\n" - "fmla v12.4s, v6.4s, v1.s[3]\n" - "ldr q6, [x10, #0x30]\n" + "fmla v8.4s, v17.4s, v0.s[3]\n" + "fmla v12.4s, v17.4s, v1.s[3]\n" + "ldr q17, [x10, #0x30]\n" "add x10, x10, #0x40\n" - "fmla v9.4s, v7.4s, v0.s[3]\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "ldr q7, [x9, #0x30]\n" + "fmla v9.4s, v16.4s, v0.s[3]\n" + "fmla v13.4s, v16.4s, v1.s[3]\n" + "ldr q16, [x9, #0x30]\n" "add x9, x9, #0x40\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" "55:" // Height 2: Multiply loop: Main loop skip "cbz x27, 57f\n" "56:" // Height 2: Multiply loop: Odd block loop - "ldr s0, [x26], #0x4\n" - "ldr s1, [x25], #0x4\n" + "ldr s19, [x26], #0x4\n" + "ldr s18, [x25], #0x4\n" "sub x27, x27, #0x1\n" - "ldr q6, [x12, #0x0]\n" - "ldr q7, [x11, #0x0]\n" - "fmla v8.4s, v6.4s, v0.s[0]\n" - "fmla v12.4s, v6.4s, v1.s[0]\n" - "ldr q6, [x10, #0x0]\n" - "fmla v9.4s, v7.4s, v0.s[0]\n" - "fmla v13.4s, v7.4s, v1.s[0]\n" - "ldr q7, [x9, #0x0]\n" - "fmla v10.4s, v6.4s, v0.s[0]\n" - "fmla v14.4s, v6.4s, v1.s[0]\n" + "ldr q17, [x12, #0x0]\n" + "ldr q16, [x11, #0x0]\n" + "fmla v8.4s, v17.4s, v19.s[0]\n" + "fmla v12.4s, v17.4s, v18.s[0]\n" + "ldr q17, [x10, #0x0]\n" + "fmla v9.4s, v16.4s, v19.s[0]\n" + "fmla v13.4s, v16.4s, v18.s[0]\n" + "ldr q16, [x9, #0x0]\n" + "fmla v10.4s, v17.4s, v19.s[0]\n" + "fmla v14.4s, v17.4s, v18.s[0]\n" "add x12, x12, #0x10\n" "add x11, x11, #0x10\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" "add x10, x10, #0x10\n" "add x9, x9, #0x10\n" "cbnz x27, 56b\n" @@ -707,25 +707,25 @@ void a64_ffhybrid_fp32_mla_6x16 ( "add x25, x13, x20, LSL #2\n" "tbz %x[flags], #1, 58f\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" "58:" // Height 2: No activation "cmp x14, #0x10\n" "bge 67f\n" @@ -970,13 +970,13 @@ void a64_ffhybrid_fp32_mla_6x16 ( "84:" // 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, 85f\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, 86f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -985,8 +985,8 @@ void a64_ffhybrid_fp32_mla_6x16 ( "b 86f\n" "85:" // 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" "86:" // Height 3: input setup done "cmp x27, #0x4\n" "blt 89f\n" @@ -1003,75 +1003,75 @@ void a64_ffhybrid_fp32_mla_6x16 ( "sub x27, x27, #0x4\n" "cmp x27, #0x8\n" "fmla v16.4s, v6.4s, v2.s[0]\n" - "ldr q6, [x10, #0x0]\n" + "ldr q21, [x10, #0x0]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "add x26, x26, #0x10\n" "fmla v13.4s, v7.4s, v1.s[0]\n" "fmla v17.4s, v7.4s, v2.s[0]\n" - "ldr q7, [x9, #0x0]\n" + "ldr q20, [x9, #0x0]\n" "add x25, x25, #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" "add x24, x24, #0x10\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "ldr q6, [x12, #0x10]\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" - "ldr q7, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" + "fmla v18.4s, v21.4s, v2.s[0]\n" + "ldr q21, [x12, #0x10]\n" + "fmla v11.4s, v20.4s, v0.s[0]\n" + "fmla v15.4s, v20.4s, v1.s[0]\n" + "fmla v19.4s, v20.4s, v2.s[0]\n" + "ldr q20, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" + "fmla v11.4s, v20.4s, v0.s[2]\n" "add x12, x12, #0x40\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "fmla v19.4s, v7.4s, v2.s[2]\n" - "ldr q7, [x11, #0x30]\n" + "fmla v15.4s, v20.4s, v1.s[2]\n" + "fmla v19.4s, v20.4s, v2.s[2]\n" + "ldr q20, [x11, #0x30]\n" "add x11, x11, #0x40\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, #0x30]\n" - "fmla v9.4s, v7.4s, v0.s[3]\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, #0x30]\n" + "fmla v9.4s, v20.4s, v0.s[3]\n" "add x10, x10, #0x40\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "fmla v17.4s, v7.4s, v2.s[3]\n" - "ldr q7, [x9, #0x30]\n" + "fmla v13.4s, v20.4s, v1.s[3]\n" + "fmla v17.4s, v20.4s, v2.s[3]\n" + "ldr q20, [x9, #0x30]\n" "add x9, x9, #0x40\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, [x12, #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, [x11, #0x0]\n" "bge 87b\n" @@ -1081,98 +1081,98 @@ void a64_ffhybrid_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, #0x0]\n" + "ldr q21, [x10, #0x0]\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, [x9, #0x0]\n" + "ldr q20, [x9, #0x0]\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 v18.4s, v6.4s, v2.s[0]\n" - "ldr q6, [x12, #0x10]\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" - "ldr q7, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" + "fmla v10.4s, v21.4s, v0.s[0]\n" + "fmla v14.4s, v21.4s, v1.s[0]\n" + "fmla v18.4s, v21.4s, v2.s[0]\n" + "ldr q21, [x12, #0x10]\n" + "fmla v11.4s, v20.4s, v0.s[0]\n" + "fmla v15.4s, v20.4s, v1.s[0]\n" + "fmla v19.4s, v20.4s, v2.s[0]\n" + "ldr q20, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" + "fmla v11.4s, v20.4s, v0.s[2]\n" "add x12, x12, #0x40\n" - "fmla v15.4s, v7.4s, v1.s[2]\n" - "fmla v19.4s, v7.4s, v2.s[2]\n" - "ldr q7, [x11, #0x30]\n" + "fmla v15.4s, v20.4s, v1.s[2]\n" + "fmla v19.4s, v20.4s, v2.s[2]\n" + "ldr q20, [x11, #0x30]\n" "add x11, x11, #0x40\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, #0x30]\n" - "fmla v9.4s, v7.4s, v0.s[3]\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, #0x30]\n" + "fmla v9.4s, v20.4s, v0.s[3]\n" "add x10, x10, #0x40\n" - "fmla v13.4s, v7.4s, v1.s[3]\n" - "fmla v17.4s, v7.4s, v2.s[3]\n" - "ldr q7, [x9, #0x30]\n" + "fmla v13.4s, v20.4s, v1.s[3]\n" + "fmla v17.4s, v20.4s, v2.s[3]\n" + "ldr q20, [x9, #0x30]\n" "add x9, x9, #0x40\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" "89:" // Height 3: Multiply loop: Main loop skip "cbz x27, 91f\n" "90:" // 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, [x12, #0x0]\n" - "fmla v8.4s, v6.4s, v0.s[0]\n" - "fmla v12.4s, v6.4s, v1.s[0]\n" - "ldr q7, [x11, #0x0]\n" - "fmla v16.4s, v6.4s, v2.s[0]\n" - "ldr q6, [x10, #0x0]\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, [x9, #0x0]\n" + "ldr s22, [x24], #0x4\n" + "ldr q21, [x12, #0x0]\n" + "fmla v8.4s, v21.4s, v24.s[0]\n" + "fmla v12.4s, v21.4s, v23.s[0]\n" + "ldr q20, [x11, #0x0]\n" + "fmla v16.4s, v21.4s, v22.s[0]\n" + "ldr q21, [x10, #0x0]\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, [x9, #0x0]\n" "add x12, x12, #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, v24.s[0]\n" + "fmla v14.4s, v21.4s, v23.s[0]\n" "add x11, x11, #0x10\n" "add x10, x10, #0x10\n" - "fmla v18.4s, v6.4s, v2.s[0]\n" - "fmla v11.4s, v7.4s, v0.s[0]\n" + "fmla v18.4s, v21.4s, v22.s[0]\n" + "fmla v11.4s, v20.4s, v24.s[0]\n" "add x9, x9, #0x10\n" - "fmla v15.4s, v7.4s, v1.s[0]\n" - "fmla v19.4s, v7.4s, v2.s[0]\n" + "fmla v15.4s, v20.4s, v23.s[0]\n" + "fmla v19.4s, v20.4s, v22.s[0]\n" "cbnz x27, 90b\n" "91:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1184,33 +1184,33 @@ void a64_ffhybrid_fp32_mla_6x16 ( "add x24, x25, x20, LSL #2\n" "tbz %x[flags], #1, 92f\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" "92:" // Height 3: No activation "cmp x14, #0x10\n" "bge 101f\n" @@ -1504,14 +1504,14 @@ void a64_ffhybrid_fp32_mla_6x16 ( "118:" // 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, 119f\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, 120f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -1521,9 +1521,9 @@ void a64_ffhybrid_fp32_mla_6x16 ( "b 120f\n" "119:" // 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" "120:" // Height 4: input setup done "cmp x27, #0x4\n" "blt 123f\n" @@ -1542,7 +1542,7 @@ void a64_ffhybrid_fp32_mla_6x16 ( "cmp x27, #0x8\n" "fmla v16.4s, v6.4s, v2.s[0]\n" "fmla v20.4s, v6.4s, v3.s[0]\n" - "ldr q6, [x10, #0x0]\n" + "ldr q25, [x10, #0x0]\n" "add x26, x26, #0x10\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "fmla v13.4s, v7.4s, v1.s[0]\n" @@ -1550,84 +1550,84 @@ void a64_ffhybrid_fp32_mla_6x16 ( "add x24, x24, #0x10\n" "fmla v17.4s, v7.4s, v2.s[0]\n" "fmla v21.4s, v7.4s, v3.s[0]\n" - "ldr q7, [x9, #0x0]\n" + "ldr q24, [x9, #0x0]\n" "add x23, x23, #0x10\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" - "ldr q6, [x12, #0x10]\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, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" + "fmla v10.4s, v25.4s, v0.s[0]\n" + "fmla v14.4s, v25.4s, v1.s[0]\n" + "fmla v18.4s, v25.4s, v2.s[0]\n" + "fmla v22.4s, v25.4s, v3.s[0]\n" + "ldr q25, [x12, #0x10]\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, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" "add x12, x12, #0x40\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, [x11, #0x30]\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, [x11, #0x30]\n" "add x11, x11, #0x40\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, #0x30]\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, #0x30]\n" "add x10, x10, #0x40\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, [x9, #0x30]\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, [x9, #0x30]\n" "add x9, x9, #0x40\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, [x12, #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, [x11, #0x0]\n" "bge 121b\n" @@ -1638,7 +1638,7 @@ void a64_ffhybrid_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, #0x0]\n" + "ldr q25, [x10, #0x0]\n" "add x25, x25, #0x10\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "fmla v13.4s, v7.4s, v1.s[0]\n" @@ -1646,111 +1646,111 @@ void a64_ffhybrid_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, [x9, #0x0]\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" - "ldr q6, [x12, #0x10]\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, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" + "ldr q24, [x9, #0x0]\n" + "fmla v10.4s, v25.4s, v0.s[0]\n" + "fmla v14.4s, v25.4s, v1.s[0]\n" + "fmla v18.4s, v25.4s, v2.s[0]\n" + "fmla v22.4s, v25.4s, v3.s[0]\n" + "ldr q25, [x12, #0x10]\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, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" "add x12, x12, #0x40\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, [x11, #0x30]\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, [x11, #0x30]\n" "add x11, x11, #0x40\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, #0x30]\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, #0x30]\n" "add x10, x10, #0x40\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, [x9, #0x30]\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, [x9, #0x30]\n" "add x9, x9, #0x40\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" "123:" // Height 4: Multiply loop: Main loop skip "cbz x27, 125f\n" "124:" // 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, [x12, #0x0]\n" - "ldr q7, [x11, #0x0]\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, #0x0]\n" + "ldr s27, [x24], #0x4\n" + "ldr s26, [x23], #0x4\n" + "ldr q25, [x12, #0x0]\n" + "ldr q24, [x11, #0x0]\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, #0x0]\n" "add x12, x12, #0x10\n" - "fmla v9.4s, v7.4s, v0.s[0]\n" - "fmla v13.4s, v7.4s, v1.s[0]\n" + "fmla v9.4s, v24.4s, v29.s[0]\n" + "fmla v13.4s, v24.4s, v28.s[0]\n" "add x11, x11, #0x10\n" "add x10, x10, #0x10\n" - "fmla v17.4s, v7.4s, v2.s[0]\n" - "fmla v21.4s, v7.4s, v3.s[0]\n" - "ldr q7, [x9, #0x0]\n" + "fmla v17.4s, v24.4s, v27.s[0]\n" + "fmla v21.4s, v24.4s, v26.s[0]\n" + "ldr q24, [x9, #0x0]\n" "add x9, x9, #0x10\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, 124b\n" "125:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1763,41 +1763,41 @@ void a64_ffhybrid_fp32_mla_6x16 ( "add x23, x24, x20, LSL #2\n" "tbz %x[flags], #1, 126f\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" "126:" // Height 4: No activation "cmp x14, #0x10\n" "bge 135f\n" @@ -2140,15 +2140,15 @@ void a64_ffhybrid_fp32_mla_6x16 ( "152:" // 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, 153f\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, 154f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -2159,10 +2159,10 @@ void a64_ffhybrid_fp32_mla_6x16 ( "b 154f\n" "153:" // 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" "154:" // Height 5: input setup done "cmp x27, #0x4\n" "blt 157f\n" @@ -2185,7 +2185,7 @@ void a64_ffhybrid_fp32_mla_6x16 ( "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" "fmla v24.4s, v6.4s, v4.s[0]\n" - "ldr q6, [x10, #0x0]\n" + "ldr q29, [x10, #0x0]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "add x24, x24, #0x10\n" "fmla v13.4s, v7.4s, v1.s[0]\n" @@ -2194,98 +2194,98 @@ void a64_ffhybrid_fp32_mla_6x16 ( "add x22, x22, #0x10\n" "fmla v21.4s, v7.4s, v3.s[0]\n" "fmla v25.4s, v7.4s, v4.s[0]\n" - "ldr q7, [x9, #0x0]\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" - "ldr q6, [x12, #0x10]\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, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" + "ldr q28, [x9, #0x0]\n" + "fmla v10.4s, v29.4s, v0.s[0]\n" + "fmla v14.4s, v29.4s, v1.s[0]\n" + "fmla v18.4s, v29.4s, v2.s[0]\n" + "fmla v22.4s, v29.4s, v3.s[0]\n" + "fmla v26.4s, v29.4s, v4.s[0]\n" + "ldr q29, [x12, #0x10]\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, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" + "fmla v11.4s, v28.4s, v0.s[2]\n" "add x12, x12, #0x40\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, [x11, #0x30]\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, [x11, #0x30]\n" "add x11, x11, #0x40\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, #0x30]\n" - "fmla v9.4s, v7.4s, v0.s[3]\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, #0x30]\n" + "fmla v9.4s, v28.4s, v0.s[3]\n" "add x10, x10, #0x40\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, [x9, #0x30]\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, [x9, #0x30]\n" "add x9, x9, #0x40\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, [x12, #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, [x11, #0x0]\n" "bge 155b\n" @@ -2299,7 +2299,7 @@ void a64_ffhybrid_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, #0x0]\n" + "ldr q29, [x10, #0x0]\n" "fmla v9.4s, v7.4s, v0.s[0]\n" "add x23, x23, #0x10\n" "fmla v13.4s, v7.4s, v1.s[0]\n" @@ -2307,130 +2307,130 @@ void a64_ffhybrid_fp32_mla_6x16 ( "add x22, x22, #0x10\n" "fmla v21.4s, v7.4s, v3.s[0]\n" "fmla v25.4s, v7.4s, v4.s[0]\n" - "ldr q7, [x9, #0x0]\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" - "ldr q6, [x12, #0x10]\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, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" - "fmla v11.4s, v7.4s, v0.s[2]\n" + "ldr q28, [x9, #0x0]\n" + "fmla v10.4s, v29.4s, v0.s[0]\n" + "fmla v14.4s, v29.4s, v1.s[0]\n" + "fmla v18.4s, v29.4s, v2.s[0]\n" + "fmla v22.4s, v29.4s, v3.s[0]\n" + "fmla v26.4s, v29.4s, v4.s[0]\n" + "ldr q29, [x12, #0x10]\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, [x11, #0x10]\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, #0x10]\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, [x9, #0x10]\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, [x12, #0x20]\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, [x11, #0x20]\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, #0x20]\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, [x9, #0x20]\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, [x12, #0x30]\n" + "fmla v11.4s, v28.4s, v0.s[2]\n" "add x12, x12, #0x40\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, [x11, #0x30]\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, [x11, #0x30]\n" "add x11, x11, #0x40\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, #0x30]\n" - "fmla v9.4s, v7.4s, v0.s[3]\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, #0x30]\n" + "fmla v9.4s, v28.4s, v0.s[3]\n" "add x10, x10, #0x40\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, [x9, #0x30]\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, [x9, #0x30]\n" "add x9, x9, #0x40\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" "157:" // Height 5: Multiply loop: Main loop skip "cbz x27, 159f\n" "158:" // 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, [x12, #0x0]\n" - "fmla v8.4s, v6.4s, v0.s[0]\n" - "fmla v12.4s, v6.4s, v1.s[0]\n" - "ldr q7, [x11, #0x0]\n" - "fmla v16.4s, v6.4s, v2.s[0]\n" - "fmla v20.4s, v6.4s, v3.s[0]\n" + "ldr s0, [x24], #0x4\n" + "ldr s31, [x23], #0x4\n" + "ldr s30, [x22], #0x4\n" + "ldr q29, [x12, #0x0]\n" + "fmla v8.4s, v29.4s, v2.s[0]\n" + "fmla v12.4s, v29.4s, v1.s[0]\n" + "ldr q28, [x11, #0x0]\n" + "fmla v16.4s, v29.4s, v0.s[0]\n" + "fmla v20.4s, v29.4s, v31.s[0]\n" "add x12, x12, #0x10\n" - "fmla v24.4s, v6.4s, v4.s[0]\n" - "ldr q6, [x10, #0x0]\n" - "fmla v9.4s, v7.4s, v0.s[0]\n" + "fmla v24.4s, v29.4s, v30.s[0]\n" + "ldr q29, [x10, #0x0]\n" + "fmla v9.4s, v28.4s, v2.s[0]\n" "add x11, x11, #0x10\n" - "fmla v13.4s, v7.4s, v1.s[0]\n" - "fmla v17.4s, v7.4s, v2.s[0]\n" + "fmla v13.4s, v28.4s, v1.s[0]\n" + "fmla v17.4s, v28.4s, v0.s[0]\n" "add x10, x10, #0x10\n" - "fmla v21.4s, v7.4s, v3.s[0]\n" - "fmla v25.4s, v7.4s, v4.s[0]\n" - "ldr q7, [x9, #0x0]\n" + "fmla v21.4s, v28.4s, v31.s[0]\n" + "fmla v25.4s, v28.4s, v30.s[0]\n" + "ldr q28, [x9, #0x0]\n" "add x9, x9, #0x10\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, 158b\n" "159:" // Height 5: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -2444,49 +2444,49 @@ void a64_ffhybrid_fp32_mla_6x16 ( "add x22, x23, x20, LSL #2\n" "tbz %x[flags], #1, 160f\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" "160:" // Height 5: No activation "cmp x14, #0x10\n" "bge 169f\n" @@ -2881,16 +2881,16 @@ void a64_ffhybrid_fp32_mla_6x16 ( "186:" // 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, 187f\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, 188f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #2\n" @@ -2902,11 +2902,11 @@ void a64_ffhybrid_fp32_mla_6x16 ( "b 188f\n" "187:" // 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" "188:" // Height 6: input setup done "cmp x27, #0x4\n" "blt 191f\n" @@ -3177,45 +3177,45 @@ void a64_ffhybrid_fp32_mla_6x16 ( "191:" // Height 6: Multiply loop: Main loop skip "cbz x27, 193f\n" "192:" // 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, [x12, #0x0]\n" - "ldr q7, [x11, #0x0]\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 s5, [x24], #0x4\n" + "ldr s4, [x23], #0x4\n" + "ldr s3, [x22], #0x4\n" + "ldr s2, [x21], #0x4\n" + "ldr q1, [x12, #0x0]\n" + "ldr q0, [x11, #0x0]\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" "add x12, x12, #0x10\n" "add x11, x11, #0x10\n" - "fmla v24.4s, v6.4s, v4.s[0]\n" - "fmla v28.4s, v6.4s, v5.s[0]\n" - "ldr q6, [x10, #0x0]\n" + "fmla v24.4s, v1.4s, v3.s[0]\n" + "fmla v28.4s, v1.4s, v2.s[0]\n" + "ldr q1, [x10, #0x0]\n" "add x10, x10, #0x10\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, [x9, #0x0]\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, [x9, #0x0]\n" "add x9, x9, #0x10\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, 192b\n" "193:" // Height 6: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" |