From 74921eee924625426429044decefe3673561b174 Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 12 Apr 2023 17:43:17 +0100 Subject: Update CPU kernel implementations and guard directives Resolves COMPMID-6023 Change-Id: I868975d14c4f98af6716726feda22405a6a4c891 Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9686 Tested-by: Arm Jenkins Reviewed-by: Viet-Hoa Do Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../kernels/a64_hybrid_fp16_mla_6x32/generic.cpp | 3349 ++++++++++---------- 1 file changed, 1674 insertions(+), 1675 deletions(-) (limited to 'src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp') diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp index 335308751f..8e5f600c83 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hybrid_fp16_mla_6x32/generic.cpp @@ -244,11 +244,11 @@ void a64_hybrid_fp16_mla_6x32 ( "23:" // 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, 24f\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, 25f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -265,69 +265,69 @@ void a64_hybrid_fp16_mla_6x32 ( "blt 27f\n" "26:" // Height 1: Multiply loop: Main loop head "fmla v8.8h, v6.8h, v0.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "fmla v9.8h, v7.8h, v0.h[0]\n" - "ldr q7, [x10, #0x30]\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.8h, v6.8h, v0.h[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.8h, v7.8h, v0.h[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.8h, v6.8h, v0.h[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.8h, v6.8h, v0.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.8h, v7.8h, v0.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.8h, v6.8h, v0.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.8h, v7.8h, v0.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.8h, v6.8h, v0.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.8h, v7.8h, v0.h[3]\n" - "ldr q7, [x10, #0xf0]\n" - "fmla v10.8h, v6.8h, v0.h[3]\n" - "ldr q6, [x10, #0x100]\n" - "fmla v11.8h, v7.8h, v0.h[3]\n" - "ldr q7, [x10, #0x110]\n" - "fmla v8.8h, v6.8h, v0.h[4]\n" - "ldr q6, [x10, #0x120]\n" - "fmla v9.8h, v7.8h, v0.h[4]\n" - "ldr q7, [x10, #0x130]\n" - "fmla v10.8h, v6.8h, v0.h[4]\n" - "ldr q6, [x10, #0x140]\n" - "fmla v11.8h, v7.8h, v0.h[4]\n" - "ldr q7, [x10, #0x150]\n" - "fmla v8.8h, v6.8h, v0.h[5]\n" - "ldr q6, [x10, #0x160]\n" - "fmla v9.8h, v7.8h, v0.h[5]\n" - "ldr q7, [x10, #0x170]\n" - "fmla v10.8h, v6.8h, v0.h[5]\n" - "ldr q6, [x10, #0x180]\n" - "fmla v11.8h, v7.8h, v0.h[5]\n" - "ldr q7, [x10, #0x190]\n" - "fmla v8.8h, v6.8h, v0.h[6]\n" - "ldr q6, [x10, #0x1a0]\n" - "fmla v9.8h, v7.8h, v0.h[6]\n" - "ldr q7, [x10, #0x1b0]\n" - "fmla v10.8h, v6.8h, v0.h[6]\n" - "ldr q6, [x10, #0x1c0]\n" - "fmla v11.8h, v7.8h, v0.h[6]\n" - "ldr q7, [x10, #0x1d0]\n" - "fmla v8.8h, v6.8h, v0.h[7]\n" - "ldr q6, [x10, #0x1e0]\n" - "fmla v9.8h, v7.8h, v0.h[7]\n" - "ldr q7, [x10, #0x1f0]\n" + "ldr q16, [x10, #0x30]\n" + "fmla v10.8h, v17.8h, v0.h[0]\n" + "ldr q17, [x10, #0x40]\n" + "fmla v11.8h, v16.8h, v0.h[0]\n" + "ldr q16, [x10, #0x50]\n" + "fmla v8.8h, v17.8h, v0.h[1]\n" + "ldr q17, [x10, #0x60]\n" + "fmla v9.8h, v16.8h, v0.h[1]\n" + "ldr q16, [x10, #0x70]\n" + "fmla v10.8h, v17.8h, v0.h[1]\n" + "ldr q17, [x10, #0x80]\n" + "fmla v11.8h, v16.8h, v0.h[1]\n" + "ldr q16, [x10, #0x90]\n" + "fmla v8.8h, v17.8h, v0.h[2]\n" + "ldr q17, [x10, #0xa0]\n" + "fmla v9.8h, v16.8h, v0.h[2]\n" + "ldr q16, [x10, #0xb0]\n" + "fmla v10.8h, v17.8h, v0.h[2]\n" + "ldr q17, [x10, #0xc0]\n" + "fmla v11.8h, v16.8h, v0.h[2]\n" + "ldr q16, [x10, #0xd0]\n" + "fmla v8.8h, v17.8h, v0.h[3]\n" + "ldr q17, [x10, #0xe0]\n" + "fmla v9.8h, v16.8h, v0.h[3]\n" + "ldr q16, [x10, #0xf0]\n" + "fmla v10.8h, v17.8h, v0.h[3]\n" + "ldr q17, [x10, #0x100]\n" + "fmla v11.8h, v16.8h, v0.h[3]\n" + "ldr q16, [x10, #0x110]\n" + "fmla v8.8h, v17.8h, v0.h[4]\n" + "ldr q17, [x10, #0x120]\n" + "fmla v9.8h, v16.8h, v0.h[4]\n" + "ldr q16, [x10, #0x130]\n" + "fmla v10.8h, v17.8h, v0.h[4]\n" + "ldr q17, [x10, #0x140]\n" + "fmla v11.8h, v16.8h, v0.h[4]\n" + "ldr q16, [x10, #0x150]\n" + "fmla v8.8h, v17.8h, v0.h[5]\n" + "ldr q17, [x10, #0x160]\n" + "fmla v9.8h, v16.8h, v0.h[5]\n" + "ldr q16, [x10, #0x170]\n" + "fmla v10.8h, v17.8h, v0.h[5]\n" + "ldr q17, [x10, #0x180]\n" + "fmla v11.8h, v16.8h, v0.h[5]\n" + "ldr q16, [x10, #0x190]\n" + "fmla v8.8h, v17.8h, v0.h[6]\n" + "ldr q17, [x10, #0x1a0]\n" + "fmla v9.8h, v16.8h, v0.h[6]\n" + "ldr q16, [x10, #0x1b0]\n" + "fmla v10.8h, v17.8h, v0.h[6]\n" + "ldr q17, [x10, #0x1c0]\n" + "fmla v11.8h, v16.8h, v0.h[6]\n" + "ldr q16, [x10, #0x1d0]\n" + "fmla v8.8h, v17.8h, v0.h[7]\n" + "ldr q17, [x10, #0x1e0]\n" + "fmla v9.8h, v16.8h, v0.h[7]\n" + "ldr q16, [x10, #0x1f0]\n" "sub x27, x27, #0x8\n" "add x26, x26, #0x10\n" - "fmla v10.8h, v6.8h, v0.h[7]\n" - "fmla v11.8h, v7.8h, v0.h[7]\n" + "fmla v10.8h, v17.8h, v0.h[7]\n" + "fmla v11.8h, v16.8h, v0.h[7]\n" "ldr q0, [x26, #0x0]\n" "cmp x27, #0x10\n" "add x10, x10, #0x200\n" @@ -337,84 +337,84 @@ void a64_hybrid_fp16_mla_6x32 ( "bge 26b\n" "27:" // Height 1: Multiply loop: Single iteration only "fmla v8.8h, v6.8h, v0.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "fmla v9.8h, v7.8h, v0.h[0]\n" - "ldr q7, [x10, #0x30]\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.8h, v6.8h, v0.h[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.8h, v7.8h, v0.h[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.8h, v6.8h, v0.h[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.8h, v6.8h, v0.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.8h, v7.8h, v0.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.8h, v6.8h, v0.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.8h, v7.8h, v0.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.8h, v6.8h, v0.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.8h, v7.8h, v0.h[3]\n" - "ldr q7, [x10, #0xf0]\n" - "fmla v10.8h, v6.8h, v0.h[3]\n" - "ldr q6, [x10, #0x100]\n" - "fmla v11.8h, v7.8h, v0.h[3]\n" - "ldr q7, [x10, #0x110]\n" - "fmla v8.8h, v6.8h, v0.h[4]\n" - "ldr q6, [x10, #0x120]\n" - "fmla v9.8h, v7.8h, v0.h[4]\n" - "ldr q7, [x10, #0x130]\n" - "fmla v10.8h, v6.8h, v0.h[4]\n" - "ldr q6, [x10, #0x140]\n" - "fmla v11.8h, v7.8h, v0.h[4]\n" - "ldr q7, [x10, #0x150]\n" - "fmla v8.8h, v6.8h, v0.h[5]\n" - "ldr q6, [x10, #0x160]\n" - "fmla v9.8h, v7.8h, v0.h[5]\n" - "ldr q7, [x10, #0x170]\n" - "fmla v10.8h, v6.8h, v0.h[5]\n" - "ldr q6, [x10, #0x180]\n" - "fmla v11.8h, v7.8h, v0.h[5]\n" - "ldr q7, [x10, #0x190]\n" - "fmla v8.8h, v6.8h, v0.h[6]\n" - "ldr q6, [x10, #0x1a0]\n" - "fmla v9.8h, v7.8h, v0.h[6]\n" - "ldr q7, [x10, #0x1b0]\n" - "fmla v10.8h, v6.8h, v0.h[6]\n" - "ldr q6, [x10, #0x1c0]\n" - "fmla v11.8h, v7.8h, v0.h[6]\n" - "ldr q7, [x10, #0x1d0]\n" - "fmla v8.8h, v6.8h, v0.h[7]\n" - "ldr q6, [x10, #0x1e0]\n" - "fmla v9.8h, v7.8h, v0.h[7]\n" - "ldr q7, [x10, #0x1f0]\n" + "ldr q16, [x10, #0x30]\n" + "fmla v10.8h, v17.8h, v0.h[0]\n" + "ldr q17, [x10, #0x40]\n" + "fmla v11.8h, v16.8h, v0.h[0]\n" + "ldr q16, [x10, #0x50]\n" + "fmla v8.8h, v17.8h, v0.h[1]\n" + "ldr q17, [x10, #0x60]\n" + "fmla v9.8h, v16.8h, v0.h[1]\n" + "ldr q16, [x10, #0x70]\n" + "fmla v10.8h, v17.8h, v0.h[1]\n" + "ldr q17, [x10, #0x80]\n" + "fmla v11.8h, v16.8h, v0.h[1]\n" + "ldr q16, [x10, #0x90]\n" + "fmla v8.8h, v17.8h, v0.h[2]\n" + "ldr q17, [x10, #0xa0]\n" + "fmla v9.8h, v16.8h, v0.h[2]\n" + "ldr q16, [x10, #0xb0]\n" + "fmla v10.8h, v17.8h, v0.h[2]\n" + "ldr q17, [x10, #0xc0]\n" + "fmla v11.8h, v16.8h, v0.h[2]\n" + "ldr q16, [x10, #0xd0]\n" + "fmla v8.8h, v17.8h, v0.h[3]\n" + "ldr q17, [x10, #0xe0]\n" + "fmla v9.8h, v16.8h, v0.h[3]\n" + "ldr q16, [x10, #0xf0]\n" + "fmla v10.8h, v17.8h, v0.h[3]\n" + "ldr q17, [x10, #0x100]\n" + "fmla v11.8h, v16.8h, v0.h[3]\n" + "ldr q16, [x10, #0x110]\n" + "fmla v8.8h, v17.8h, v0.h[4]\n" + "ldr q17, [x10, #0x120]\n" + "fmla v9.8h, v16.8h, v0.h[4]\n" + "ldr q16, [x10, #0x130]\n" + "fmla v10.8h, v17.8h, v0.h[4]\n" + "ldr q17, [x10, #0x140]\n" + "fmla v11.8h, v16.8h, v0.h[4]\n" + "ldr q16, [x10, #0x150]\n" + "fmla v8.8h, v17.8h, v0.h[5]\n" + "ldr q17, [x10, #0x160]\n" + "fmla v9.8h, v16.8h, v0.h[5]\n" + "ldr q16, [x10, #0x170]\n" + "fmla v10.8h, v17.8h, v0.h[5]\n" + "ldr q17, [x10, #0x180]\n" + "fmla v11.8h, v16.8h, v0.h[5]\n" + "ldr q16, [x10, #0x190]\n" + "fmla v8.8h, v17.8h, v0.h[6]\n" + "ldr q17, [x10, #0x1a0]\n" + "fmla v9.8h, v16.8h, v0.h[6]\n" + "ldr q16, [x10, #0x1b0]\n" + "fmla v10.8h, v17.8h, v0.h[6]\n" + "ldr q17, [x10, #0x1c0]\n" + "fmla v11.8h, v16.8h, v0.h[6]\n" + "ldr q16, [x10, #0x1d0]\n" + "fmla v8.8h, v17.8h, v0.h[7]\n" + "ldr q17, [x10, #0x1e0]\n" + "fmla v9.8h, v16.8h, v0.h[7]\n" + "ldr q16, [x10, #0x1f0]\n" "add x26, x26, #0x10\n" "sub x27, x27, #0x8\n" - "fmla v10.8h, v6.8h, v0.h[7]\n" - "fmla v11.8h, v7.8h, v0.h[7]\n" + "fmla v10.8h, v17.8h, v0.h[7]\n" + "fmla v11.8h, v16.8h, v0.h[7]\n" "prfm pldl1keep, [x26, #0x80]\n" "add x10, x10, #0x200\n" "28:" // Height 1: Multiply loop: Main loop skip "cbz x27, 30f\n" "29:" // Height 1: Multiply loop: Odd block loop "ldr h0, [x26], #0x2\n" - "ldr q6, [x10, #0x0]\n" - "fmla v8.8h, v6.8h, v0.h[0]\n" + "ldr q16, [x10, #0x0]\n" + "fmla v8.8h, v16.8h, v0.h[0]\n" "sub x27, x27, #0x1\n" - "ldr q7, [x10, #0x10]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.8h, v7.8h, v0.h[0]\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "ldr q7, [x10, #0x30]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" + "ldr q17, [x10, #0x10]\n" + "ldr q16, [x10, #0x20]\n" + "fmla v9.8h, v17.8h, v0.h[0]\n" + "fmla v10.8h, v16.8h, v0.h[0]\n" + "ldr q16, [x10, #0x30]\n" + "fmla v11.8h, v16.8h, v0.h[0]\n" "add x10, x10, #0x40\n" "cbnz x27, 29b\n" "30:" // Height 1: Multiply loop: No odd multiplies @@ -425,17 +425,17 @@ void a64_hybrid_fp16_mla_6x32 ( "prfm pstl1keep, [x9, #0x0]\n" "tbz %x[flags], #1, 31f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.8h }, [x20]\n" + "ld1r { v17.8h }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.8h }, [x20]\n" - "fmin v8.8h, v8.8h, v1.8h\n" - "fmin v9.8h, v9.8h, v1.8h\n" - "fmin v10.8h, v10.8h, v1.8h\n" - "fmin v11.8h, v11.8h, v1.8h\n" - "fmax v8.8h, v8.8h, v0.8h\n" - "fmax v9.8h, v9.8h, v0.8h\n" - "fmax v10.8h, v10.8h, v0.8h\n" - "fmax v11.8h, v11.8h, v0.8h\n" + "ld1r { v16.8h }, [x20]\n" + "fmin v8.8h, v8.8h, v17.8h\n" + "fmin v9.8h, v9.8h, v17.8h\n" + "fmin v10.8h, v10.8h, v17.8h\n" + "fmin v11.8h, v11.8h, v17.8h\n" + "fmax v8.8h, v8.8h, v16.8h\n" + "fmax v9.8h, v9.8h, v16.8h\n" + "fmax v10.8h, v10.8h, v16.8h\n" + "fmax v11.8h, v11.8h, v16.8h\n" "31:" // Height 1: No activation "cmp x11, #0x20\n" "bge 48f\n" @@ -733,12 +733,12 @@ void a64_hybrid_fp16_mla_6x32 ( "72:" // 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, 73f\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, 74f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -746,7 +746,7 @@ void a64_hybrid_fp16_mla_6x32 ( "b 74f\n" "73:" // Height 2: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" "74:" // Height 2: input setup done "cmp x27, #0x8\n" "blt 77f\n" @@ -759,230 +759,230 @@ void a64_hybrid_fp16_mla_6x32 ( "75:" // Height 2: Multiply loop: Main loop head "fmla v8.8h, v6.8h, v0.h[0]\n" "fmla v12.8h, v6.8h, v1.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "sub x27, x27, #0x8\n" "fmla v9.8h, v7.8h, v0.h[0]\n" "fmla v13.8h, v7.8h, v1.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q16, [x10, #0x30]\n" "add x26, x26, #0x10\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" - "ldr q6, [x10, #0x40]\n" + "fmla v10.8h, v17.8h, v0.h[0]\n" + "fmla v14.8h, v17.8h, v1.h[0]\n" + "ldr q17, [x10, #0x40]\n" "add x25, x25, #0x10\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" - "ldr q7, [x10, #0x50]\n" + "fmla v11.8h, v16.8h, v0.h[0]\n" + "fmla v15.8h, v16.8h, v1.h[0]\n" + "ldr q16, [x10, #0x50]\n" "cmp x27, #0x10\n" - "fmla v8.8h, v6.8h, v0.h[1]\n" - "fmla v12.8h, v6.8h, v1.h[1]\n" - "ldr q6, [x10, #0x60]\n" + "fmla v8.8h, v17.8h, v0.h[1]\n" + "fmla v12.8h, v17.8h, v1.h[1]\n" + "ldr q17, [x10, #0x60]\n" "prfm pldl1keep, [x26, #0x80]\n" - "fmla v9.8h, v7.8h, v0.h[1]\n" - "fmla v13.8h, v7.8h, v1.h[1]\n" - "ldr q7, [x10, #0x70]\n" + "fmla v9.8h, v16.8h, v0.h[1]\n" + "fmla v13.8h, v16.8h, v1.h[1]\n" + "ldr q16, [x10, #0x70]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v10.8h, v6.8h, v0.h[1]\n" - "fmla v14.8h, v6.8h, v1.h[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[1]\n" - "fmla v15.8h, v7.8h, v1.h[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.8h, v6.8h, v0.h[2]\n" - "fmla v12.8h, v6.8h, v1.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.8h, v7.8h, v0.h[2]\n" - "fmla v13.8h, v7.8h, v1.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.8h, v6.8h, v0.h[2]\n" - "fmla v14.8h, v6.8h, v1.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.8h, v7.8h, v0.h[2]\n" - "fmla v15.8h, v7.8h, v1.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.8h, v6.8h, v0.h[3]\n" - "fmla v12.8h, v6.8h, v1.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.8h, v7.8h, v0.h[3]\n" - "fmla v13.8h, v7.8h, v1.h[3]\n" - "ldr q7, [x10, #0xf0]\n" - "fmla v10.8h, v6.8h, v0.h[3]\n" - "fmla v14.8h, v6.8h, v1.h[3]\n" - "ldr q6, [x10, #0x100]\n" - "fmla v11.8h, v7.8h, v0.h[3]\n" - "fmla v15.8h, v7.8h, v1.h[3]\n" - "ldr q7, [x10, #0x110]\n" - "fmla v8.8h, v6.8h, v0.h[4]\n" - "fmla v12.8h, v6.8h, v1.h[4]\n" - "ldr q6, [x10, #0x120]\n" - "fmla v9.8h, v7.8h, v0.h[4]\n" - "fmla v13.8h, v7.8h, v1.h[4]\n" - "ldr q7, [x10, #0x130]\n" - "fmla v10.8h, v6.8h, v0.h[4]\n" - "fmla v14.8h, v6.8h, v1.h[4]\n" - "ldr q6, [x10, #0x140]\n" - "fmla v11.8h, v7.8h, v0.h[4]\n" - "fmla v15.8h, v7.8h, v1.h[4]\n" - "ldr q7, [x10, #0x150]\n" - "fmla v8.8h, v6.8h, v0.h[5]\n" - "fmla v12.8h, v6.8h, v1.h[5]\n" - "ldr q6, [x10, #0x160]\n" - "fmla v9.8h, v7.8h, v0.h[5]\n" - "fmla v13.8h, v7.8h, v1.h[5]\n" - "ldr q7, [x10, #0x170]\n" - "fmla v10.8h, v6.8h, v0.h[5]\n" - "fmla v14.8h, v6.8h, v1.h[5]\n" - "ldr q6, [x10, #0x180]\n" - "fmla v11.8h, v7.8h, v0.h[5]\n" - "fmla v15.8h, v7.8h, v1.h[5]\n" - "ldr q7, [x10, #0x190]\n" - "fmla v8.8h, v6.8h, v0.h[6]\n" - "fmla v12.8h, v6.8h, v1.h[6]\n" - "ldr q6, [x10, #0x1a0]\n" - "fmla v9.8h, v7.8h, v0.h[6]\n" - "fmla v13.8h, v7.8h, v1.h[6]\n" - "ldr q7, [x10, #0x1b0]\n" - "fmla v10.8h, v6.8h, v0.h[6]\n" - "fmla v14.8h, v6.8h, v1.h[6]\n" - "ldr q6, [x10, #0x1c0]\n" - "fmla v11.8h, v7.8h, v0.h[6]\n" - "fmla v15.8h, v7.8h, v1.h[6]\n" - "ldr q7, [x10, #0x1d0]\n" - "fmla v8.8h, v6.8h, v0.h[7]\n" - "fmla v12.8h, v6.8h, v1.h[7]\n" - "ldr q6, [x10, #0x1e0]\n" - "fmla v9.8h, v7.8h, v0.h[7]\n" - "fmla v13.8h, v7.8h, v1.h[7]\n" - "ldr q7, [x10, #0x1f0]\n" + "fmla v10.8h, v17.8h, v0.h[1]\n" + "fmla v14.8h, v17.8h, v1.h[1]\n" + "ldr q17, [x10, #0x80]\n" + "fmla v11.8h, v16.8h, v0.h[1]\n" + "fmla v15.8h, v16.8h, v1.h[1]\n" + "ldr q16, [x10, #0x90]\n" + "fmla v8.8h, v17.8h, v0.h[2]\n" + "fmla v12.8h, v17.8h, v1.h[2]\n" + "ldr q17, [x10, #0xa0]\n" + "fmla v9.8h, v16.8h, v0.h[2]\n" + "fmla v13.8h, v16.8h, v1.h[2]\n" + "ldr q16, [x10, #0xb0]\n" + "fmla v10.8h, v17.8h, v0.h[2]\n" + "fmla v14.8h, v17.8h, v1.h[2]\n" + "ldr q17, [x10, #0xc0]\n" + "fmla v11.8h, v16.8h, v0.h[2]\n" + "fmla v15.8h, v16.8h, v1.h[2]\n" + "ldr q16, [x10, #0xd0]\n" + "fmla v8.8h, v17.8h, v0.h[3]\n" + "fmla v12.8h, v17.8h, v1.h[3]\n" + "ldr q17, [x10, #0xe0]\n" + "fmla v9.8h, v16.8h, v0.h[3]\n" + "fmla v13.8h, v16.8h, v1.h[3]\n" + "ldr q16, [x10, #0xf0]\n" + "fmla v10.8h, v17.8h, v0.h[3]\n" + "fmla v14.8h, v17.8h, v1.h[3]\n" + "ldr q17, [x10, #0x100]\n" + "fmla v11.8h, v16.8h, v0.h[3]\n" + "fmla v15.8h, v16.8h, v1.h[3]\n" + "ldr q16, [x10, #0x110]\n" + "fmla v8.8h, v17.8h, v0.h[4]\n" + "fmla v12.8h, v17.8h, v1.h[4]\n" + "ldr q17, [x10, #0x120]\n" + "fmla v9.8h, v16.8h, v0.h[4]\n" + "fmla v13.8h, v16.8h, v1.h[4]\n" + "ldr q16, [x10, #0x130]\n" + "fmla v10.8h, v17.8h, v0.h[4]\n" + "fmla v14.8h, v17.8h, v1.h[4]\n" + "ldr q17, [x10, #0x140]\n" + "fmla v11.8h, v16.8h, v0.h[4]\n" + "fmla v15.8h, v16.8h, v1.h[4]\n" + "ldr q16, [x10, #0x150]\n" + "fmla v8.8h, v17.8h, v0.h[5]\n" + "fmla v12.8h, v17.8h, v1.h[5]\n" + "ldr q17, [x10, #0x160]\n" + "fmla v9.8h, v16.8h, v0.h[5]\n" + "fmla v13.8h, v16.8h, v1.h[5]\n" + "ldr q16, [x10, #0x170]\n" + "fmla v10.8h, v17.8h, v0.h[5]\n" + "fmla v14.8h, v17.8h, v1.h[5]\n" + "ldr q17, [x10, #0x180]\n" + "fmla v11.8h, v16.8h, v0.h[5]\n" + "fmla v15.8h, v16.8h, v1.h[5]\n" + "ldr q16, [x10, #0x190]\n" + "fmla v8.8h, v17.8h, v0.h[6]\n" + "fmla v12.8h, v17.8h, v1.h[6]\n" + "ldr q17, [x10, #0x1a0]\n" + "fmla v9.8h, v16.8h, v0.h[6]\n" + "fmla v13.8h, v16.8h, v1.h[6]\n" + "ldr q16, [x10, #0x1b0]\n" + "fmla v10.8h, v17.8h, v0.h[6]\n" + "fmla v14.8h, v17.8h, v1.h[6]\n" + "ldr q17, [x10, #0x1c0]\n" + "fmla v11.8h, v16.8h, v0.h[6]\n" + "fmla v15.8h, v16.8h, v1.h[6]\n" + "ldr q16, [x10, #0x1d0]\n" + "fmla v8.8h, v17.8h, v0.h[7]\n" + "fmla v12.8h, v17.8h, v1.h[7]\n" + "ldr q17, [x10, #0x1e0]\n" + "fmla v9.8h, v16.8h, v0.h[7]\n" + "fmla v13.8h, v16.8h, v1.h[7]\n" + "ldr q16, [x10, #0x1f0]\n" "add x10, x10, #0x200\n" - "fmla v10.8h, v6.8h, v0.h[7]\n" - "fmla v14.8h, v6.8h, v1.h[7]\n" + "fmla v10.8h, v17.8h, v0.h[7]\n" + "fmla v14.8h, v17.8h, v1.h[7]\n" "ldr q6, [x10, #0x0]\n" - "fmla v11.8h, v7.8h, v0.h[7]\n" + "fmla v11.8h, v16.8h, v0.h[7]\n" "ldr q0, [x26, #0x0]\n" - "fmla v15.8h, v7.8h, v1.h[7]\n" + "fmla v15.8h, v16.8h, v1.h[7]\n" "ldr q1, [x25, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 75b\n" "76:" // Height 2: Multiply loop: Single iteration only "fmla v8.8h, v6.8h, v0.h[0]\n" "fmla v12.8h, v6.8h, v1.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q17, [x10, #0x20]\n" "add x26, x26, #0x10\n" "fmla v9.8h, v7.8h, v0.h[0]\n" "fmla v13.8h, v7.8h, v1.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q16, [x10, #0x30]\n" "add x25, x25, #0x10\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" - "ldr q6, [x10, #0x40]\n" + "fmla v10.8h, v17.8h, v0.h[0]\n" + "fmla v14.8h, v17.8h, v1.h[0]\n" + "ldr q17, [x10, #0x40]\n" "sub x27, x27, #0x8\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" - "ldr q7, [x10, #0x50]\n" + "fmla v11.8h, v16.8h, v0.h[0]\n" + "fmla v15.8h, v16.8h, v1.h[0]\n" + "ldr q16, [x10, #0x50]\n" "prfm pldl1keep, [x26, #0x80]\n" - "fmla v8.8h, v6.8h, v0.h[1]\n" - "fmla v12.8h, v6.8h, v1.h[1]\n" - "ldr q6, [x10, #0x60]\n" + "fmla v8.8h, v17.8h, v0.h[1]\n" + "fmla v12.8h, v17.8h, v1.h[1]\n" + "ldr q17, [x10, #0x60]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v9.8h, v7.8h, v0.h[1]\n" - "fmla v13.8h, v7.8h, v1.h[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.8h, v6.8h, v0.h[1]\n" - "fmla v14.8h, v6.8h, v1.h[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[1]\n" - "fmla v15.8h, v7.8h, v1.h[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.8h, v6.8h, v0.h[2]\n" - "fmla v12.8h, v6.8h, v1.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.8h, v7.8h, v0.h[2]\n" - "fmla v13.8h, v7.8h, v1.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.8h, v6.8h, v0.h[2]\n" - "fmla v14.8h, v6.8h, v1.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.8h, v7.8h, v0.h[2]\n" - "fmla v15.8h, v7.8h, v1.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.8h, v6.8h, v0.h[3]\n" - "fmla v12.8h, v6.8h, v1.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.8h, v7.8h, v0.h[3]\n" - "fmla v13.8h, v7.8h, v1.h[3]\n" - "ldr q7, [x10, #0xf0]\n" - "fmla v10.8h, v6.8h, v0.h[3]\n" - "fmla v14.8h, v6.8h, v1.h[3]\n" - "ldr q6, [x10, #0x100]\n" - "fmla v11.8h, v7.8h, v0.h[3]\n" - "fmla v15.8h, v7.8h, v1.h[3]\n" - "ldr q7, [x10, #0x110]\n" - "fmla v8.8h, v6.8h, v0.h[4]\n" - "fmla v12.8h, v6.8h, v1.h[4]\n" - "ldr q6, [x10, #0x120]\n" - "fmla v9.8h, v7.8h, v0.h[4]\n" - "fmla v13.8h, v7.8h, v1.h[4]\n" - "ldr q7, [x10, #0x130]\n" - "fmla v10.8h, v6.8h, v0.h[4]\n" - "fmla v14.8h, v6.8h, v1.h[4]\n" - "ldr q6, [x10, #0x140]\n" - "fmla v11.8h, v7.8h, v0.h[4]\n" - "fmla v15.8h, v7.8h, v1.h[4]\n" - "ldr q7, [x10, #0x150]\n" - "fmla v8.8h, v6.8h, v0.h[5]\n" - "fmla v12.8h, v6.8h, v1.h[5]\n" - "ldr q6, [x10, #0x160]\n" - "fmla v9.8h, v7.8h, v0.h[5]\n" - "fmla v13.8h, v7.8h, v1.h[5]\n" - "ldr q7, [x10, #0x170]\n" - "fmla v10.8h, v6.8h, v0.h[5]\n" - "fmla v14.8h, v6.8h, v1.h[5]\n" - "ldr q6, [x10, #0x180]\n" - "fmla v11.8h, v7.8h, v0.h[5]\n" - "fmla v15.8h, v7.8h, v1.h[5]\n" - "ldr q7, [x10, #0x190]\n" - "fmla v8.8h, v6.8h, v0.h[6]\n" - "fmla v12.8h, v6.8h, v1.h[6]\n" - "ldr q6, [x10, #0x1a0]\n" - "fmla v9.8h, v7.8h, v0.h[6]\n" - "fmla v13.8h, v7.8h, v1.h[6]\n" - "ldr q7, [x10, #0x1b0]\n" - "fmla v10.8h, v6.8h, v0.h[6]\n" - "fmla v14.8h, v6.8h, v1.h[6]\n" - "ldr q6, [x10, #0x1c0]\n" - "fmla v11.8h, v7.8h, v0.h[6]\n" - "fmla v15.8h, v7.8h, v1.h[6]\n" - "ldr q7, [x10, #0x1d0]\n" - "fmla v8.8h, v6.8h, v0.h[7]\n" - "fmla v12.8h, v6.8h, v1.h[7]\n" - "ldr q6, [x10, #0x1e0]\n" - "fmla v9.8h, v7.8h, v0.h[7]\n" - "fmla v13.8h, v7.8h, v1.h[7]\n" - "ldr q7, [x10, #0x1f0]\n" + "fmla v9.8h, v16.8h, v0.h[1]\n" + "fmla v13.8h, v16.8h, v1.h[1]\n" + "ldr q16, [x10, #0x70]\n" + "fmla v10.8h, v17.8h, v0.h[1]\n" + "fmla v14.8h, v17.8h, v1.h[1]\n" + "ldr q17, [x10, #0x80]\n" + "fmla v11.8h, v16.8h, v0.h[1]\n" + "fmla v15.8h, v16.8h, v1.h[1]\n" + "ldr q16, [x10, #0x90]\n" + "fmla v8.8h, v17.8h, v0.h[2]\n" + "fmla v12.8h, v17.8h, v1.h[2]\n" + "ldr q17, [x10, #0xa0]\n" + "fmla v9.8h, v16.8h, v0.h[2]\n" + "fmla v13.8h, v16.8h, v1.h[2]\n" + "ldr q16, [x10, #0xb0]\n" + "fmla v10.8h, v17.8h, v0.h[2]\n" + "fmla v14.8h, v17.8h, v1.h[2]\n" + "ldr q17, [x10, #0xc0]\n" + "fmla v11.8h, v16.8h, v0.h[2]\n" + "fmla v15.8h, v16.8h, v1.h[2]\n" + "ldr q16, [x10, #0xd0]\n" + "fmla v8.8h, v17.8h, v0.h[3]\n" + "fmla v12.8h, v17.8h, v1.h[3]\n" + "ldr q17, [x10, #0xe0]\n" + "fmla v9.8h, v16.8h, v0.h[3]\n" + "fmla v13.8h, v16.8h, v1.h[3]\n" + "ldr q16, [x10, #0xf0]\n" + "fmla v10.8h, v17.8h, v0.h[3]\n" + "fmla v14.8h, v17.8h, v1.h[3]\n" + "ldr q17, [x10, #0x100]\n" + "fmla v11.8h, v16.8h, v0.h[3]\n" + "fmla v15.8h, v16.8h, v1.h[3]\n" + "ldr q16, [x10, #0x110]\n" + "fmla v8.8h, v17.8h, v0.h[4]\n" + "fmla v12.8h, v17.8h, v1.h[4]\n" + "ldr q17, [x10, #0x120]\n" + "fmla v9.8h, v16.8h, v0.h[4]\n" + "fmla v13.8h, v16.8h, v1.h[4]\n" + "ldr q16, [x10, #0x130]\n" + "fmla v10.8h, v17.8h, v0.h[4]\n" + "fmla v14.8h, v17.8h, v1.h[4]\n" + "ldr q17, [x10, #0x140]\n" + "fmla v11.8h, v16.8h, v0.h[4]\n" + "fmla v15.8h, v16.8h, v1.h[4]\n" + "ldr q16, [x10, #0x150]\n" + "fmla v8.8h, v17.8h, v0.h[5]\n" + "fmla v12.8h, v17.8h, v1.h[5]\n" + "ldr q17, [x10, #0x160]\n" + "fmla v9.8h, v16.8h, v0.h[5]\n" + "fmla v13.8h, v16.8h, v1.h[5]\n" + "ldr q16, [x10, #0x170]\n" + "fmla v10.8h, v17.8h, v0.h[5]\n" + "fmla v14.8h, v17.8h, v1.h[5]\n" + "ldr q17, [x10, #0x180]\n" + "fmla v11.8h, v16.8h, v0.h[5]\n" + "fmla v15.8h, v16.8h, v1.h[5]\n" + "ldr q16, [x10, #0x190]\n" + "fmla v8.8h, v17.8h, v0.h[6]\n" + "fmla v12.8h, v17.8h, v1.h[6]\n" + "ldr q17, [x10, #0x1a0]\n" + "fmla v9.8h, v16.8h, v0.h[6]\n" + "fmla v13.8h, v16.8h, v1.h[6]\n" + "ldr q16, [x10, #0x1b0]\n" + "fmla v10.8h, v17.8h, v0.h[6]\n" + "fmla v14.8h, v17.8h, v1.h[6]\n" + "ldr q17, [x10, #0x1c0]\n" + "fmla v11.8h, v16.8h, v0.h[6]\n" + "fmla v15.8h, v16.8h, v1.h[6]\n" + "ldr q16, [x10, #0x1d0]\n" + "fmla v8.8h, v17.8h, v0.h[7]\n" + "fmla v12.8h, v17.8h, v1.h[7]\n" + "ldr q17, [x10, #0x1e0]\n" + "fmla v9.8h, v16.8h, v0.h[7]\n" + "fmla v13.8h, v16.8h, v1.h[7]\n" + "ldr q16, [x10, #0x1f0]\n" "add x10, x10, #0x200\n" - "fmla v10.8h, v6.8h, v0.h[7]\n" - "fmla v14.8h, v6.8h, v1.h[7]\n" - "fmla v11.8h, v7.8h, v0.h[7]\n" - "fmla v15.8h, v7.8h, v1.h[7]\n" + "fmla v10.8h, v17.8h, v0.h[7]\n" + "fmla v14.8h, v17.8h, v1.h[7]\n" + "fmla v11.8h, v16.8h, v0.h[7]\n" + "fmla v15.8h, v16.8h, v1.h[7]\n" "77:" // Height 2: Multiply loop: Main loop skip "cbz x27, 79f\n" "78:" // Height 2: Multiply loop: Odd block loop - "ldr h0, [x26], #0x2\n" - "ldr h1, [x25], #0x2\n" + "ldr h1, [x26], #0x2\n" + "ldr h0, [x25], #0x2\n" "sub x27, x27, #0x1\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - "fmla v8.8h, v6.8h, v0.h[0]\n" - "fmla v12.8h, v6.8h, v1.h[0]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.8h, v7.8h, v0.h[0]\n" - "fmla v13.8h, v7.8h, v1.h[0]\n" - "ldr q7, [x10, #0x30]\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" + "ldr q17, [x10, #0x0]\n" + "ldr q16, [x10, #0x10]\n" + "fmla v8.8h, v17.8h, v1.h[0]\n" + "fmla v12.8h, v17.8h, v0.h[0]\n" + "ldr q17, [x10, #0x20]\n" + "fmla v9.8h, v16.8h, v1.h[0]\n" + "fmla v13.8h, v16.8h, v0.h[0]\n" + "ldr q16, [x10, #0x30]\n" + "fmla v10.8h, v17.8h, v1.h[0]\n" + "fmla v14.8h, v17.8h, v0.h[0]\n" "add x10, x10, #0x40\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" + "fmla v11.8h, v16.8h, v1.h[0]\n" + "fmla v15.8h, v16.8h, v0.h[0]\n" "cbnz x27, 78b\n" "79:" // Height 2: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -995,25 +995,25 @@ void a64_hybrid_fp16_mla_6x32 ( "prfm pstl1keep, [x25, #0x0]\n" "tbz %x[flags], #1, 80f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.8h }, [x20]\n" + "ld1r { v17.8h }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.8h }, [x20]\n" - "fmin v8.8h, v8.8h, v1.8h\n" - "fmin v9.8h, v9.8h, v1.8h\n" - "fmin v10.8h, v10.8h, v1.8h\n" - "fmin v11.8h, v11.8h, v1.8h\n" - "fmin v12.8h, v12.8h, v1.8h\n" - "fmin v13.8h, v13.8h, v1.8h\n" - "fmin v14.8h, v14.8h, v1.8h\n" - "fmin v15.8h, v15.8h, v1.8h\n" - "fmax v8.8h, v8.8h, v0.8h\n" - "fmax v9.8h, v9.8h, v0.8h\n" - "fmax v10.8h, v10.8h, v0.8h\n" - "fmax v11.8h, v11.8h, v0.8h\n" - "fmax v12.8h, v12.8h, v0.8h\n" - "fmax v13.8h, v13.8h, v0.8h\n" - "fmax v14.8h, v14.8h, v0.8h\n" - "fmax v15.8h, v15.8h, v0.8h\n" + "ld1r { v16.8h }, [x20]\n" + "fmin v8.8h, v8.8h, v17.8h\n" + "fmin v9.8h, v9.8h, v17.8h\n" + "fmin v10.8h, v10.8h, v17.8h\n" + "fmin v11.8h, v11.8h, v17.8h\n" + "fmin v12.8h, v12.8h, v17.8h\n" + "fmin v13.8h, v13.8h, v17.8h\n" + "fmin v14.8h, v14.8h, v17.8h\n" + "fmin v15.8h, v15.8h, v17.8h\n" + "fmax v8.8h, v8.8h, v16.8h\n" + "fmax v9.8h, v9.8h, v16.8h\n" + "fmax v10.8h, v10.8h, v16.8h\n" + "fmax v11.8h, v11.8h, v16.8h\n" + "fmax v12.8h, v12.8h, v16.8h\n" + "fmax v13.8h, v13.8h, v16.8h\n" + "fmax v14.8h, v14.8h, v16.8h\n" + "fmax v15.8h, v15.8h, v16.8h\n" "80:" // Height 2: No activation "cmp x11, #0x20\n" "bge 97f\n" @@ -1392,13 +1392,13 @@ void a64_hybrid_fp16_mla_6x32 ( "121:" // 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, 122f\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, 123f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -1407,8 +1407,8 @@ void a64_hybrid_fp16_mla_6x32 ( "b 123f\n" "122:" // Height 3: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" "123:" // Height 3: input setup done "cmp x27, #0x8\n" "blt 126f\n" @@ -1425,139 +1425,139 @@ void a64_hybrid_fp16_mla_6x32 ( "sub x27, x27, #0x8\n" "add x26, x26, #0x10\n" "fmla v16.8h, v6.8h, v2.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q21, [x10, #0x20]\n" "fmla v9.8h, v7.8h, v0.h[0]\n" "add x25, x25, #0x10\n" "fmla v13.8h, v7.8h, v1.h[0]\n" "fmla v17.8h, v7.8h, v2.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q20, [x10, #0x30]\n" "add x24, x24, #0x10\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" + "fmla v10.8h, v21.8h, v0.h[0]\n" + "fmla v14.8h, v21.8h, v1.h[0]\n" "cmp x27, #0x10\n" "prfm pldl1keep, [x26, #0x80]\n" - "fmla v18.8h, v6.8h, v2.h[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" + "fmla v18.8h, v21.8h, v2.h[0]\n" + "ldr q21, [x10, #0x40]\n" + "fmla v11.8h, v20.8h, v0.h[0]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" - "fmla v19.8h, v7.8h, v2.h[0]\n" - "ldr q7, [x10, #0x50]\n" + "fmla v15.8h, v20.8h, v1.h[0]\n" + "fmla v19.8h, v20.8h, v2.h[0]\n" + "ldr q20, [x10, #0x50]\n" "prfm pldl1keep, [x24, #0x80]\n" - "fmla v8.8h, v6.8h, v0.h[1]\n" - "fmla v12.8h, v6.8h, v1.h[1]\n" - "fmla v16.8h, v6.8h, v2.h[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.8h, v7.8h, v0.h[1]\n" - "fmla v13.8h, v7.8h, v1.h[1]\n" - "fmla v17.8h, v7.8h, v2.h[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.8h, v6.8h, v0.h[1]\n" - "fmla v14.8h, v6.8h, v1.h[1]\n" - "fmla v18.8h, v6.8h, v2.h[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[1]\n" - "fmla v15.8h, v7.8h, v1.h[1]\n" - "fmla v19.8h, v7.8h, v2.h[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.8h, v6.8h, v0.h[2]\n" - "fmla v12.8h, v6.8h, v1.h[2]\n" - "fmla v16.8h, v6.8h, v2.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.8h, v7.8h, v0.h[2]\n" - "fmla v13.8h, v7.8h, v1.h[2]\n" - "fmla v17.8h, v7.8h, v2.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.8h, v6.8h, v0.h[2]\n" - "fmla v14.8h, v6.8h, v1.h[2]\n" - "fmla v18.8h, v6.8h, v2.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.8h, v7.8h, v0.h[2]\n" - "fmla v15.8h, v7.8h, v1.h[2]\n" - "fmla v19.8h, v7.8h, v2.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.8h, v6.8h, v0.h[3]\n" - "fmla v12.8h, v6.8h, v1.h[3]\n" - "fmla v16.8h, v6.8h, v2.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.8h, v7.8h, v0.h[3]\n" - "fmla v13.8h, v7.8h, v1.h[3]\n" - "fmla v17.8h, v7.8h, v2.h[3]\n" - "ldr q7, [x10, #0xf0]\n" - "fmla v10.8h, v6.8h, v0.h[3]\n" - "fmla v14.8h, v6.8h, v1.h[3]\n" - "fmla v18.8h, v6.8h, v2.h[3]\n" - "ldr q6, [x10, #0x100]\n" - "fmla v11.8h, v7.8h, v0.h[3]\n" - "fmla v15.8h, v7.8h, v1.h[3]\n" - "fmla v19.8h, v7.8h, v2.h[3]\n" - "ldr q7, [x10, #0x110]\n" - "fmla v8.8h, v6.8h, v0.h[4]\n" - "fmla v12.8h, v6.8h, v1.h[4]\n" - "fmla v16.8h, v6.8h, v2.h[4]\n" - "ldr q6, [x10, #0x120]\n" - "fmla v9.8h, v7.8h, v0.h[4]\n" - "fmla v13.8h, v7.8h, v1.h[4]\n" - "fmla v17.8h, v7.8h, v2.h[4]\n" - "ldr q7, [x10, #0x130]\n" - "fmla v10.8h, v6.8h, v0.h[4]\n" - "fmla v14.8h, v6.8h, v1.h[4]\n" - "fmla v18.8h, v6.8h, v2.h[4]\n" - "ldr q6, [x10, #0x140]\n" - "fmla v11.8h, v7.8h, v0.h[4]\n" - "fmla v15.8h, v7.8h, v1.h[4]\n" - "fmla v19.8h, v7.8h, v2.h[4]\n" - "ldr q7, [x10, #0x150]\n" - "fmla v8.8h, v6.8h, v0.h[5]\n" - "fmla v12.8h, v6.8h, v1.h[5]\n" - "fmla v16.8h, v6.8h, v2.h[5]\n" - "ldr q6, [x10, #0x160]\n" - "fmla v9.8h, v7.8h, v0.h[5]\n" - "fmla v13.8h, v7.8h, v1.h[5]\n" - "fmla v17.8h, v7.8h, v2.h[5]\n" - "ldr q7, [x10, #0x170]\n" - "fmla v10.8h, v6.8h, v0.h[5]\n" - "fmla v14.8h, v6.8h, v1.h[5]\n" - "fmla v18.8h, v6.8h, v2.h[5]\n" - "ldr q6, [x10, #0x180]\n" - "fmla v11.8h, v7.8h, v0.h[5]\n" - "fmla v15.8h, v7.8h, v1.h[5]\n" - "fmla v19.8h, v7.8h, v2.h[5]\n" - "ldr q7, [x10, #0x190]\n" - "fmla v8.8h, v6.8h, v0.h[6]\n" - "fmla v12.8h, v6.8h, v1.h[6]\n" - "fmla v16.8h, v6.8h, v2.h[6]\n" - "ldr q6, [x10, #0x1a0]\n" - "fmla v9.8h, v7.8h, v0.h[6]\n" - "fmla v13.8h, v7.8h, v1.h[6]\n" - "fmla v17.8h, v7.8h, v2.h[6]\n" - "ldr q7, [x10, #0x1b0]\n" - "fmla v10.8h, v6.8h, v0.h[6]\n" - "fmla v14.8h, v6.8h, v1.h[6]\n" - "fmla v18.8h, v6.8h, v2.h[6]\n" - "ldr q6, [x10, #0x1c0]\n" - "fmla v11.8h, v7.8h, v0.h[6]\n" - "fmla v15.8h, v7.8h, v1.h[6]\n" - "fmla v19.8h, v7.8h, v2.h[6]\n" - "ldr q7, [x10, #0x1d0]\n" - "fmla v8.8h, v6.8h, v0.h[7]\n" - "fmla v12.8h, v6.8h, v1.h[7]\n" - "fmla v16.8h, v6.8h, v2.h[7]\n" - "ldr q6, [x10, #0x1e0]\n" - "fmla v9.8h, v7.8h, v0.h[7]\n" - "fmla v13.8h, v7.8h, v1.h[7]\n" - "fmla v17.8h, v7.8h, v2.h[7]\n" - "ldr q7, [x10, #0x1f0]\n" + "fmla v8.8h, v21.8h, v0.h[1]\n" + "fmla v12.8h, v21.8h, v1.h[1]\n" + "fmla v16.8h, v21.8h, v2.h[1]\n" + "ldr q21, [x10, #0x60]\n" + "fmla v9.8h, v20.8h, v0.h[1]\n" + "fmla v13.8h, v20.8h, v1.h[1]\n" + "fmla v17.8h, v20.8h, v2.h[1]\n" + "ldr q20, [x10, #0x70]\n" + "fmla v10.8h, v21.8h, v0.h[1]\n" + "fmla v14.8h, v21.8h, v1.h[1]\n" + "fmla v18.8h, v21.8h, v2.h[1]\n" + "ldr q21, [x10, #0x80]\n" + "fmla v11.8h, v20.8h, v0.h[1]\n" + "fmla v15.8h, v20.8h, v1.h[1]\n" + "fmla v19.8h, v20.8h, v2.h[1]\n" + "ldr q20, [x10, #0x90]\n" + "fmla v8.8h, v21.8h, v0.h[2]\n" + "fmla v12.8h, v21.8h, v1.h[2]\n" + "fmla v16.8h, v21.8h, v2.h[2]\n" + "ldr q21, [x10, #0xa0]\n" + "fmla v9.8h, v20.8h, v0.h[2]\n" + "fmla v13.8h, v20.8h, v1.h[2]\n" + "fmla v17.8h, v20.8h, v2.h[2]\n" + "ldr q20, [x10, #0xb0]\n" + "fmla v10.8h, v21.8h, v0.h[2]\n" + "fmla v14.8h, v21.8h, v1.h[2]\n" + "fmla v18.8h, v21.8h, v2.h[2]\n" + "ldr q21, [x10, #0xc0]\n" + "fmla v11.8h, v20.8h, v0.h[2]\n" + "fmla v15.8h, v20.8h, v1.h[2]\n" + "fmla v19.8h, v20.8h, v2.h[2]\n" + "ldr q20, [x10, #0xd0]\n" + "fmla v8.8h, v21.8h, v0.h[3]\n" + "fmla v12.8h, v21.8h, v1.h[3]\n" + "fmla v16.8h, v21.8h, v2.h[3]\n" + "ldr q21, [x10, #0xe0]\n" + "fmla v9.8h, v20.8h, v0.h[3]\n" + "fmla v13.8h, v20.8h, v1.h[3]\n" + "fmla v17.8h, v20.8h, v2.h[3]\n" + "ldr q20, [x10, #0xf0]\n" + "fmla v10.8h, v21.8h, v0.h[3]\n" + "fmla v14.8h, v21.8h, v1.h[3]\n" + "fmla v18.8h, v21.8h, v2.h[3]\n" + "ldr q21, [x10, #0x100]\n" + "fmla v11.8h, v20.8h, v0.h[3]\n" + "fmla v15.8h, v20.8h, v1.h[3]\n" + "fmla v19.8h, v20.8h, v2.h[3]\n" + "ldr q20, [x10, #0x110]\n" + "fmla v8.8h, v21.8h, v0.h[4]\n" + "fmla v12.8h, v21.8h, v1.h[4]\n" + "fmla v16.8h, v21.8h, v2.h[4]\n" + "ldr q21, [x10, #0x120]\n" + "fmla v9.8h, v20.8h, v0.h[4]\n" + "fmla v13.8h, v20.8h, v1.h[4]\n" + "fmla v17.8h, v20.8h, v2.h[4]\n" + "ldr q20, [x10, #0x130]\n" + "fmla v10.8h, v21.8h, v0.h[4]\n" + "fmla v14.8h, v21.8h, v1.h[4]\n" + "fmla v18.8h, v21.8h, v2.h[4]\n" + "ldr q21, [x10, #0x140]\n" + "fmla v11.8h, v20.8h, v0.h[4]\n" + "fmla v15.8h, v20.8h, v1.h[4]\n" + "fmla v19.8h, v20.8h, v2.h[4]\n" + "ldr q20, [x10, #0x150]\n" + "fmla v8.8h, v21.8h, v0.h[5]\n" + "fmla v12.8h, v21.8h, v1.h[5]\n" + "fmla v16.8h, v21.8h, v2.h[5]\n" + "ldr q21, [x10, #0x160]\n" + "fmla v9.8h, v20.8h, v0.h[5]\n" + "fmla v13.8h, v20.8h, v1.h[5]\n" + "fmla v17.8h, v20.8h, v2.h[5]\n" + "ldr q20, [x10, #0x170]\n" + "fmla v10.8h, v21.8h, v0.h[5]\n" + "fmla v14.8h, v21.8h, v1.h[5]\n" + "fmla v18.8h, v21.8h, v2.h[5]\n" + "ldr q21, [x10, #0x180]\n" + "fmla v11.8h, v20.8h, v0.h[5]\n" + "fmla v15.8h, v20.8h, v1.h[5]\n" + "fmla v19.8h, v20.8h, v2.h[5]\n" + "ldr q20, [x10, #0x190]\n" + "fmla v8.8h, v21.8h, v0.h[6]\n" + "fmla v12.8h, v21.8h, v1.h[6]\n" + "fmla v16.8h, v21.8h, v2.h[6]\n" + "ldr q21, [x10, #0x1a0]\n" + "fmla v9.8h, v20.8h, v0.h[6]\n" + "fmla v13.8h, v20.8h, v1.h[6]\n" + "fmla v17.8h, v20.8h, v2.h[6]\n" + "ldr q20, [x10, #0x1b0]\n" + "fmla v10.8h, v21.8h, v0.h[6]\n" + "fmla v14.8h, v21.8h, v1.h[6]\n" + "fmla v18.8h, v21.8h, v2.h[6]\n" + "ldr q21, [x10, #0x1c0]\n" + "fmla v11.8h, v20.8h, v0.h[6]\n" + "fmla v15.8h, v20.8h, v1.h[6]\n" + "fmla v19.8h, v20.8h, v2.h[6]\n" + "ldr q20, [x10, #0x1d0]\n" + "fmla v8.8h, v21.8h, v0.h[7]\n" + "fmla v12.8h, v21.8h, v1.h[7]\n" + "fmla v16.8h, v21.8h, v2.h[7]\n" + "ldr q21, [x10, #0x1e0]\n" + "fmla v9.8h, v20.8h, v0.h[7]\n" + "fmla v13.8h, v20.8h, v1.h[7]\n" + "fmla v17.8h, v20.8h, v2.h[7]\n" + "ldr q20, [x10, #0x1f0]\n" "add x10, x10, #0x200\n" - "fmla v10.8h, v6.8h, v0.h[7]\n" - "fmla v14.8h, v6.8h, v1.h[7]\n" - "fmla v18.8h, v6.8h, v2.h[7]\n" + "fmla v10.8h, v21.8h, v0.h[7]\n" + "fmla v14.8h, v21.8h, v1.h[7]\n" + "fmla v18.8h, v21.8h, v2.h[7]\n" "ldr q6, [x10, #0x0]\n" - "fmla v11.8h, v7.8h, v0.h[7]\n" + "fmla v11.8h, v20.8h, v0.h[7]\n" "ldr q0, [x26, #0x0]\n" - "fmla v15.8h, v7.8h, v1.h[7]\n" + "fmla v15.8h, v20.8h, v1.h[7]\n" "ldr q1, [x25, #0x0]\n" - "fmla v19.8h, v7.8h, v2.h[7]\n" + "fmla v19.8h, v20.8h, v2.h[7]\n" "ldr q2, [x24, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 124b\n" @@ -1567,159 +1567,159 @@ void a64_hybrid_fp16_mla_6x32 ( "add x26, x26, #0x10\n" "add x25, x25, #0x10\n" "fmla v16.8h, v6.8h, v2.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q21, [x10, #0x20]\n" "fmla v9.8h, v7.8h, v0.h[0]\n" "add x24, x24, #0x10\n" "fmla v13.8h, v7.8h, v1.h[0]\n" "fmla v17.8h, v7.8h, v2.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q20, [x10, #0x30]\n" "sub x27, x27, #0x8\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" + "fmla v10.8h, v21.8h, v0.h[0]\n" + "fmla v14.8h, v21.8h, v1.h[0]\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v18.8h, v6.8h, v2.h[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" + "fmla v18.8h, v21.8h, v2.h[0]\n" + "ldr q21, [x10, #0x40]\n" + "fmla v11.8h, v20.8h, v0.h[0]\n" "prfm pldl1keep, [x24, #0x80]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" - "fmla v19.8h, v7.8h, v2.h[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.8h, v6.8h, v0.h[1]\n" - "fmla v12.8h, v6.8h, v1.h[1]\n" - "fmla v16.8h, v6.8h, v2.h[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.8h, v7.8h, v0.h[1]\n" - "fmla v13.8h, v7.8h, v1.h[1]\n" - "fmla v17.8h, v7.8h, v2.h[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.8h, v6.8h, v0.h[1]\n" - "fmla v14.8h, v6.8h, v1.h[1]\n" - "fmla v18.8h, v6.8h, v2.h[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[1]\n" - "fmla v15.8h, v7.8h, v1.h[1]\n" - "fmla v19.8h, v7.8h, v2.h[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.8h, v6.8h, v0.h[2]\n" - "fmla v12.8h, v6.8h, v1.h[2]\n" - "fmla v16.8h, v6.8h, v2.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.8h, v7.8h, v0.h[2]\n" - "fmla v13.8h, v7.8h, v1.h[2]\n" - "fmla v17.8h, v7.8h, v2.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.8h, v6.8h, v0.h[2]\n" - "fmla v14.8h, v6.8h, v1.h[2]\n" - "fmla v18.8h, v6.8h, v2.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.8h, v7.8h, v0.h[2]\n" - "fmla v15.8h, v7.8h, v1.h[2]\n" - "fmla v19.8h, v7.8h, v2.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.8h, v6.8h, v0.h[3]\n" - "fmla v12.8h, v6.8h, v1.h[3]\n" - "fmla v16.8h, v6.8h, v2.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.8h, v7.8h, v0.h[3]\n" - "fmla v13.8h, v7.8h, v1.h[3]\n" - "fmla v17.8h, v7.8h, v2.h[3]\n" - "ldr q7, [x10, #0xf0]\n" - "fmla v10.8h, v6.8h, v0.h[3]\n" - "fmla v14.8h, v6.8h, v1.h[3]\n" - "fmla v18.8h, v6.8h, v2.h[3]\n" - "ldr q6, [x10, #0x100]\n" - "fmla v11.8h, v7.8h, v0.h[3]\n" - "fmla v15.8h, v7.8h, v1.h[3]\n" - "fmla v19.8h, v7.8h, v2.h[3]\n" - "ldr q7, [x10, #0x110]\n" - "fmla v8.8h, v6.8h, v0.h[4]\n" - "fmla v12.8h, v6.8h, v1.h[4]\n" - "fmla v16.8h, v6.8h, v2.h[4]\n" - "ldr q6, [x10, #0x120]\n" - "fmla v9.8h, v7.8h, v0.h[4]\n" - "fmla v13.8h, v7.8h, v1.h[4]\n" - "fmla v17.8h, v7.8h, v2.h[4]\n" - "ldr q7, [x10, #0x130]\n" - "fmla v10.8h, v6.8h, v0.h[4]\n" - "fmla v14.8h, v6.8h, v1.h[4]\n" - "fmla v18.8h, v6.8h, v2.h[4]\n" - "ldr q6, [x10, #0x140]\n" - "fmla v11.8h, v7.8h, v0.h[4]\n" - "fmla v15.8h, v7.8h, v1.h[4]\n" - "fmla v19.8h, v7.8h, v2.h[4]\n" - "ldr q7, [x10, #0x150]\n" - "fmla v8.8h, v6.8h, v0.h[5]\n" - "fmla v12.8h, v6.8h, v1.h[5]\n" - "fmla v16.8h, v6.8h, v2.h[5]\n" - "ldr q6, [x10, #0x160]\n" - "fmla v9.8h, v7.8h, v0.h[5]\n" - "fmla v13.8h, v7.8h, v1.h[5]\n" - "fmla v17.8h, v7.8h, v2.h[5]\n" - "ldr q7, [x10, #0x170]\n" - "fmla v10.8h, v6.8h, v0.h[5]\n" - "fmla v14.8h, v6.8h, v1.h[5]\n" - "fmla v18.8h, v6.8h, v2.h[5]\n" - "ldr q6, [x10, #0x180]\n" - "fmla v11.8h, v7.8h, v0.h[5]\n" - "fmla v15.8h, v7.8h, v1.h[5]\n" - "fmla v19.8h, v7.8h, v2.h[5]\n" - "ldr q7, [x10, #0x190]\n" - "fmla v8.8h, v6.8h, v0.h[6]\n" - "fmla v12.8h, v6.8h, v1.h[6]\n" - "fmla v16.8h, v6.8h, v2.h[6]\n" - "ldr q6, [x10, #0x1a0]\n" - "fmla v9.8h, v7.8h, v0.h[6]\n" - "fmla v13.8h, v7.8h, v1.h[6]\n" - "fmla v17.8h, v7.8h, v2.h[6]\n" - "ldr q7, [x10, #0x1b0]\n" - "fmla v10.8h, v6.8h, v0.h[6]\n" - "fmla v14.8h, v6.8h, v1.h[6]\n" - "fmla v18.8h, v6.8h, v2.h[6]\n" - "ldr q6, [x10, #0x1c0]\n" - "fmla v11.8h, v7.8h, v0.h[6]\n" - "fmla v15.8h, v7.8h, v1.h[6]\n" - "fmla v19.8h, v7.8h, v2.h[6]\n" - "ldr q7, [x10, #0x1d0]\n" - "fmla v8.8h, v6.8h, v0.h[7]\n" - "fmla v12.8h, v6.8h, v1.h[7]\n" - "fmla v16.8h, v6.8h, v2.h[7]\n" - "ldr q6, [x10, #0x1e0]\n" - "fmla v9.8h, v7.8h, v0.h[7]\n" - "fmla v13.8h, v7.8h, v1.h[7]\n" - "fmla v17.8h, v7.8h, v2.h[7]\n" - "ldr q7, [x10, #0x1f0]\n" + "fmla v15.8h, v20.8h, v1.h[0]\n" + "fmla v19.8h, v20.8h, v2.h[0]\n" + "ldr q20, [x10, #0x50]\n" + "fmla v8.8h, v21.8h, v0.h[1]\n" + "fmla v12.8h, v21.8h, v1.h[1]\n" + "fmla v16.8h, v21.8h, v2.h[1]\n" + "ldr q21, [x10, #0x60]\n" + "fmla v9.8h, v20.8h, v0.h[1]\n" + "fmla v13.8h, v20.8h, v1.h[1]\n" + "fmla v17.8h, v20.8h, v2.h[1]\n" + "ldr q20, [x10, #0x70]\n" + "fmla v10.8h, v21.8h, v0.h[1]\n" + "fmla v14.8h, v21.8h, v1.h[1]\n" + "fmla v18.8h, v21.8h, v2.h[1]\n" + "ldr q21, [x10, #0x80]\n" + "fmla v11.8h, v20.8h, v0.h[1]\n" + "fmla v15.8h, v20.8h, v1.h[1]\n" + "fmla v19.8h, v20.8h, v2.h[1]\n" + "ldr q20, [x10, #0x90]\n" + "fmla v8.8h, v21.8h, v0.h[2]\n" + "fmla v12.8h, v21.8h, v1.h[2]\n" + "fmla v16.8h, v21.8h, v2.h[2]\n" + "ldr q21, [x10, #0xa0]\n" + "fmla v9.8h, v20.8h, v0.h[2]\n" + "fmla v13.8h, v20.8h, v1.h[2]\n" + "fmla v17.8h, v20.8h, v2.h[2]\n" + "ldr q20, [x10, #0xb0]\n" + "fmla v10.8h, v21.8h, v0.h[2]\n" + "fmla v14.8h, v21.8h, v1.h[2]\n" + "fmla v18.8h, v21.8h, v2.h[2]\n" + "ldr q21, [x10, #0xc0]\n" + "fmla v11.8h, v20.8h, v0.h[2]\n" + "fmla v15.8h, v20.8h, v1.h[2]\n" + "fmla v19.8h, v20.8h, v2.h[2]\n" + "ldr q20, [x10, #0xd0]\n" + "fmla v8.8h, v21.8h, v0.h[3]\n" + "fmla v12.8h, v21.8h, v1.h[3]\n" + "fmla v16.8h, v21.8h, v2.h[3]\n" + "ldr q21, [x10, #0xe0]\n" + "fmla v9.8h, v20.8h, v0.h[3]\n" + "fmla v13.8h, v20.8h, v1.h[3]\n" + "fmla v17.8h, v20.8h, v2.h[3]\n" + "ldr q20, [x10, #0xf0]\n" + "fmla v10.8h, v21.8h, v0.h[3]\n" + "fmla v14.8h, v21.8h, v1.h[3]\n" + "fmla v18.8h, v21.8h, v2.h[3]\n" + "ldr q21, [x10, #0x100]\n" + "fmla v11.8h, v20.8h, v0.h[3]\n" + "fmla v15.8h, v20.8h, v1.h[3]\n" + "fmla v19.8h, v20.8h, v2.h[3]\n" + "ldr q20, [x10, #0x110]\n" + "fmla v8.8h, v21.8h, v0.h[4]\n" + "fmla v12.8h, v21.8h, v1.h[4]\n" + "fmla v16.8h, v21.8h, v2.h[4]\n" + "ldr q21, [x10, #0x120]\n" + "fmla v9.8h, v20.8h, v0.h[4]\n" + "fmla v13.8h, v20.8h, v1.h[4]\n" + "fmla v17.8h, v20.8h, v2.h[4]\n" + "ldr q20, [x10, #0x130]\n" + "fmla v10.8h, v21.8h, v0.h[4]\n" + "fmla v14.8h, v21.8h, v1.h[4]\n" + "fmla v18.8h, v21.8h, v2.h[4]\n" + "ldr q21, [x10, #0x140]\n" + "fmla v11.8h, v20.8h, v0.h[4]\n" + "fmla v15.8h, v20.8h, v1.h[4]\n" + "fmla v19.8h, v20.8h, v2.h[4]\n" + "ldr q20, [x10, #0x150]\n" + "fmla v8.8h, v21.8h, v0.h[5]\n" + "fmla v12.8h, v21.8h, v1.h[5]\n" + "fmla v16.8h, v21.8h, v2.h[5]\n" + "ldr q21, [x10, #0x160]\n" + "fmla v9.8h, v20.8h, v0.h[5]\n" + "fmla v13.8h, v20.8h, v1.h[5]\n" + "fmla v17.8h, v20.8h, v2.h[5]\n" + "ldr q20, [x10, #0x170]\n" + "fmla v10.8h, v21.8h, v0.h[5]\n" + "fmla v14.8h, v21.8h, v1.h[5]\n" + "fmla v18.8h, v21.8h, v2.h[5]\n" + "ldr q21, [x10, #0x180]\n" + "fmla v11.8h, v20.8h, v0.h[5]\n" + "fmla v15.8h, v20.8h, v1.h[5]\n" + "fmla v19.8h, v20.8h, v2.h[5]\n" + "ldr q20, [x10, #0x190]\n" + "fmla v8.8h, v21.8h, v0.h[6]\n" + "fmla v12.8h, v21.8h, v1.h[6]\n" + "fmla v16.8h, v21.8h, v2.h[6]\n" + "ldr q21, [x10, #0x1a0]\n" + "fmla v9.8h, v20.8h, v0.h[6]\n" + "fmla v13.8h, v20.8h, v1.h[6]\n" + "fmla v17.8h, v20.8h, v2.h[6]\n" + "ldr q20, [x10, #0x1b0]\n" + "fmla v10.8h, v21.8h, v0.h[6]\n" + "fmla v14.8h, v21.8h, v1.h[6]\n" + "fmla v18.8h, v21.8h, v2.h[6]\n" + "ldr q21, [x10, #0x1c0]\n" + "fmla v11.8h, v20.8h, v0.h[6]\n" + "fmla v15.8h, v20.8h, v1.h[6]\n" + "fmla v19.8h, v20.8h, v2.h[6]\n" + "ldr q20, [x10, #0x1d0]\n" + "fmla v8.8h, v21.8h, v0.h[7]\n" + "fmla v12.8h, v21.8h, v1.h[7]\n" + "fmla v16.8h, v21.8h, v2.h[7]\n" + "ldr q21, [x10, #0x1e0]\n" + "fmla v9.8h, v20.8h, v0.h[7]\n" + "fmla v13.8h, v20.8h, v1.h[7]\n" + "fmla v17.8h, v20.8h, v2.h[7]\n" + "ldr q20, [x10, #0x1f0]\n" "add x10, x10, #0x200\n" - "fmla v10.8h, v6.8h, v0.h[7]\n" - "fmla v14.8h, v6.8h, v1.h[7]\n" - "fmla v18.8h, v6.8h, v2.h[7]\n" - "fmla v11.8h, v7.8h, v0.h[7]\n" - "fmla v15.8h, v7.8h, v1.h[7]\n" - "fmla v19.8h, v7.8h, v2.h[7]\n" + "fmla v10.8h, v21.8h, v0.h[7]\n" + "fmla v14.8h, v21.8h, v1.h[7]\n" + "fmla v18.8h, v21.8h, v2.h[7]\n" + "fmla v11.8h, v20.8h, v0.h[7]\n" + "fmla v15.8h, v20.8h, v1.h[7]\n" + "fmla v19.8h, v20.8h, v2.h[7]\n" "126:" // Height 3: Multiply loop: Main loop skip "cbz x27, 128f\n" "127:" // Height 3: Multiply loop: Odd block loop - "ldr h0, [x26], #0x2\n" + "ldr h2, [x26], #0x2\n" "ldr h1, [x25], #0x2\n" "sub x27, x27, #0x1\n" - "ldr h2, [x24], #0x2\n" - "ldr q6, [x10, #0x0]\n" - "fmla v8.8h, v6.8h, v0.h[0]\n" - "fmla v12.8h, v6.8h, v1.h[0]\n" - "ldr q7, [x10, #0x10]\n" - "fmla v16.8h, v6.8h, v2.h[0]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.8h, v7.8h, v0.h[0]\n" - "fmla v13.8h, v7.8h, v1.h[0]\n" - "fmla v17.8h, v7.8h, v2.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr h0, [x24], #0x2\n" + "ldr q21, [x10, #0x0]\n" + "fmla v8.8h, v21.8h, v2.h[0]\n" + "fmla v12.8h, v21.8h, v1.h[0]\n" + "ldr q20, [x10, #0x10]\n" + "fmla v16.8h, v21.8h, v0.h[0]\n" + "ldr q21, [x10, #0x20]\n" + "fmla v9.8h, v20.8h, v2.h[0]\n" + "fmla v13.8h, v20.8h, v1.h[0]\n" + "fmla v17.8h, v20.8h, v0.h[0]\n" + "ldr q20, [x10, #0x30]\n" "add x10, x10, #0x40\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" - "fmla v18.8h, v6.8h, v2.h[0]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" - "fmla v19.8h, v7.8h, v2.h[0]\n" + "fmla v10.8h, v21.8h, v2.h[0]\n" + "fmla v14.8h, v21.8h, v1.h[0]\n" + "fmla v18.8h, v21.8h, v0.h[0]\n" + "fmla v11.8h, v20.8h, v2.h[0]\n" + "fmla v15.8h, v20.8h, v1.h[0]\n" + "fmla v19.8h, v20.8h, v0.h[0]\n" "cbnz x27, 127b\n" "128:" // Height 3: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -1734,33 +1734,33 @@ void a64_hybrid_fp16_mla_6x32 ( "prfm pstl1keep, [x24, #0x0]\n" "tbz %x[flags], #1, 129f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.8h }, [x20]\n" + "ld1r { v21.8h }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.8h }, [x20]\n" - "fmin v8.8h, v8.8h, v1.8h\n" - "fmin v9.8h, v9.8h, v1.8h\n" - "fmin v10.8h, v10.8h, v1.8h\n" - "fmin v11.8h, v11.8h, v1.8h\n" - "fmin v12.8h, v12.8h, v1.8h\n" - "fmin v13.8h, v13.8h, v1.8h\n" - "fmin v14.8h, v14.8h, v1.8h\n" - "fmin v15.8h, v15.8h, v1.8h\n" - "fmin v16.8h, v16.8h, v1.8h\n" - "fmin v17.8h, v17.8h, v1.8h\n" - "fmin v18.8h, v18.8h, v1.8h\n" - "fmin v19.8h, v19.8h, v1.8h\n" - "fmax v8.8h, v8.8h, v0.8h\n" - "fmax v9.8h, v9.8h, v0.8h\n" - "fmax v10.8h, v10.8h, v0.8h\n" - "fmax v11.8h, v11.8h, v0.8h\n" - "fmax v12.8h, v12.8h, v0.8h\n" - "fmax v13.8h, v13.8h, v0.8h\n" - "fmax v14.8h, v14.8h, v0.8h\n" - "fmax v15.8h, v15.8h, v0.8h\n" - "fmax v16.8h, v16.8h, v0.8h\n" - "fmax v17.8h, v17.8h, v0.8h\n" - "fmax v18.8h, v18.8h, v0.8h\n" - "fmax v19.8h, v19.8h, v0.8h\n" + "ld1r { v20.8h }, [x20]\n" + "fmin v8.8h, v8.8h, v21.8h\n" + "fmin v9.8h, v9.8h, v21.8h\n" + "fmin v10.8h, v10.8h, v21.8h\n" + "fmin v11.8h, v11.8h, v21.8h\n" + "fmin v12.8h, v12.8h, v21.8h\n" + "fmin v13.8h, v13.8h, v21.8h\n" + "fmin v14.8h, v14.8h, v21.8h\n" + "fmin v15.8h, v15.8h, v21.8h\n" + "fmin v16.8h, v16.8h, v21.8h\n" + "fmin v17.8h, v17.8h, v21.8h\n" + "fmin v18.8h, v18.8h, v21.8h\n" + "fmin v19.8h, v19.8h, v21.8h\n" + "fmax v8.8h, v8.8h, v20.8h\n" + "fmax v9.8h, v9.8h, v20.8h\n" + "fmax v10.8h, v10.8h, v20.8h\n" + "fmax v11.8h, v11.8h, v20.8h\n" + "fmax v12.8h, v12.8h, v20.8h\n" + "fmax v13.8h, v13.8h, v20.8h\n" + "fmax v14.8h, v14.8h, v20.8h\n" + "fmax v15.8h, v15.8h, v20.8h\n" + "fmax v16.8h, v16.8h, v20.8h\n" + "fmax v17.8h, v17.8h, v20.8h\n" + "fmax v18.8h, v18.8h, v20.8h\n" + "fmax v19.8h, v19.8h, v20.8h\n" "129:" // Height 3: No activation "cmp x11, #0x20\n" "bge 146f\n" @@ -2220,14 +2220,14 @@ void a64_hybrid_fp16_mla_6x32 ( "170:" // 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, 171f\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, 172f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -2237,9 +2237,9 @@ void a64_hybrid_fp16_mla_6x32 ( "b 172f\n" "171:" // Height 4: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" "172:" // Height 4: input setup done "cmp x27, #0x8\n" "blt 175f\n" @@ -2258,7 +2258,7 @@ void a64_hybrid_fp16_mla_6x32 ( "add x26, x26, #0x10\n" "fmla v16.8h, v6.8h, v2.h[0]\n" "fmla v20.8h, v6.8h, v3.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q25, [x10, #0x20]\n" "add x25, x25, #0x10\n" "fmla v9.8h, v7.8h, v0.h[0]\n" "fmla v13.8h, v7.8h, v1.h[0]\n" @@ -2266,165 +2266,165 @@ void a64_hybrid_fp16_mla_6x32 ( "add x23, x23, #0x10\n" "fmla v17.8h, v7.8h, v2.h[0]\n" "fmla v21.8h, v7.8h, v3.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q24, [x10, #0x30]\n" "cmp x27, #0x10\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" + "fmla v10.8h, v25.8h, v0.h[0]\n" + "fmla v14.8h, v25.8h, v1.h[0]\n" "prfm pldl1keep, [x26, #0x80]\n" "prfm pldl1keep, [x25, #0x80]\n" - "fmla v18.8h, v6.8h, v2.h[0]\n" - "fmla v22.8h, v6.8h, v3.h[0]\n" - "ldr q6, [x10, #0x40]\n" + "fmla v18.8h, v25.8h, v2.h[0]\n" + "fmla v22.8h, v25.8h, v3.h[0]\n" + "ldr q25, [x10, #0x40]\n" "prfm pldl1keep, [x24, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" + "fmla v11.8h, v24.8h, v0.h[0]\n" + "fmla v15.8h, v24.8h, v1.h[0]\n" "prfm pldl1keep, [x23, #0x80]\n" - "fmla v19.8h, v7.8h, v2.h[0]\n" - "fmla v23.8h, v7.8h, v3.h[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.8h, v6.8h, v0.h[1]\n" - "fmla v12.8h, v6.8h, v1.h[1]\n" - "fmla v16.8h, v6.8h, v2.h[1]\n" - "fmla v20.8h, v6.8h, v3.h[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.8h, v7.8h, v0.h[1]\n" - "fmla v13.8h, v7.8h, v1.h[1]\n" - "fmla v17.8h, v7.8h, v2.h[1]\n" - "fmla v21.8h, v7.8h, v3.h[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.8h, v6.8h, v0.h[1]\n" - "fmla v14.8h, v6.8h, v1.h[1]\n" - "fmla v18.8h, v6.8h, v2.h[1]\n" - "fmla v22.8h, v6.8h, v3.h[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[1]\n" - "fmla v15.8h, v7.8h, v1.h[1]\n" - "fmla v19.8h, v7.8h, v2.h[1]\n" - "fmla v23.8h, v7.8h, v3.h[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.8h, v6.8h, v0.h[2]\n" - "fmla v12.8h, v6.8h, v1.h[2]\n" - "fmla v16.8h, v6.8h, v2.h[2]\n" - "fmla v20.8h, v6.8h, v3.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.8h, v7.8h, v0.h[2]\n" - "fmla v13.8h, v7.8h, v1.h[2]\n" - "fmla v17.8h, v7.8h, v2.h[2]\n" - "fmla v21.8h, v7.8h, v3.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.8h, v6.8h, v0.h[2]\n" - "fmla v14.8h, v6.8h, v1.h[2]\n" - "fmla v18.8h, v6.8h, v2.h[2]\n" - "fmla v22.8h, v6.8h, v3.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.8h, v7.8h, v0.h[2]\n" - "fmla v15.8h, v7.8h, v1.h[2]\n" - "fmla v19.8h, v7.8h, v2.h[2]\n" - "fmla v23.8h, v7.8h, v3.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.8h, v6.8h, v0.h[3]\n" - "fmla v12.8h, v6.8h, v1.h[3]\n" - "fmla v16.8h, v6.8h, v2.h[3]\n" - "fmla v20.8h, v6.8h, v3.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.8h, v7.8h, v0.h[3]\n" - "fmla v13.8h, v7.8h, v1.h[3]\n" - "fmla v17.8h, v7.8h, v2.h[3]\n" - "fmla v21.8h, v7.8h, v3.h[3]\n" - "ldr q7, [x10, #0xf0]\n" - "fmla v10.8h, v6.8h, v0.h[3]\n" - "fmla v14.8h, v6.8h, v1.h[3]\n" - "fmla v18.8h, v6.8h, v2.h[3]\n" - "fmla v22.8h, v6.8h, v3.h[3]\n" - "ldr q6, [x10, #0x100]\n" - "fmla v11.8h, v7.8h, v0.h[3]\n" - "fmla v15.8h, v7.8h, v1.h[3]\n" - "fmla v19.8h, v7.8h, v2.h[3]\n" - "fmla v23.8h, v7.8h, v3.h[3]\n" - "ldr q7, [x10, #0x110]\n" - "fmla v8.8h, v6.8h, v0.h[4]\n" - "fmla v12.8h, v6.8h, v1.h[4]\n" - "fmla v16.8h, v6.8h, v2.h[4]\n" - "fmla v20.8h, v6.8h, v3.h[4]\n" - "ldr q6, [x10, #0x120]\n" - "fmla v9.8h, v7.8h, v0.h[4]\n" - "fmla v13.8h, v7.8h, v1.h[4]\n" - "fmla v17.8h, v7.8h, v2.h[4]\n" - "fmla v21.8h, v7.8h, v3.h[4]\n" - "ldr q7, [x10, #0x130]\n" - "fmla v10.8h, v6.8h, v0.h[4]\n" - "fmla v14.8h, v6.8h, v1.h[4]\n" - "fmla v18.8h, v6.8h, v2.h[4]\n" - "fmla v22.8h, v6.8h, v3.h[4]\n" - "ldr q6, [x10, #0x140]\n" - "fmla v11.8h, v7.8h, v0.h[4]\n" - "fmla v15.8h, v7.8h, v1.h[4]\n" - "fmla v19.8h, v7.8h, v2.h[4]\n" - "fmla v23.8h, v7.8h, v3.h[4]\n" - "ldr q7, [x10, #0x150]\n" - "fmla v8.8h, v6.8h, v0.h[5]\n" - "fmla v12.8h, v6.8h, v1.h[5]\n" - "fmla v16.8h, v6.8h, v2.h[5]\n" - "fmla v20.8h, v6.8h, v3.h[5]\n" - "ldr q6, [x10, #0x160]\n" - "fmla v9.8h, v7.8h, v0.h[5]\n" - "fmla v13.8h, v7.8h, v1.h[5]\n" - "fmla v17.8h, v7.8h, v2.h[5]\n" - "fmla v21.8h, v7.8h, v3.h[5]\n" - "ldr q7, [x10, #0x170]\n" - "fmla v10.8h, v6.8h, v0.h[5]\n" - "fmla v14.8h, v6.8h, v1.h[5]\n" - "fmla v18.8h, v6.8h, v2.h[5]\n" - "fmla v22.8h, v6.8h, v3.h[5]\n" - "ldr q6, [x10, #0x180]\n" - "fmla v11.8h, v7.8h, v0.h[5]\n" - "fmla v15.8h, v7.8h, v1.h[5]\n" - "fmla v19.8h, v7.8h, v2.h[5]\n" - "fmla v23.8h, v7.8h, v3.h[5]\n" - "ldr q7, [x10, #0x190]\n" - "fmla v8.8h, v6.8h, v0.h[6]\n" - "fmla v12.8h, v6.8h, v1.h[6]\n" - "fmla v16.8h, v6.8h, v2.h[6]\n" - "fmla v20.8h, v6.8h, v3.h[6]\n" - "ldr q6, [x10, #0x1a0]\n" - "fmla v9.8h, v7.8h, v0.h[6]\n" - "fmla v13.8h, v7.8h, v1.h[6]\n" - "fmla v17.8h, v7.8h, v2.h[6]\n" - "fmla v21.8h, v7.8h, v3.h[6]\n" - "ldr q7, [x10, #0x1b0]\n" - "fmla v10.8h, v6.8h, v0.h[6]\n" - "fmla v14.8h, v6.8h, v1.h[6]\n" - "fmla v18.8h, v6.8h, v2.h[6]\n" - "fmla v22.8h, v6.8h, v3.h[6]\n" - "ldr q6, [x10, #0x1c0]\n" - "fmla v11.8h, v7.8h, v0.h[6]\n" - "fmla v15.8h, v7.8h, v1.h[6]\n" - "fmla v19.8h, v7.8h, v2.h[6]\n" - "fmla v23.8h, v7.8h, v3.h[6]\n" - "ldr q7, [x10, #0x1d0]\n" - "fmla v8.8h, v6.8h, v0.h[7]\n" - "fmla v12.8h, v6.8h, v1.h[7]\n" - "fmla v16.8h, v6.8h, v2.h[7]\n" - "fmla v20.8h, v6.8h, v3.h[7]\n" - "ldr q6, [x10, #0x1e0]\n" - "fmla v9.8h, v7.8h, v0.h[7]\n" - "fmla v13.8h, v7.8h, v1.h[7]\n" - "fmla v17.8h, v7.8h, v2.h[7]\n" - "fmla v21.8h, v7.8h, v3.h[7]\n" - "ldr q7, [x10, #0x1f0]\n" + "fmla v19.8h, v24.8h, v2.h[0]\n" + "fmla v23.8h, v24.8h, v3.h[0]\n" + "ldr q24, [x10, #0x50]\n" + "fmla v8.8h, v25.8h, v0.h[1]\n" + "fmla v12.8h, v25.8h, v1.h[1]\n" + "fmla v16.8h, v25.8h, v2.h[1]\n" + "fmla v20.8h, v25.8h, v3.h[1]\n" + "ldr q25, [x10, #0x60]\n" + "fmla v9.8h, v24.8h, v0.h[1]\n" + "fmla v13.8h, v24.8h, v1.h[1]\n" + "fmla v17.8h, v24.8h, v2.h[1]\n" + "fmla v21.8h, v24.8h, v3.h[1]\n" + "ldr q24, [x10, #0x70]\n" + "fmla v10.8h, v25.8h, v0.h[1]\n" + "fmla v14.8h, v25.8h, v1.h[1]\n" + "fmla v18.8h, v25.8h, v2.h[1]\n" + "fmla v22.8h, v25.8h, v3.h[1]\n" + "ldr q25, [x10, #0x80]\n" + "fmla v11.8h, v24.8h, v0.h[1]\n" + "fmla v15.8h, v24.8h, v1.h[1]\n" + "fmla v19.8h, v24.8h, v2.h[1]\n" + "fmla v23.8h, v24.8h, v3.h[1]\n" + "ldr q24, [x10, #0x90]\n" + "fmla v8.8h, v25.8h, v0.h[2]\n" + "fmla v12.8h, v25.8h, v1.h[2]\n" + "fmla v16.8h, v25.8h, v2.h[2]\n" + "fmla v20.8h, v25.8h, v3.h[2]\n" + "ldr q25, [x10, #0xa0]\n" + "fmla v9.8h, v24.8h, v0.h[2]\n" + "fmla v13.8h, v24.8h, v1.h[2]\n" + "fmla v17.8h, v24.8h, v2.h[2]\n" + "fmla v21.8h, v24.8h, v3.h[2]\n" + "ldr q24, [x10, #0xb0]\n" + "fmla v10.8h, v25.8h, v0.h[2]\n" + "fmla v14.8h, v25.8h, v1.h[2]\n" + "fmla v18.8h, v25.8h, v2.h[2]\n" + "fmla v22.8h, v25.8h, v3.h[2]\n" + "ldr q25, [x10, #0xc0]\n" + "fmla v11.8h, v24.8h, v0.h[2]\n" + "fmla v15.8h, v24.8h, v1.h[2]\n" + "fmla v19.8h, v24.8h, v2.h[2]\n" + "fmla v23.8h, v24.8h, v3.h[2]\n" + "ldr q24, [x10, #0xd0]\n" + "fmla v8.8h, v25.8h, v0.h[3]\n" + "fmla v12.8h, v25.8h, v1.h[3]\n" + "fmla v16.8h, v25.8h, v2.h[3]\n" + "fmla v20.8h, v25.8h, v3.h[3]\n" + "ldr q25, [x10, #0xe0]\n" + "fmla v9.8h, v24.8h, v0.h[3]\n" + "fmla v13.8h, v24.8h, v1.h[3]\n" + "fmla v17.8h, v24.8h, v2.h[3]\n" + "fmla v21.8h, v24.8h, v3.h[3]\n" + "ldr q24, [x10, #0xf0]\n" + "fmla v10.8h, v25.8h, v0.h[3]\n" + "fmla v14.8h, v25.8h, v1.h[3]\n" + "fmla v18.8h, v25.8h, v2.h[3]\n" + "fmla v22.8h, v25.8h, v3.h[3]\n" + "ldr q25, [x10, #0x100]\n" + "fmla v11.8h, v24.8h, v0.h[3]\n" + "fmla v15.8h, v24.8h, v1.h[3]\n" + "fmla v19.8h, v24.8h, v2.h[3]\n" + "fmla v23.8h, v24.8h, v3.h[3]\n" + "ldr q24, [x10, #0x110]\n" + "fmla v8.8h, v25.8h, v0.h[4]\n" + "fmla v12.8h, v25.8h, v1.h[4]\n" + "fmla v16.8h, v25.8h, v2.h[4]\n" + "fmla v20.8h, v25.8h, v3.h[4]\n" + "ldr q25, [x10, #0x120]\n" + "fmla v9.8h, v24.8h, v0.h[4]\n" + "fmla v13.8h, v24.8h, v1.h[4]\n" + "fmla v17.8h, v24.8h, v2.h[4]\n" + "fmla v21.8h, v24.8h, v3.h[4]\n" + "ldr q24, [x10, #0x130]\n" + "fmla v10.8h, v25.8h, v0.h[4]\n" + "fmla v14.8h, v25.8h, v1.h[4]\n" + "fmla v18.8h, v25.8h, v2.h[4]\n" + "fmla v22.8h, v25.8h, v3.h[4]\n" + "ldr q25, [x10, #0x140]\n" + "fmla v11.8h, v24.8h, v0.h[4]\n" + "fmla v15.8h, v24.8h, v1.h[4]\n" + "fmla v19.8h, v24.8h, v2.h[4]\n" + "fmla v23.8h, v24.8h, v3.h[4]\n" + "ldr q24, [x10, #0x150]\n" + "fmla v8.8h, v25.8h, v0.h[5]\n" + "fmla v12.8h, v25.8h, v1.h[5]\n" + "fmla v16.8h, v25.8h, v2.h[5]\n" + "fmla v20.8h, v25.8h, v3.h[5]\n" + "ldr q25, [x10, #0x160]\n" + "fmla v9.8h, v24.8h, v0.h[5]\n" + "fmla v13.8h, v24.8h, v1.h[5]\n" + "fmla v17.8h, v24.8h, v2.h[5]\n" + "fmla v21.8h, v24.8h, v3.h[5]\n" + "ldr q24, [x10, #0x170]\n" + "fmla v10.8h, v25.8h, v0.h[5]\n" + "fmla v14.8h, v25.8h, v1.h[5]\n" + "fmla v18.8h, v25.8h, v2.h[5]\n" + "fmla v22.8h, v25.8h, v3.h[5]\n" + "ldr q25, [x10, #0x180]\n" + "fmla v11.8h, v24.8h, v0.h[5]\n" + "fmla v15.8h, v24.8h, v1.h[5]\n" + "fmla v19.8h, v24.8h, v2.h[5]\n" + "fmla v23.8h, v24.8h, v3.h[5]\n" + "ldr q24, [x10, #0x190]\n" + "fmla v8.8h, v25.8h, v0.h[6]\n" + "fmla v12.8h, v25.8h, v1.h[6]\n" + "fmla v16.8h, v25.8h, v2.h[6]\n" + "fmla v20.8h, v25.8h, v3.h[6]\n" + "ldr q25, [x10, #0x1a0]\n" + "fmla v9.8h, v24.8h, v0.h[6]\n" + "fmla v13.8h, v24.8h, v1.h[6]\n" + "fmla v17.8h, v24.8h, v2.h[6]\n" + "fmla v21.8h, v24.8h, v3.h[6]\n" + "ldr q24, [x10, #0x1b0]\n" + "fmla v10.8h, v25.8h, v0.h[6]\n" + "fmla v14.8h, v25.8h, v1.h[6]\n" + "fmla v18.8h, v25.8h, v2.h[6]\n" + "fmla v22.8h, v25.8h, v3.h[6]\n" + "ldr q25, [x10, #0x1c0]\n" + "fmla v11.8h, v24.8h, v0.h[6]\n" + "fmla v15.8h, v24.8h, v1.h[6]\n" + "fmla v19.8h, v24.8h, v2.h[6]\n" + "fmla v23.8h, v24.8h, v3.h[6]\n" + "ldr q24, [x10, #0x1d0]\n" + "fmla v8.8h, v25.8h, v0.h[7]\n" + "fmla v12.8h, v25.8h, v1.h[7]\n" + "fmla v16.8h, v25.8h, v2.h[7]\n" + "fmla v20.8h, v25.8h, v3.h[7]\n" + "ldr q25, [x10, #0x1e0]\n" + "fmla v9.8h, v24.8h, v0.h[7]\n" + "fmla v13.8h, v24.8h, v1.h[7]\n" + "fmla v17.8h, v24.8h, v2.h[7]\n" + "fmla v21.8h, v24.8h, v3.h[7]\n" + "ldr q24, [x10, #0x1f0]\n" "add x10, x10, #0x200\n" - "fmla v10.8h, v6.8h, v0.h[7]\n" - "fmla v14.8h, v6.8h, v1.h[7]\n" - "fmla v18.8h, v6.8h, v2.h[7]\n" - "fmla v22.8h, v6.8h, v3.h[7]\n" + "fmla v10.8h, v25.8h, v0.h[7]\n" + "fmla v14.8h, v25.8h, v1.h[7]\n" + "fmla v18.8h, v25.8h, v2.h[7]\n" + "fmla v22.8h, v25.8h, v3.h[7]\n" "ldr q6, [x10, #0x0]\n" - "fmla v11.8h, v7.8h, v0.h[7]\n" + "fmla v11.8h, v24.8h, v0.h[7]\n" "ldr q0, [x26, #0x0]\n" - "fmla v15.8h, v7.8h, v1.h[7]\n" + "fmla v15.8h, v24.8h, v1.h[7]\n" "ldr q1, [x25, #0x0]\n" - "fmla v19.8h, v7.8h, v2.h[7]\n" + "fmla v19.8h, v24.8h, v2.h[7]\n" "ldr q2, [x24, #0x0]\n" - "fmla v23.8h, v7.8h, v3.h[7]\n" + "fmla v23.8h, v24.8h, v3.h[7]\n" "ldr q3, [x23, #0x0]\n" "ldr q7, [x10, #0x10]\n" "bge 173b\n" @@ -2435,7 +2435,7 @@ void a64_hybrid_fp16_mla_6x32 ( "add x25, x25, #0x10\n" "fmla v16.8h, v6.8h, v2.h[0]\n" "fmla v20.8h, v6.8h, v3.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q25, [x10, #0x20]\n" "add x24, x24, #0x10\n" "fmla v9.8h, v7.8h, v0.h[0]\n" "fmla v13.8h, v7.8h, v1.h[0]\n" @@ -2443,189 +2443,189 @@ void a64_hybrid_fp16_mla_6x32 ( "sub x27, x27, #0x8\n" "fmla v17.8h, v7.8h, v2.h[0]\n" "fmla v21.8h, v7.8h, v3.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q24, [x10, #0x30]\n" "prfm pldl1keep, [x26, #0x80]\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" + "fmla v10.8h, v25.8h, v0.h[0]\n" + "fmla v14.8h, v25.8h, v1.h[0]\n" "prfm pldl1keep, [x25, #0x80]\n" "prfm pldl1keep, [x24, #0x80]\n" - "fmla v18.8h, v6.8h, v2.h[0]\n" - "fmla v22.8h, v6.8h, v3.h[0]\n" - "ldr q6, [x10, #0x40]\n" + "fmla v18.8h, v25.8h, v2.h[0]\n" + "fmla v22.8h, v25.8h, v3.h[0]\n" + "ldr q25, [x10, #0x40]\n" "prfm pldl1keep, [x23, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" - "fmla v19.8h, v7.8h, v2.h[0]\n" - "fmla v23.8h, v7.8h, v3.h[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.8h, v6.8h, v0.h[1]\n" - "fmla v12.8h, v6.8h, v1.h[1]\n" - "fmla v16.8h, v6.8h, v2.h[1]\n" - "fmla v20.8h, v6.8h, v3.h[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.8h, v7.8h, v0.h[1]\n" - "fmla v13.8h, v7.8h, v1.h[1]\n" - "fmla v17.8h, v7.8h, v2.h[1]\n" - "fmla v21.8h, v7.8h, v3.h[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.8h, v6.8h, v0.h[1]\n" - "fmla v14.8h, v6.8h, v1.h[1]\n" - "fmla v18.8h, v6.8h, v2.h[1]\n" - "fmla v22.8h, v6.8h, v3.h[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[1]\n" - "fmla v15.8h, v7.8h, v1.h[1]\n" - "fmla v19.8h, v7.8h, v2.h[1]\n" - "fmla v23.8h, v7.8h, v3.h[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.8h, v6.8h, v0.h[2]\n" - "fmla v12.8h, v6.8h, v1.h[2]\n" - "fmla v16.8h, v6.8h, v2.h[2]\n" - "fmla v20.8h, v6.8h, v3.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.8h, v7.8h, v0.h[2]\n" - "fmla v13.8h, v7.8h, v1.h[2]\n" - "fmla v17.8h, v7.8h, v2.h[2]\n" - "fmla v21.8h, v7.8h, v3.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.8h, v6.8h, v0.h[2]\n" - "fmla v14.8h, v6.8h, v1.h[2]\n" - "fmla v18.8h, v6.8h, v2.h[2]\n" - "fmla v22.8h, v6.8h, v3.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.8h, v7.8h, v0.h[2]\n" - "fmla v15.8h, v7.8h, v1.h[2]\n" - "fmla v19.8h, v7.8h, v2.h[2]\n" - "fmla v23.8h, v7.8h, v3.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.8h, v6.8h, v0.h[3]\n" - "fmla v12.8h, v6.8h, v1.h[3]\n" - "fmla v16.8h, v6.8h, v2.h[3]\n" - "fmla v20.8h, v6.8h, v3.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.8h, v7.8h, v0.h[3]\n" - "fmla v13.8h, v7.8h, v1.h[3]\n" - "fmla v17.8h, v7.8h, v2.h[3]\n" - "fmla v21.8h, v7.8h, v3.h[3]\n" - "ldr q7, [x10, #0xf0]\n" - "fmla v10.8h, v6.8h, v0.h[3]\n" - "fmla v14.8h, v6.8h, v1.h[3]\n" - "fmla v18.8h, v6.8h, v2.h[3]\n" - "fmla v22.8h, v6.8h, v3.h[3]\n" - "ldr q6, [x10, #0x100]\n" - "fmla v11.8h, v7.8h, v0.h[3]\n" - "fmla v15.8h, v7.8h, v1.h[3]\n" - "fmla v19.8h, v7.8h, v2.h[3]\n" - "fmla v23.8h, v7.8h, v3.h[3]\n" - "ldr q7, [x10, #0x110]\n" - "fmla v8.8h, v6.8h, v0.h[4]\n" - "fmla v12.8h, v6.8h, v1.h[4]\n" - "fmla v16.8h, v6.8h, v2.h[4]\n" - "fmla v20.8h, v6.8h, v3.h[4]\n" - "ldr q6, [x10, #0x120]\n" - "fmla v9.8h, v7.8h, v0.h[4]\n" - "fmla v13.8h, v7.8h, v1.h[4]\n" - "fmla v17.8h, v7.8h, v2.h[4]\n" - "fmla v21.8h, v7.8h, v3.h[4]\n" - "ldr q7, [x10, #0x130]\n" - "fmla v10.8h, v6.8h, v0.h[4]\n" - "fmla v14.8h, v6.8h, v1.h[4]\n" - "fmla v18.8h, v6.8h, v2.h[4]\n" - "fmla v22.8h, v6.8h, v3.h[4]\n" - "ldr q6, [x10, #0x140]\n" - "fmla v11.8h, v7.8h, v0.h[4]\n" - "fmla v15.8h, v7.8h, v1.h[4]\n" - "fmla v19.8h, v7.8h, v2.h[4]\n" - "fmla v23.8h, v7.8h, v3.h[4]\n" - "ldr q7, [x10, #0x150]\n" - "fmla v8.8h, v6.8h, v0.h[5]\n" - "fmla v12.8h, v6.8h, v1.h[5]\n" - "fmla v16.8h, v6.8h, v2.h[5]\n" - "fmla v20.8h, v6.8h, v3.h[5]\n" - "ldr q6, [x10, #0x160]\n" - "fmla v9.8h, v7.8h, v0.h[5]\n" - "fmla v13.8h, v7.8h, v1.h[5]\n" - "fmla v17.8h, v7.8h, v2.h[5]\n" - "fmla v21.8h, v7.8h, v3.h[5]\n" - "ldr q7, [x10, #0x170]\n" - "fmla v10.8h, v6.8h, v0.h[5]\n" - "fmla v14.8h, v6.8h, v1.h[5]\n" - "fmla v18.8h, v6.8h, v2.h[5]\n" - "fmla v22.8h, v6.8h, v3.h[5]\n" - "ldr q6, [x10, #0x180]\n" - "fmla v11.8h, v7.8h, v0.h[5]\n" - "fmla v15.8h, v7.8h, v1.h[5]\n" - "fmla v19.8h, v7.8h, v2.h[5]\n" - "fmla v23.8h, v7.8h, v3.h[5]\n" - "ldr q7, [x10, #0x190]\n" - "fmla v8.8h, v6.8h, v0.h[6]\n" - "fmla v12.8h, v6.8h, v1.h[6]\n" - "fmla v16.8h, v6.8h, v2.h[6]\n" - "fmla v20.8h, v6.8h, v3.h[6]\n" - "ldr q6, [x10, #0x1a0]\n" - "fmla v9.8h, v7.8h, v0.h[6]\n" - "fmla v13.8h, v7.8h, v1.h[6]\n" - "fmla v17.8h, v7.8h, v2.h[6]\n" - "fmla v21.8h, v7.8h, v3.h[6]\n" - "ldr q7, [x10, #0x1b0]\n" - "fmla v10.8h, v6.8h, v0.h[6]\n" - "fmla v14.8h, v6.8h, v1.h[6]\n" - "fmla v18.8h, v6.8h, v2.h[6]\n" - "fmla v22.8h, v6.8h, v3.h[6]\n" - "ldr q6, [x10, #0x1c0]\n" - "fmla v11.8h, v7.8h, v0.h[6]\n" - "fmla v15.8h, v7.8h, v1.h[6]\n" - "fmla v19.8h, v7.8h, v2.h[6]\n" - "fmla v23.8h, v7.8h, v3.h[6]\n" - "ldr q7, [x10, #0x1d0]\n" - "fmla v8.8h, v6.8h, v0.h[7]\n" - "fmla v12.8h, v6.8h, v1.h[7]\n" - "fmla v16.8h, v6.8h, v2.h[7]\n" - "fmla v20.8h, v6.8h, v3.h[7]\n" - "ldr q6, [x10, #0x1e0]\n" - "fmla v9.8h, v7.8h, v0.h[7]\n" - "fmla v13.8h, v7.8h, v1.h[7]\n" - "fmla v17.8h, v7.8h, v2.h[7]\n" - "fmla v21.8h, v7.8h, v3.h[7]\n" - "ldr q7, [x10, #0x1f0]\n" + "fmla v11.8h, v24.8h, v0.h[0]\n" + "fmla v15.8h, v24.8h, v1.h[0]\n" + "fmla v19.8h, v24.8h, v2.h[0]\n" + "fmla v23.8h, v24.8h, v3.h[0]\n" + "ldr q24, [x10, #0x50]\n" + "fmla v8.8h, v25.8h, v0.h[1]\n" + "fmla v12.8h, v25.8h, v1.h[1]\n" + "fmla v16.8h, v25.8h, v2.h[1]\n" + "fmla v20.8h, v25.8h, v3.h[1]\n" + "ldr q25, [x10, #0x60]\n" + "fmla v9.8h, v24.8h, v0.h[1]\n" + "fmla v13.8h, v24.8h, v1.h[1]\n" + "fmla v17.8h, v24.8h, v2.h[1]\n" + "fmla v21.8h, v24.8h, v3.h[1]\n" + "ldr q24, [x10, #0x70]\n" + "fmla v10.8h, v25.8h, v0.h[1]\n" + "fmla v14.8h, v25.8h, v1.h[1]\n" + "fmla v18.8h, v25.8h, v2.h[1]\n" + "fmla v22.8h, v25.8h, v3.h[1]\n" + "ldr q25, [x10, #0x80]\n" + "fmla v11.8h, v24.8h, v0.h[1]\n" + "fmla v15.8h, v24.8h, v1.h[1]\n" + "fmla v19.8h, v24.8h, v2.h[1]\n" + "fmla v23.8h, v24.8h, v3.h[1]\n" + "ldr q24, [x10, #0x90]\n" + "fmla v8.8h, v25.8h, v0.h[2]\n" + "fmla v12.8h, v25.8h, v1.h[2]\n" + "fmla v16.8h, v25.8h, v2.h[2]\n" + "fmla v20.8h, v25.8h, v3.h[2]\n" + "ldr q25, [x10, #0xa0]\n" + "fmla v9.8h, v24.8h, v0.h[2]\n" + "fmla v13.8h, v24.8h, v1.h[2]\n" + "fmla v17.8h, v24.8h, v2.h[2]\n" + "fmla v21.8h, v24.8h, v3.h[2]\n" + "ldr q24, [x10, #0xb0]\n" + "fmla v10.8h, v25.8h, v0.h[2]\n" + "fmla v14.8h, v25.8h, v1.h[2]\n" + "fmla v18.8h, v25.8h, v2.h[2]\n" + "fmla v22.8h, v25.8h, v3.h[2]\n" + "ldr q25, [x10, #0xc0]\n" + "fmla v11.8h, v24.8h, v0.h[2]\n" + "fmla v15.8h, v24.8h, v1.h[2]\n" + "fmla v19.8h, v24.8h, v2.h[2]\n" + "fmla v23.8h, v24.8h, v3.h[2]\n" + "ldr q24, [x10, #0xd0]\n" + "fmla v8.8h, v25.8h, v0.h[3]\n" + "fmla v12.8h, v25.8h, v1.h[3]\n" + "fmla v16.8h, v25.8h, v2.h[3]\n" + "fmla v20.8h, v25.8h, v3.h[3]\n" + "ldr q25, [x10, #0xe0]\n" + "fmla v9.8h, v24.8h, v0.h[3]\n" + "fmla v13.8h, v24.8h, v1.h[3]\n" + "fmla v17.8h, v24.8h, v2.h[3]\n" + "fmla v21.8h, v24.8h, v3.h[3]\n" + "ldr q24, [x10, #0xf0]\n" + "fmla v10.8h, v25.8h, v0.h[3]\n" + "fmla v14.8h, v25.8h, v1.h[3]\n" + "fmla v18.8h, v25.8h, v2.h[3]\n" + "fmla v22.8h, v25.8h, v3.h[3]\n" + "ldr q25, [x10, #0x100]\n" + "fmla v11.8h, v24.8h, v0.h[3]\n" + "fmla v15.8h, v24.8h, v1.h[3]\n" + "fmla v19.8h, v24.8h, v2.h[3]\n" + "fmla v23.8h, v24.8h, v3.h[3]\n" + "ldr q24, [x10, #0x110]\n" + "fmla v8.8h, v25.8h, v0.h[4]\n" + "fmla v12.8h, v25.8h, v1.h[4]\n" + "fmla v16.8h, v25.8h, v2.h[4]\n" + "fmla v20.8h, v25.8h, v3.h[4]\n" + "ldr q25, [x10, #0x120]\n" + "fmla v9.8h, v24.8h, v0.h[4]\n" + "fmla v13.8h, v24.8h, v1.h[4]\n" + "fmla v17.8h, v24.8h, v2.h[4]\n" + "fmla v21.8h, v24.8h, v3.h[4]\n" + "ldr q24, [x10, #0x130]\n" + "fmla v10.8h, v25.8h, v0.h[4]\n" + "fmla v14.8h, v25.8h, v1.h[4]\n" + "fmla v18.8h, v25.8h, v2.h[4]\n" + "fmla v22.8h, v25.8h, v3.h[4]\n" + "ldr q25, [x10, #0x140]\n" + "fmla v11.8h, v24.8h, v0.h[4]\n" + "fmla v15.8h, v24.8h, v1.h[4]\n" + "fmla v19.8h, v24.8h, v2.h[4]\n" + "fmla v23.8h, v24.8h, v3.h[4]\n" + "ldr q24, [x10, #0x150]\n" + "fmla v8.8h, v25.8h, v0.h[5]\n" + "fmla v12.8h, v25.8h, v1.h[5]\n" + "fmla v16.8h, v25.8h, v2.h[5]\n" + "fmla v20.8h, v25.8h, v3.h[5]\n" + "ldr q25, [x10, #0x160]\n" + "fmla v9.8h, v24.8h, v0.h[5]\n" + "fmla v13.8h, v24.8h, v1.h[5]\n" + "fmla v17.8h, v24.8h, v2.h[5]\n" + "fmla v21.8h, v24.8h, v3.h[5]\n" + "ldr q24, [x10, #0x170]\n" + "fmla v10.8h, v25.8h, v0.h[5]\n" + "fmla v14.8h, v25.8h, v1.h[5]\n" + "fmla v18.8h, v25.8h, v2.h[5]\n" + "fmla v22.8h, v25.8h, v3.h[5]\n" + "ldr q25, [x10, #0x180]\n" + "fmla v11.8h, v24.8h, v0.h[5]\n" + "fmla v15.8h, v24.8h, v1.h[5]\n" + "fmla v19.8h, v24.8h, v2.h[5]\n" + "fmla v23.8h, v24.8h, v3.h[5]\n" + "ldr q24, [x10, #0x190]\n" + "fmla v8.8h, v25.8h, v0.h[6]\n" + "fmla v12.8h, v25.8h, v1.h[6]\n" + "fmla v16.8h, v25.8h, v2.h[6]\n" + "fmla v20.8h, v25.8h, v3.h[6]\n" + "ldr q25, [x10, #0x1a0]\n" + "fmla v9.8h, v24.8h, v0.h[6]\n" + "fmla v13.8h, v24.8h, v1.h[6]\n" + "fmla v17.8h, v24.8h, v2.h[6]\n" + "fmla v21.8h, v24.8h, v3.h[6]\n" + "ldr q24, [x10, #0x1b0]\n" + "fmla v10.8h, v25.8h, v0.h[6]\n" + "fmla v14.8h, v25.8h, v1.h[6]\n" + "fmla v18.8h, v25.8h, v2.h[6]\n" + "fmla v22.8h, v25.8h, v3.h[6]\n" + "ldr q25, [x10, #0x1c0]\n" + "fmla v11.8h, v24.8h, v0.h[6]\n" + "fmla v15.8h, v24.8h, v1.h[6]\n" + "fmla v19.8h, v24.8h, v2.h[6]\n" + "fmla v23.8h, v24.8h, v3.h[6]\n" + "ldr q24, [x10, #0x1d0]\n" + "fmla v8.8h, v25.8h, v0.h[7]\n" + "fmla v12.8h, v25.8h, v1.h[7]\n" + "fmla v16.8h, v25.8h, v2.h[7]\n" + "fmla v20.8h, v25.8h, v3.h[7]\n" + "ldr q25, [x10, #0x1e0]\n" + "fmla v9.8h, v24.8h, v0.h[7]\n" + "fmla v13.8h, v24.8h, v1.h[7]\n" + "fmla v17.8h, v24.8h, v2.h[7]\n" + "fmla v21.8h, v24.8h, v3.h[7]\n" + "ldr q24, [x10, #0x1f0]\n" "add x10, x10, #0x200\n" - "fmla v10.8h, v6.8h, v0.h[7]\n" - "fmla v14.8h, v6.8h, v1.h[7]\n" - "fmla v18.8h, v6.8h, v2.h[7]\n" - "fmla v22.8h, v6.8h, v3.h[7]\n" - "fmla v11.8h, v7.8h, v0.h[7]\n" - "fmla v15.8h, v7.8h, v1.h[7]\n" - "fmla v19.8h, v7.8h, v2.h[7]\n" - "fmla v23.8h, v7.8h, v3.h[7]\n" + "fmla v10.8h, v25.8h, v0.h[7]\n" + "fmla v14.8h, v25.8h, v1.h[7]\n" + "fmla v18.8h, v25.8h, v2.h[7]\n" + "fmla v22.8h, v25.8h, v3.h[7]\n" + "fmla v11.8h, v24.8h, v0.h[7]\n" + "fmla v15.8h, v24.8h, v1.h[7]\n" + "fmla v19.8h, v24.8h, v2.h[7]\n" + "fmla v23.8h, v24.8h, v3.h[7]\n" "175:" // Height 4: Multiply loop: Main loop skip "cbz x27, 177f\n" "176:" // Height 4: Multiply loop: Odd block loop - "ldr h0, [x26], #0x2\n" - "ldr h1, [x25], #0x2\n" + "ldr h3, [x26], #0x2\n" + "ldr h2, [x25], #0x2\n" "sub x27, x27, #0x1\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - "fmla v8.8h, v6.8h, v0.h[0]\n" - "fmla v12.8h, v6.8h, v1.h[0]\n" - "fmla v16.8h, v6.8h, v2.h[0]\n" - "fmla v20.8h, v6.8h, v3.h[0]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.8h, v7.8h, v0.h[0]\n" - "fmla v13.8h, v7.8h, v1.h[0]\n" - "fmla v17.8h, v7.8h, v2.h[0]\n" - "fmla v21.8h, v7.8h, v3.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr h1, [x24], #0x2\n" + "ldr h0, [x23], #0x2\n" + "ldr q25, [x10, #0x0]\n" + "ldr q24, [x10, #0x10]\n" + "fmla v8.8h, v25.8h, v3.h[0]\n" + "fmla v12.8h, v25.8h, v2.h[0]\n" + "fmla v16.8h, v25.8h, v1.h[0]\n" + "fmla v20.8h, v25.8h, v0.h[0]\n" + "ldr q25, [x10, #0x20]\n" + "fmla v9.8h, v24.8h, v3.h[0]\n" + "fmla v13.8h, v24.8h, v2.h[0]\n" + "fmla v17.8h, v24.8h, v1.h[0]\n" + "fmla v21.8h, v24.8h, v0.h[0]\n" + "ldr q24, [x10, #0x30]\n" "add x10, x10, #0x40\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" - "fmla v18.8h, v6.8h, v2.h[0]\n" - "fmla v22.8h, v6.8h, v3.h[0]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" - "fmla v19.8h, v7.8h, v2.h[0]\n" - "fmla v23.8h, v7.8h, v3.h[0]\n" + "fmla v10.8h, v25.8h, v3.h[0]\n" + "fmla v14.8h, v25.8h, v2.h[0]\n" + "fmla v18.8h, v25.8h, v1.h[0]\n" + "fmla v22.8h, v25.8h, v0.h[0]\n" + "fmla v11.8h, v24.8h, v3.h[0]\n" + "fmla v15.8h, v24.8h, v2.h[0]\n" + "fmla v19.8h, v24.8h, v1.h[0]\n" + "fmla v23.8h, v24.8h, v0.h[0]\n" "cbnz x27, 176b\n" "177:" // Height 4: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -2642,41 +2642,41 @@ void a64_hybrid_fp16_mla_6x32 ( "prfm pstl1keep, [x23, #0x0]\n" "tbz %x[flags], #1, 178f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.8h }, [x20]\n" + "ld1r { v25.8h }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.8h }, [x20]\n" - "fmin v8.8h, v8.8h, v1.8h\n" - "fmin v9.8h, v9.8h, v1.8h\n" - "fmin v10.8h, v10.8h, v1.8h\n" - "fmin v11.8h, v11.8h, v1.8h\n" - "fmin v12.8h, v12.8h, v1.8h\n" - "fmin v13.8h, v13.8h, v1.8h\n" - "fmin v14.8h, v14.8h, v1.8h\n" - "fmin v15.8h, v15.8h, v1.8h\n" - "fmin v16.8h, v16.8h, v1.8h\n" - "fmin v17.8h, v17.8h, v1.8h\n" - "fmin v18.8h, v18.8h, v1.8h\n" - "fmin v19.8h, v19.8h, v1.8h\n" - "fmin v20.8h, v20.8h, v1.8h\n" - "fmin v21.8h, v21.8h, v1.8h\n" - "fmin v22.8h, v22.8h, v1.8h\n" - "fmin v23.8h, v23.8h, v1.8h\n" - "fmax v8.8h, v8.8h, v0.8h\n" - "fmax v9.8h, v9.8h, v0.8h\n" - "fmax v10.8h, v10.8h, v0.8h\n" - "fmax v11.8h, v11.8h, v0.8h\n" - "fmax v12.8h, v12.8h, v0.8h\n" - "fmax v13.8h, v13.8h, v0.8h\n" - "fmax v14.8h, v14.8h, v0.8h\n" - "fmax v15.8h, v15.8h, v0.8h\n" - "fmax v16.8h, v16.8h, v0.8h\n" - "fmax v17.8h, v17.8h, v0.8h\n" - "fmax v18.8h, v18.8h, v0.8h\n" - "fmax v19.8h, v19.8h, v0.8h\n" - "fmax v20.8h, v20.8h, v0.8h\n" - "fmax v21.8h, v21.8h, v0.8h\n" - "fmax v22.8h, v22.8h, v0.8h\n" - "fmax v23.8h, v23.8h, v0.8h\n" + "ld1r { v24.8h }, [x20]\n" + "fmin v8.8h, v8.8h, v25.8h\n" + "fmin v9.8h, v9.8h, v25.8h\n" + "fmin v10.8h, v10.8h, v25.8h\n" + "fmin v11.8h, v11.8h, v25.8h\n" + "fmin v12.8h, v12.8h, v25.8h\n" + "fmin v13.8h, v13.8h, v25.8h\n" + "fmin v14.8h, v14.8h, v25.8h\n" + "fmin v15.8h, v15.8h, v25.8h\n" + "fmin v16.8h, v16.8h, v25.8h\n" + "fmin v17.8h, v17.8h, v25.8h\n" + "fmin v18.8h, v18.8h, v25.8h\n" + "fmin v19.8h, v19.8h, v25.8h\n" + "fmin v20.8h, v20.8h, v25.8h\n" + "fmin v21.8h, v21.8h, v25.8h\n" + "fmin v22.8h, v22.8h, v25.8h\n" + "fmin v23.8h, v23.8h, v25.8h\n" + "fmax v8.8h, v8.8h, v24.8h\n" + "fmax v9.8h, v9.8h, v24.8h\n" + "fmax v10.8h, v10.8h, v24.8h\n" + "fmax v11.8h, v11.8h, v24.8h\n" + "fmax v12.8h, v12.8h, v24.8h\n" + "fmax v13.8h, v13.8h, v24.8h\n" + "fmax v14.8h, v14.8h, v24.8h\n" + "fmax v15.8h, v15.8h, v24.8h\n" + "fmax v16.8h, v16.8h, v24.8h\n" + "fmax v17.8h, v17.8h, v24.8h\n" + "fmax v18.8h, v18.8h, v24.8h\n" + "fmax v19.8h, v19.8h, v24.8h\n" + "fmax v20.8h, v20.8h, v24.8h\n" + "fmax v21.8h, v21.8h, v24.8h\n" + "fmax v22.8h, v22.8h, v24.8h\n" + "fmax v23.8h, v23.8h, v24.8h\n" "178:" // Height 4: No activation "cmp x11, #0x20\n" "bge 195f\n" @@ -3124,583 +3124,583 @@ void a64_hybrid_fp16_mla_6x32 ( "tbz x11, #1, 212f\n" "ld1 { v8.s }[2], [x9], #0x4\n" "ld1 { v12.s }[2], [x25], #0x4\n" - "mov x20, #0xc\n" - "ld1 { v16.s }[2], [x24], #0x4\n" - "ld1 { v20.s }[2], [x23], #0x4\n" - "ld1 { v24.s }[2], [x22], #0x4\n" - "tbz x11, #0, 215f\n" - "ld1 { v8.h }[6], [x9]\n" - "ld1 { v12.h }[6], [x25]\n" - "ld1 { v16.h }[6], [x24]\n" - "ld1 { v20.h }[6], [x23]\n" - "ld1 { v24.h }[6], [x22]\n" - "b 215f\n" - "212:" // Height 5: Partial accumulate: partial_1_4 - "mov x20, #0x8\n" - "tbz x11, #0, 215f\n" - "ld1 { v8.h }[4], [x9]\n" - "ld1 { v12.h }[4], [x25]\n" - "ld1 { v16.h }[4], [x24]\n" - "ld1 { v20.h }[4], [x23]\n" - "ld1 { v24.h }[4], [x22]\n" - "b 215f\n" - "213:" // Height 5: Partial accumulate: partial_2_0 - "tbz x11, #1, 214f\n" - "ldr s8, [x9], #0x4\n" - "ldr s12, [x25], #0x4\n" - "mov x20, #0x4\n" - "ldr s16, [x24], #0x4\n" - "ldr s20, [x23], #0x4\n" - "ldr s24, [x22], #0x4\n" - "tbz x11, #0, 215f\n" - "ld1 { v8.h }[2], [x9]\n" - "ld1 { v12.h }[2], [x25]\n" - "ld1 { v16.h }[2], [x24]\n" - "ld1 { v20.h }[2], [x23]\n" - "ld1 { v24.h }[2], [x22]\n" - "b 215f\n" - "214:" // Height 5: Partial accumulate: partial_1_0 - "ldr h8, [x9, #0x0]\n" - "ldr h12, [x25, #0x0]\n" - "mov x20, #0x0\n" - "ldr h16, [x24, #0x0]\n" - "ldr h20, [x23, #0x0]\n" - "ldr h24, [x22, #0x0]\n" - "215:" // Height 5: Partial accumulate: Done - "sub x9, x9, x20\n" - "b 218f\n" - "216:" // Height 5: full accumulate - "ldr q8, [x9, #0x0]\n" - "ldr q9, [x9, #0x10]\n" - "ldr q10, [x9, #0x20]\n" - "ldr q11, [x9, #0x30]\n" - "ldr q12, [x25, #0x0]\n" - "ldr q13, [x25, #0x10]\n" - "ldr q14, [x25, #0x20]\n" - "ldr q15, [x25, #0x30]\n" - "ldr q16, [x24, #0x0]\n" - "ldr q17, [x24, #0x10]\n" - "ldr q18, [x24, #0x20]\n" - "ldr q19, [x24, #0x30]\n" - "ldr q20, [x23, #0x0]\n" - "ldr q21, [x23, #0x10]\n" - "ldr q22, [x23, #0x20]\n" - "ldr q23, [x23, #0x30]\n" - "ldr q24, [x22, #0x0]\n" - "ldr q25, [x22, #0x10]\n" - "ldr q26, [x22, #0x20]\n" - "ldr q27, [x22, #0x30]\n" - "b 218f\n" - "217:" // Height 5: no accumulate - "movi v8.16b, #0x0\n" - "movi v9.16b, #0x0\n" - "movi v10.16b, #0x0\n" - "movi v11.16b, #0x0\n" - "movi v12.16b, #0x0\n" - "movi v13.16b, #0x0\n" - "movi v14.16b, #0x0\n" - "movi v15.16b, #0x0\n" - "movi v16.16b, #0x0\n" - "movi v17.16b, #0x0\n" - "movi v18.16b, #0x0\n" - "movi v19.16b, #0x0\n" - "movi v20.16b, #0x0\n" - "movi v21.16b, #0x0\n" - "movi v22.16b, #0x0\n" - "movi v23.16b, #0x0\n" - "movi v24.16b, #0x0\n" - "movi v25.16b, #0x0\n" - "movi v26.16b, #0x0\n" - "movi v27.16b, #0x0\n" - "218:" // Height 5: setup done - "mov x28, #0x0\n" - "219:" // Height 5: String loop - "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" - "ldr w27, [x20, x28, LSL #0x2]\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n" - "tbz %x[flags], #3, 220f\n" - "ldr x21, [%x[input_ptr], x28, LSL #0x3]\n" - "add x21, x21, x20, LSL #3\n" - "ldr x26, [x21, #0x0]\n" - "ldr x25, [x21, #0x8]\n" - "ldr x24, [x21, #0x10]\n" - "ldr x23, [x21, #0x18]\n" - "ldr x22, [x21, #0x20]\n" - "cbnz x28, 221f\n" - "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" - "add x26, x26, x20, LSL #1\n" - "add x25, x25, x20, LSL #1\n" - "add x24, x24, x20, LSL #1\n" - "add x23, x23, x20, LSL #1\n" - "add x22, x22, x20, LSL #1\n" - "b 221f\n" - "220:" // Height 5: setup direct input - "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" - "add x22, x23, x20, LSL #1\n" - "221:" // Height 5: input setup done - "cmp x27, #0x8\n" - "blt 224f\n" - "ldr q0, [x26, #0x0]\n" - "ldr q1, [x25, #0x0]\n" - "cmp x27, #0x10\n" - "ldr q2, [x24, #0x0]\n" - "ldr q3, [x23, #0x0]\n" - "ldr q4, [x22, #0x0]\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - "blt 223f\n" - "222:" // Height 5: Multiply loop: Main loop head - "fmla v8.8h, v6.8h, v0.h[0]\n" - "fmla v12.8h, v6.8h, v1.h[0]\n" - "sub x27, x27, #0x8\n" - "add x26, x26, #0x10\n" - "fmla v16.8h, v6.8h, v2.h[0]\n" - "fmla v20.8h, v6.8h, v3.h[0]\n" - "add x25, x25, #0x10\n" - "add x24, x24, #0x10\n" - "fmla v24.8h, v6.8h, v4.h[0]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.8h, v7.8h, v0.h[0]\n" - "add x23, x23, #0x10\n" - "fmla v13.8h, v7.8h, v1.h[0]\n" - "fmla v17.8h, v7.8h, v2.h[0]\n" - "add x22, x22, #0x10\n" - "cmp x27, #0x10\n" - "fmla v21.8h, v7.8h, v3.h[0]\n" - "fmla v25.8h, v7.8h, v4.h[0]\n" - "ldr q7, [x10, #0x30]\n" - "prfm pldl1keep, [x26, #0x80]\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" - "prfm pldl1keep, [x25, #0x80]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "fmla v18.8h, v6.8h, v2.h[0]\n" - "fmla v22.8h, v6.8h, v3.h[0]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "fmla v26.8h, v6.8h, v4.h[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" - "fmla v19.8h, v7.8h, v2.h[0]\n" - "fmla v23.8h, v7.8h, v3.h[0]\n" - "fmla v27.8h, v7.8h, v4.h[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.8h, v6.8h, v0.h[1]\n" - "fmla v12.8h, v6.8h, v1.h[1]\n" - "fmla v16.8h, v6.8h, v2.h[1]\n" - "fmla v20.8h, v6.8h, v3.h[1]\n" - "fmla v24.8h, v6.8h, v4.h[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.8h, v7.8h, v0.h[1]\n" - "fmla v13.8h, v7.8h, v1.h[1]\n" - "fmla v17.8h, v7.8h, v2.h[1]\n" - "fmla v21.8h, v7.8h, v3.h[1]\n" - "fmla v25.8h, v7.8h, v4.h[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.8h, v6.8h, v0.h[1]\n" - "fmla v14.8h, v6.8h, v1.h[1]\n" - "fmla v18.8h, v6.8h, v2.h[1]\n" - "fmla v22.8h, v6.8h, v3.h[1]\n" - "fmla v26.8h, v6.8h, v4.h[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[1]\n" - "fmla v15.8h, v7.8h, v1.h[1]\n" - "fmla v19.8h, v7.8h, v2.h[1]\n" - "fmla v23.8h, v7.8h, v3.h[1]\n" - "fmla v27.8h, v7.8h, v4.h[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.8h, v6.8h, v0.h[2]\n" - "fmla v12.8h, v6.8h, v1.h[2]\n" - "fmla v16.8h, v6.8h, v2.h[2]\n" - "fmla v20.8h, v6.8h, v3.h[2]\n" - "fmla v24.8h, v6.8h, v4.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.8h, v7.8h, v0.h[2]\n" - "fmla v13.8h, v7.8h, v1.h[2]\n" - "fmla v17.8h, v7.8h, v2.h[2]\n" - "fmla v21.8h, v7.8h, v3.h[2]\n" - "fmla v25.8h, v7.8h, v4.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.8h, v6.8h, v0.h[2]\n" - "fmla v14.8h, v6.8h, v1.h[2]\n" - "fmla v18.8h, v6.8h, v2.h[2]\n" - "fmla v22.8h, v6.8h, v3.h[2]\n" - "fmla v26.8h, v6.8h, v4.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.8h, v7.8h, v0.h[2]\n" - "fmla v15.8h, v7.8h, v1.h[2]\n" - "fmla v19.8h, v7.8h, v2.h[2]\n" - "fmla v23.8h, v7.8h, v3.h[2]\n" - "fmla v27.8h, v7.8h, v4.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.8h, v6.8h, v0.h[3]\n" - "fmla v12.8h, v6.8h, v1.h[3]\n" - "fmla v16.8h, v6.8h, v2.h[3]\n" - "fmla v20.8h, v6.8h, v3.h[3]\n" - "fmla v24.8h, v6.8h, v4.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.8h, v7.8h, v0.h[3]\n" - "fmla v13.8h, v7.8h, v1.h[3]\n" - "fmla v17.8h, v7.8h, v2.h[3]\n" - "fmla v21.8h, v7.8h, v3.h[3]\n" - "fmla v25.8h, v7.8h, v4.h[3]\n" - "ldr q7, [x10, #0xf0]\n" - "fmla v10.8h, v6.8h, v0.h[3]\n" - "fmla v14.8h, v6.8h, v1.h[3]\n" - "fmla v18.8h, v6.8h, v2.h[3]\n" - "fmla v22.8h, v6.8h, v3.h[3]\n" - "fmla v26.8h, v6.8h, v4.h[3]\n" - "ldr q6, [x10, #0x100]\n" - "fmla v11.8h, v7.8h, v0.h[3]\n" - "fmla v15.8h, v7.8h, v1.h[3]\n" - "fmla v19.8h, v7.8h, v2.h[3]\n" - "fmla v23.8h, v7.8h, v3.h[3]\n" - "fmla v27.8h, v7.8h, v4.h[3]\n" - "ldr q7, [x10, #0x110]\n" - "fmla v8.8h, v6.8h, v0.h[4]\n" - "fmla v12.8h, v6.8h, v1.h[4]\n" - "fmla v16.8h, v6.8h, v2.h[4]\n" - "fmla v20.8h, v6.8h, v3.h[4]\n" - "fmla v24.8h, v6.8h, v4.h[4]\n" - "ldr q6, [x10, #0x120]\n" - "fmla v9.8h, v7.8h, v0.h[4]\n" - "fmla v13.8h, v7.8h, v1.h[4]\n" - "fmla v17.8h, v7.8h, v2.h[4]\n" - "fmla v21.8h, v7.8h, v3.h[4]\n" - "fmla v25.8h, v7.8h, v4.h[4]\n" - "ldr q7, [x10, #0x130]\n" - "fmla v10.8h, v6.8h, v0.h[4]\n" - "fmla v14.8h, v6.8h, v1.h[4]\n" - "fmla v18.8h, v6.8h, v2.h[4]\n" - "fmla v22.8h, v6.8h, v3.h[4]\n" - "fmla v26.8h, v6.8h, v4.h[4]\n" - "ldr q6, [x10, #0x140]\n" - "fmla v11.8h, v7.8h, v0.h[4]\n" - "fmla v15.8h, v7.8h, v1.h[4]\n" - "fmla v19.8h, v7.8h, v2.h[4]\n" - "fmla v23.8h, v7.8h, v3.h[4]\n" - "fmla v27.8h, v7.8h, v4.h[4]\n" - "ldr q7, [x10, #0x150]\n" - "fmla v8.8h, v6.8h, v0.h[5]\n" - "fmla v12.8h, v6.8h, v1.h[5]\n" - "fmla v16.8h, v6.8h, v2.h[5]\n" - "fmla v20.8h, v6.8h, v3.h[5]\n" - "fmla v24.8h, v6.8h, v4.h[5]\n" - "ldr q6, [x10, #0x160]\n" - "fmla v9.8h, v7.8h, v0.h[5]\n" - "fmla v13.8h, v7.8h, v1.h[5]\n" - "fmla v17.8h, v7.8h, v2.h[5]\n" - "fmla v21.8h, v7.8h, v3.h[5]\n" - "fmla v25.8h, v7.8h, v4.h[5]\n" - "ldr q7, [x10, #0x170]\n" - "fmla v10.8h, v6.8h, v0.h[5]\n" - "fmla v14.8h, v6.8h, v1.h[5]\n" - "fmla v18.8h, v6.8h, v2.h[5]\n" - "fmla v22.8h, v6.8h, v3.h[5]\n" - "fmla v26.8h, v6.8h, v4.h[5]\n" - "ldr q6, [x10, #0x180]\n" - "fmla v11.8h, v7.8h, v0.h[5]\n" - "fmla v15.8h, v7.8h, v1.h[5]\n" - "fmla v19.8h, v7.8h, v2.h[5]\n" - "fmla v23.8h, v7.8h, v3.h[5]\n" - "fmla v27.8h, v7.8h, v4.h[5]\n" - "ldr q7, [x10, #0x190]\n" - "fmla v8.8h, v6.8h, v0.h[6]\n" - "fmla v12.8h, v6.8h, v1.h[6]\n" - "fmla v16.8h, v6.8h, v2.h[6]\n" - "fmla v20.8h, v6.8h, v3.h[6]\n" - "fmla v24.8h, v6.8h, v4.h[6]\n" - "ldr q6, [x10, #0x1a0]\n" - "fmla v9.8h, v7.8h, v0.h[6]\n" - "fmla v13.8h, v7.8h, v1.h[6]\n" - "fmla v17.8h, v7.8h, v2.h[6]\n" - "fmla v21.8h, v7.8h, v3.h[6]\n" - "fmla v25.8h, v7.8h, v4.h[6]\n" - "ldr q7, [x10, #0x1b0]\n" - "fmla v10.8h, v6.8h, v0.h[6]\n" - "fmla v14.8h, v6.8h, v1.h[6]\n" - "fmla v18.8h, v6.8h, v2.h[6]\n" - "fmla v22.8h, v6.8h, v3.h[6]\n" - "fmla v26.8h, v6.8h, v4.h[6]\n" - "ldr q6, [x10, #0x1c0]\n" - "fmla v11.8h, v7.8h, v0.h[6]\n" - "fmla v15.8h, v7.8h, v1.h[6]\n" - "fmla v19.8h, v7.8h, v2.h[6]\n" - "fmla v23.8h, v7.8h, v3.h[6]\n" - "fmla v27.8h, v7.8h, v4.h[6]\n" - "ldr q7, [x10, #0x1d0]\n" - "fmla v8.8h, v6.8h, v0.h[7]\n" - "fmla v12.8h, v6.8h, v1.h[7]\n" - "fmla v16.8h, v6.8h, v2.h[7]\n" - "fmla v20.8h, v6.8h, v3.h[7]\n" - "fmla v24.8h, v6.8h, v4.h[7]\n" - "ldr q6, [x10, #0x1e0]\n" - "fmla v9.8h, v7.8h, v0.h[7]\n" - "fmla v13.8h, v7.8h, v1.h[7]\n" - "fmla v17.8h, v7.8h, v2.h[7]\n" - "fmla v21.8h, v7.8h, v3.h[7]\n" - "fmla v25.8h, v7.8h, v4.h[7]\n" - "ldr q7, [x10, #0x1f0]\n" - "add x10, x10, #0x200\n" - "fmla v10.8h, v6.8h, v0.h[7]\n" - "fmla v14.8h, v6.8h, v1.h[7]\n" - "fmla v18.8h, v6.8h, v2.h[7]\n" - "fmla v22.8h, v6.8h, v3.h[7]\n" - "fmla v26.8h, v6.8h, v4.h[7]\n" - "ldr q6, [x10, #0x0]\n" - "fmla v11.8h, v7.8h, v0.h[7]\n" - "ldr q0, [x26, #0x0]\n" - "fmla v15.8h, v7.8h, v1.h[7]\n" - "ldr q1, [x25, #0x0]\n" - "fmla v19.8h, v7.8h, v2.h[7]\n" - "ldr q2, [x24, #0x0]\n" - "fmla v23.8h, v7.8h, v3.h[7]\n" - "ldr q3, [x23, #0x0]\n" - "fmla v27.8h, v7.8h, v4.h[7]\n" - "ldr q4, [x22, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - "bge 222b\n" - "223:" // Height 5: Multiply loop: Single iteration only - "fmla v8.8h, v6.8h, v0.h[0]\n" - "fmla v12.8h, v6.8h, v1.h[0]\n" - "add x26, x26, #0x10\n" - "add x25, x25, #0x10\n" - "fmla v16.8h, v6.8h, v2.h[0]\n" - "fmla v20.8h, v6.8h, v3.h[0]\n" - "add x24, x24, #0x10\n" - "add x23, x23, #0x10\n" - "fmla v24.8h, v6.8h, v4.h[0]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.8h, v7.8h, v0.h[0]\n" - "add x22, x22, #0x10\n" - "fmla v13.8h, v7.8h, v1.h[0]\n" - "fmla v17.8h, v7.8h, v2.h[0]\n" - "sub x27, x27, #0x8\n" - "prfm pldl1keep, [x26, #0x80]\n" - "fmla v21.8h, v7.8h, v3.h[0]\n" - "fmla v25.8h, v7.8h, v4.h[0]\n" - "ldr q7, [x10, #0x30]\n" - "prfm pldl1keep, [x25, #0x80]\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" - "prfm pldl1keep, [x24, #0x80]\n" - "prfm pldl1keep, [x23, #0x80]\n" - "fmla v18.8h, v6.8h, v2.h[0]\n" - "fmla v22.8h, v6.8h, v3.h[0]\n" - "prfm pldl1keep, [x22, #0x80]\n" - "fmla v26.8h, v6.8h, v4.h[0]\n" - "ldr q6, [x10, #0x40]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" - "fmla v19.8h, v7.8h, v2.h[0]\n" - "fmla v23.8h, v7.8h, v3.h[0]\n" - "fmla v27.8h, v7.8h, v4.h[0]\n" - "ldr q7, [x10, #0x50]\n" - "fmla v8.8h, v6.8h, v0.h[1]\n" - "fmla v12.8h, v6.8h, v1.h[1]\n" - "fmla v16.8h, v6.8h, v2.h[1]\n" - "fmla v20.8h, v6.8h, v3.h[1]\n" - "fmla v24.8h, v6.8h, v4.h[1]\n" - "ldr q6, [x10, #0x60]\n" - "fmla v9.8h, v7.8h, v0.h[1]\n" - "fmla v13.8h, v7.8h, v1.h[1]\n" - "fmla v17.8h, v7.8h, v2.h[1]\n" - "fmla v21.8h, v7.8h, v3.h[1]\n" - "fmla v25.8h, v7.8h, v4.h[1]\n" - "ldr q7, [x10, #0x70]\n" - "fmla v10.8h, v6.8h, v0.h[1]\n" - "fmla v14.8h, v6.8h, v1.h[1]\n" - "fmla v18.8h, v6.8h, v2.h[1]\n" - "fmla v22.8h, v6.8h, v3.h[1]\n" - "fmla v26.8h, v6.8h, v4.h[1]\n" - "ldr q6, [x10, #0x80]\n" - "fmla v11.8h, v7.8h, v0.h[1]\n" - "fmla v15.8h, v7.8h, v1.h[1]\n" - "fmla v19.8h, v7.8h, v2.h[1]\n" - "fmla v23.8h, v7.8h, v3.h[1]\n" - "fmla v27.8h, v7.8h, v4.h[1]\n" - "ldr q7, [x10, #0x90]\n" - "fmla v8.8h, v6.8h, v0.h[2]\n" - "fmla v12.8h, v6.8h, v1.h[2]\n" - "fmla v16.8h, v6.8h, v2.h[2]\n" - "fmla v20.8h, v6.8h, v3.h[2]\n" - "fmla v24.8h, v6.8h, v4.h[2]\n" - "ldr q6, [x10, #0xa0]\n" - "fmla v9.8h, v7.8h, v0.h[2]\n" - "fmla v13.8h, v7.8h, v1.h[2]\n" - "fmla v17.8h, v7.8h, v2.h[2]\n" - "fmla v21.8h, v7.8h, v3.h[2]\n" - "fmla v25.8h, v7.8h, v4.h[2]\n" - "ldr q7, [x10, #0xb0]\n" - "fmla v10.8h, v6.8h, v0.h[2]\n" - "fmla v14.8h, v6.8h, v1.h[2]\n" - "fmla v18.8h, v6.8h, v2.h[2]\n" - "fmla v22.8h, v6.8h, v3.h[2]\n" - "fmla v26.8h, v6.8h, v4.h[2]\n" - "ldr q6, [x10, #0xc0]\n" - "fmla v11.8h, v7.8h, v0.h[2]\n" - "fmla v15.8h, v7.8h, v1.h[2]\n" - "fmla v19.8h, v7.8h, v2.h[2]\n" - "fmla v23.8h, v7.8h, v3.h[2]\n" - "fmla v27.8h, v7.8h, v4.h[2]\n" - "ldr q7, [x10, #0xd0]\n" - "fmla v8.8h, v6.8h, v0.h[3]\n" - "fmla v12.8h, v6.8h, v1.h[3]\n" - "fmla v16.8h, v6.8h, v2.h[3]\n" - "fmla v20.8h, v6.8h, v3.h[3]\n" - "fmla v24.8h, v6.8h, v4.h[3]\n" - "ldr q6, [x10, #0xe0]\n" - "fmla v9.8h, v7.8h, v0.h[3]\n" - "fmla v13.8h, v7.8h, v1.h[3]\n" - "fmla v17.8h, v7.8h, v2.h[3]\n" - "fmla v21.8h, v7.8h, v3.h[3]\n" - "fmla v25.8h, v7.8h, v4.h[3]\n" - "ldr q7, [x10, #0xf0]\n" - "fmla v10.8h, v6.8h, v0.h[3]\n" - "fmla v14.8h, v6.8h, v1.h[3]\n" - "fmla v18.8h, v6.8h, v2.h[3]\n" - "fmla v22.8h, v6.8h, v3.h[3]\n" - "fmla v26.8h, v6.8h, v4.h[3]\n" - "ldr q6, [x10, #0x100]\n" - "fmla v11.8h, v7.8h, v0.h[3]\n" - "fmla v15.8h, v7.8h, v1.h[3]\n" - "fmla v19.8h, v7.8h, v2.h[3]\n" - "fmla v23.8h, v7.8h, v3.h[3]\n" - "fmla v27.8h, v7.8h, v4.h[3]\n" - "ldr q7, [x10, #0x110]\n" - "fmla v8.8h, v6.8h, v0.h[4]\n" - "fmla v12.8h, v6.8h, v1.h[4]\n" - "fmla v16.8h, v6.8h, v2.h[4]\n" - "fmla v20.8h, v6.8h, v3.h[4]\n" - "fmla v24.8h, v6.8h, v4.h[4]\n" - "ldr q6, [x10, #0x120]\n" - "fmla v9.8h, v7.8h, v0.h[4]\n" - "fmla v13.8h, v7.8h, v1.h[4]\n" - "fmla v17.8h, v7.8h, v2.h[4]\n" - "fmla v21.8h, v7.8h, v3.h[4]\n" - "fmla v25.8h, v7.8h, v4.h[4]\n" - "ldr q7, [x10, #0x130]\n" - "fmla v10.8h, v6.8h, v0.h[4]\n" - "fmla v14.8h, v6.8h, v1.h[4]\n" - "fmla v18.8h, v6.8h, v2.h[4]\n" - "fmla v22.8h, v6.8h, v3.h[4]\n" - "fmla v26.8h, v6.8h, v4.h[4]\n" - "ldr q6, [x10, #0x140]\n" - "fmla v11.8h, v7.8h, v0.h[4]\n" - "fmla v15.8h, v7.8h, v1.h[4]\n" - "fmla v19.8h, v7.8h, v2.h[4]\n" - "fmla v23.8h, v7.8h, v3.h[4]\n" - "fmla v27.8h, v7.8h, v4.h[4]\n" - "ldr q7, [x10, #0x150]\n" - "fmla v8.8h, v6.8h, v0.h[5]\n" - "fmla v12.8h, v6.8h, v1.h[5]\n" - "fmla v16.8h, v6.8h, v2.h[5]\n" - "fmla v20.8h, v6.8h, v3.h[5]\n" - "fmla v24.8h, v6.8h, v4.h[5]\n" - "ldr q6, [x10, #0x160]\n" - "fmla v9.8h, v7.8h, v0.h[5]\n" - "fmla v13.8h, v7.8h, v1.h[5]\n" - "fmla v17.8h, v7.8h, v2.h[5]\n" - "fmla v21.8h, v7.8h, v3.h[5]\n" - "fmla v25.8h, v7.8h, v4.h[5]\n" - "ldr q7, [x10, #0x170]\n" - "fmla v10.8h, v6.8h, v0.h[5]\n" - "fmla v14.8h, v6.8h, v1.h[5]\n" - "fmla v18.8h, v6.8h, v2.h[5]\n" - "fmla v22.8h, v6.8h, v3.h[5]\n" - "fmla v26.8h, v6.8h, v4.h[5]\n" - "ldr q6, [x10, #0x180]\n" - "fmla v11.8h, v7.8h, v0.h[5]\n" - "fmla v15.8h, v7.8h, v1.h[5]\n" - "fmla v19.8h, v7.8h, v2.h[5]\n" - "fmla v23.8h, v7.8h, v3.h[5]\n" - "fmla v27.8h, v7.8h, v4.h[5]\n" - "ldr q7, [x10, #0x190]\n" - "fmla v8.8h, v6.8h, v0.h[6]\n" - "fmla v12.8h, v6.8h, v1.h[6]\n" - "fmla v16.8h, v6.8h, v2.h[6]\n" - "fmla v20.8h, v6.8h, v3.h[6]\n" - "fmla v24.8h, v6.8h, v4.h[6]\n" - "ldr q6, [x10, #0x1a0]\n" - "fmla v9.8h, v7.8h, v0.h[6]\n" - "fmla v13.8h, v7.8h, v1.h[6]\n" - "fmla v17.8h, v7.8h, v2.h[6]\n" - "fmla v21.8h, v7.8h, v3.h[6]\n" - "fmla v25.8h, v7.8h, v4.h[6]\n" - "ldr q7, [x10, #0x1b0]\n" - "fmla v10.8h, v6.8h, v0.h[6]\n" - "fmla v14.8h, v6.8h, v1.h[6]\n" - "fmla v18.8h, v6.8h, v2.h[6]\n" - "fmla v22.8h, v6.8h, v3.h[6]\n" - "fmla v26.8h, v6.8h, v4.h[6]\n" - "ldr q6, [x10, #0x1c0]\n" - "fmla v11.8h, v7.8h, v0.h[6]\n" - "fmla v15.8h, v7.8h, v1.h[6]\n" - "fmla v19.8h, v7.8h, v2.h[6]\n" - "fmla v23.8h, v7.8h, v3.h[6]\n" - "fmla v27.8h, v7.8h, v4.h[6]\n" - "ldr q7, [x10, #0x1d0]\n" - "fmla v8.8h, v6.8h, v0.h[7]\n" - "fmla v12.8h, v6.8h, v1.h[7]\n" - "fmla v16.8h, v6.8h, v2.h[7]\n" - "fmla v20.8h, v6.8h, v3.h[7]\n" - "fmla v24.8h, v6.8h, v4.h[7]\n" - "ldr q6, [x10, #0x1e0]\n" - "fmla v9.8h, v7.8h, v0.h[7]\n" - "fmla v13.8h, v7.8h, v1.h[7]\n" - "fmla v17.8h, v7.8h, v2.h[7]\n" - "fmla v21.8h, v7.8h, v3.h[7]\n" - "fmla v25.8h, v7.8h, v4.h[7]\n" - "ldr q7, [x10, #0x1f0]\n" - "add x10, x10, #0x200\n" - "fmla v10.8h, v6.8h, v0.h[7]\n" - "fmla v14.8h, v6.8h, v1.h[7]\n" - "fmla v18.8h, v6.8h, v2.h[7]\n" - "fmla v22.8h, v6.8h, v3.h[7]\n" - "fmla v26.8h, v6.8h, v4.h[7]\n" - "fmla v11.8h, v7.8h, v0.h[7]\n" - "fmla v15.8h, v7.8h, v1.h[7]\n" - "fmla v19.8h, v7.8h, v2.h[7]\n" - "fmla v23.8h, v7.8h, v3.h[7]\n" - "fmla v27.8h, v7.8h, v4.h[7]\n" - "224:" // Height 5: Multiply loop: Main loop skip - "cbz x27, 226f\n" - "225:" // Height 5: Multiply loop: Odd block loop - "ldr h0, [x26], #0x2\n" - "ldr h1, [x25], #0x2\n" - "sub x27, x27, #0x1\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" - "ldr h4, [x22], #0x2\n" + "mov x20, #0xc\n" + "ld1 { v16.s }[2], [x24], #0x4\n" + "ld1 { v20.s }[2], [x23], #0x4\n" + "ld1 { v24.s }[2], [x22], #0x4\n" + "tbz x11, #0, 215f\n" + "ld1 { v8.h }[6], [x9]\n" + "ld1 { v12.h }[6], [x25]\n" + "ld1 { v16.h }[6], [x24]\n" + "ld1 { v20.h }[6], [x23]\n" + "ld1 { v24.h }[6], [x22]\n" + "b 215f\n" + "212:" // Height 5: Partial accumulate: partial_1_4 + "mov x20, #0x8\n" + "tbz x11, #0, 215f\n" + "ld1 { v8.h }[4], [x9]\n" + "ld1 { v12.h }[4], [x25]\n" + "ld1 { v16.h }[4], [x24]\n" + "ld1 { v20.h }[4], [x23]\n" + "ld1 { v24.h }[4], [x22]\n" + "b 215f\n" + "213:" // Height 5: Partial accumulate: partial_2_0 + "tbz x11, #1, 214f\n" + "ldr s8, [x9], #0x4\n" + "ldr s12, [x25], #0x4\n" + "mov x20, #0x4\n" + "ldr s16, [x24], #0x4\n" + "ldr s20, [x23], #0x4\n" + "ldr s24, [x22], #0x4\n" + "tbz x11, #0, 215f\n" + "ld1 { v8.h }[2], [x9]\n" + "ld1 { v12.h }[2], [x25]\n" + "ld1 { v16.h }[2], [x24]\n" + "ld1 { v20.h }[2], [x23]\n" + "ld1 { v24.h }[2], [x22]\n" + "b 215f\n" + "214:" // Height 5: Partial accumulate: partial_1_0 + "ldr h8, [x9, #0x0]\n" + "ldr h12, [x25, #0x0]\n" + "mov x20, #0x0\n" + "ldr h16, [x24, #0x0]\n" + "ldr h20, [x23, #0x0]\n" + "ldr h24, [x22, #0x0]\n" + "215:" // Height 5: Partial accumulate: Done + "sub x9, x9, x20\n" + "b 218f\n" + "216:" // Height 5: full accumulate + "ldr q8, [x9, #0x0]\n" + "ldr q9, [x9, #0x10]\n" + "ldr q10, [x9, #0x20]\n" + "ldr q11, [x9, #0x30]\n" + "ldr q12, [x25, #0x0]\n" + "ldr q13, [x25, #0x10]\n" + "ldr q14, [x25, #0x20]\n" + "ldr q15, [x25, #0x30]\n" + "ldr q16, [x24, #0x0]\n" + "ldr q17, [x24, #0x10]\n" + "ldr q18, [x24, #0x20]\n" + "ldr q19, [x24, #0x30]\n" + "ldr q20, [x23, #0x0]\n" + "ldr q21, [x23, #0x10]\n" + "ldr q22, [x23, #0x20]\n" + "ldr q23, [x23, #0x30]\n" + "ldr q24, [x22, #0x0]\n" + "ldr q25, [x22, #0x10]\n" + "ldr q26, [x22, #0x20]\n" + "ldr q27, [x22, #0x30]\n" + "b 218f\n" + "217:" // Height 5: no accumulate + "movi v8.16b, #0x0\n" + "movi v9.16b, #0x0\n" + "movi v10.16b, #0x0\n" + "movi v11.16b, #0x0\n" + "movi v12.16b, #0x0\n" + "movi v13.16b, #0x0\n" + "movi v14.16b, #0x0\n" + "movi v15.16b, #0x0\n" + "movi v16.16b, #0x0\n" + "movi v17.16b, #0x0\n" + "movi v18.16b, #0x0\n" + "movi v19.16b, #0x0\n" + "movi v20.16b, #0x0\n" + "movi v21.16b, #0x0\n" + "movi v22.16b, #0x0\n" + "movi v23.16b, #0x0\n" + "movi v24.16b, #0x0\n" + "movi v25.16b, #0x0\n" + "movi v26.16b, #0x0\n" + "movi v27.16b, #0x0\n" + "218:" // Height 5: setup done + "mov x28, #0x0\n" + "219:" // Height 5: String loop + "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n" + "ldr w27, [x20, x28, LSL #0x2]\n" + "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n" + "tbz %x[flags], #3, 220f\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, 221f\n" + "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" + "add x26, x26, x20, LSL #1\n" + "add x25, x25, x20, LSL #1\n" + "add x24, x24, x20, LSL #1\n" + "add x23, x23, x20, LSL #1\n" + "add x22, x22, x20, LSL #1\n" + "b 221f\n" + "220:" // Height 5: setup direct input + "mov x26, %x[input_ptr]\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" + "add x22, x23, x21, LSL #1\n" + "221:" // Height 5: input setup done + "cmp x27, #0x8\n" + "blt 224f\n" + "ldr q0, [x26, #0x0]\n" + "ldr q1, [x25, #0x0]\n" + "cmp x27, #0x10\n" + "ldr q2, [x24, #0x0]\n" + "ldr q3, [x23, #0x0]\n" + "ldr q4, [x22, #0x0]\n" "ldr q6, [x10, #0x0]\n" + "ldr q7, [x10, #0x10]\n" + "blt 223f\n" + "222:" // Height 5: Multiply loop: Main loop head "fmla v8.8h, v6.8h, v0.h[0]\n" "fmla v12.8h, v6.8h, v1.h[0]\n" + "sub x27, x27, #0x8\n" + "add x26, x26, #0x10\n" + "fmla v16.8h, v6.8h, v2.h[0]\n" + "fmla v20.8h, v6.8h, v3.h[0]\n" + "add x25, x25, #0x10\n" + "add x24, x24, #0x10\n" + "fmla v24.8h, v6.8h, v4.h[0]\n" + "ldr q29, [x10, #0x20]\n" + "fmla v9.8h, v7.8h, v0.h[0]\n" + "add x23, x23, #0x10\n" + "fmla v13.8h, v7.8h, v1.h[0]\n" + "fmla v17.8h, v7.8h, v2.h[0]\n" + "add x22, x22, #0x10\n" + "cmp x27, #0x10\n" + "fmla v21.8h, v7.8h, v3.h[0]\n" + "fmla v25.8h, v7.8h, v4.h[0]\n" + "ldr q28, [x10, #0x30]\n" + "prfm pldl1keep, [x26, #0x80]\n" + "fmla v10.8h, v29.8h, v0.h[0]\n" + "fmla v14.8h, v29.8h, v1.h[0]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "fmla v18.8h, v29.8h, v2.h[0]\n" + "fmla v22.8h, v29.8h, v3.h[0]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "prfm pldl1keep, [x22, #0x80]\n" + "fmla v26.8h, v29.8h, v4.h[0]\n" + "ldr q29, [x10, #0x40]\n" + "fmla v11.8h, v28.8h, v0.h[0]\n" + "fmla v15.8h, v28.8h, v1.h[0]\n" + "fmla v19.8h, v28.8h, v2.h[0]\n" + "fmla v23.8h, v28.8h, v3.h[0]\n" + "fmla v27.8h, v28.8h, v4.h[0]\n" + "ldr q28, [x10, #0x50]\n" + "fmla v8.8h, v29.8h, v0.h[1]\n" + "fmla v12.8h, v29.8h, v1.h[1]\n" + "fmla v16.8h, v29.8h, v2.h[1]\n" + "fmla v20.8h, v29.8h, v3.h[1]\n" + "fmla v24.8h, v29.8h, v4.h[1]\n" + "ldr q29, [x10, #0x60]\n" + "fmla v9.8h, v28.8h, v0.h[1]\n" + "fmla v13.8h, v28.8h, v1.h[1]\n" + "fmla v17.8h, v28.8h, v2.h[1]\n" + "fmla v21.8h, v28.8h, v3.h[1]\n" + "fmla v25.8h, v28.8h, v4.h[1]\n" + "ldr q28, [x10, #0x70]\n" + "fmla v10.8h, v29.8h, v0.h[1]\n" + "fmla v14.8h, v29.8h, v1.h[1]\n" + "fmla v18.8h, v29.8h, v2.h[1]\n" + "fmla v22.8h, v29.8h, v3.h[1]\n" + "fmla v26.8h, v29.8h, v4.h[1]\n" + "ldr q29, [x10, #0x80]\n" + "fmla v11.8h, v28.8h, v0.h[1]\n" + "fmla v15.8h, v28.8h, v1.h[1]\n" + "fmla v19.8h, v28.8h, v2.h[1]\n" + "fmla v23.8h, v28.8h, v3.h[1]\n" + "fmla v27.8h, v28.8h, v4.h[1]\n" + "ldr q28, [x10, #0x90]\n" + "fmla v8.8h, v29.8h, v0.h[2]\n" + "fmla v12.8h, v29.8h, v1.h[2]\n" + "fmla v16.8h, v29.8h, v2.h[2]\n" + "fmla v20.8h, v29.8h, v3.h[2]\n" + "fmla v24.8h, v29.8h, v4.h[2]\n" + "ldr q29, [x10, #0xa0]\n" + "fmla v9.8h, v28.8h, v0.h[2]\n" + "fmla v13.8h, v28.8h, v1.h[2]\n" + "fmla v17.8h, v28.8h, v2.h[2]\n" + "fmla v21.8h, v28.8h, v3.h[2]\n" + "fmla v25.8h, v28.8h, v4.h[2]\n" + "ldr q28, [x10, #0xb0]\n" + "fmla v10.8h, v29.8h, v0.h[2]\n" + "fmla v14.8h, v29.8h, v1.h[2]\n" + "fmla v18.8h, v29.8h, v2.h[2]\n" + "fmla v22.8h, v29.8h, v3.h[2]\n" + "fmla v26.8h, v29.8h, v4.h[2]\n" + "ldr q29, [x10, #0xc0]\n" + "fmla v11.8h, v28.8h, v0.h[2]\n" + "fmla v15.8h, v28.8h, v1.h[2]\n" + "fmla v19.8h, v28.8h, v2.h[2]\n" + "fmla v23.8h, v28.8h, v3.h[2]\n" + "fmla v27.8h, v28.8h, v4.h[2]\n" + "ldr q28, [x10, #0xd0]\n" + "fmla v8.8h, v29.8h, v0.h[3]\n" + "fmla v12.8h, v29.8h, v1.h[3]\n" + "fmla v16.8h, v29.8h, v2.h[3]\n" + "fmla v20.8h, v29.8h, v3.h[3]\n" + "fmla v24.8h, v29.8h, v4.h[3]\n" + "ldr q29, [x10, #0xe0]\n" + "fmla v9.8h, v28.8h, v0.h[3]\n" + "fmla v13.8h, v28.8h, v1.h[3]\n" + "fmla v17.8h, v28.8h, v2.h[3]\n" + "fmla v21.8h, v28.8h, v3.h[3]\n" + "fmla v25.8h, v28.8h, v4.h[3]\n" + "ldr q28, [x10, #0xf0]\n" + "fmla v10.8h, v29.8h, v0.h[3]\n" + "fmla v14.8h, v29.8h, v1.h[3]\n" + "fmla v18.8h, v29.8h, v2.h[3]\n" + "fmla v22.8h, v29.8h, v3.h[3]\n" + "fmla v26.8h, v29.8h, v4.h[3]\n" + "ldr q29, [x10, #0x100]\n" + "fmla v11.8h, v28.8h, v0.h[3]\n" + "fmla v15.8h, v28.8h, v1.h[3]\n" + "fmla v19.8h, v28.8h, v2.h[3]\n" + "fmla v23.8h, v28.8h, v3.h[3]\n" + "fmla v27.8h, v28.8h, v4.h[3]\n" + "ldr q28, [x10, #0x110]\n" + "fmla v8.8h, v29.8h, v0.h[4]\n" + "fmla v12.8h, v29.8h, v1.h[4]\n" + "fmla v16.8h, v29.8h, v2.h[4]\n" + "fmla v20.8h, v29.8h, v3.h[4]\n" + "fmla v24.8h, v29.8h, v4.h[4]\n" + "ldr q29, [x10, #0x120]\n" + "fmla v9.8h, v28.8h, v0.h[4]\n" + "fmla v13.8h, v28.8h, v1.h[4]\n" + "fmla v17.8h, v28.8h, v2.h[4]\n" + "fmla v21.8h, v28.8h, v3.h[4]\n" + "fmla v25.8h, v28.8h, v4.h[4]\n" + "ldr q28, [x10, #0x130]\n" + "fmla v10.8h, v29.8h, v0.h[4]\n" + "fmla v14.8h, v29.8h, v1.h[4]\n" + "fmla v18.8h, v29.8h, v2.h[4]\n" + "fmla v22.8h, v29.8h, v3.h[4]\n" + "fmla v26.8h, v29.8h, v4.h[4]\n" + "ldr q29, [x10, #0x140]\n" + "fmla v11.8h, v28.8h, v0.h[4]\n" + "fmla v15.8h, v28.8h, v1.h[4]\n" + "fmla v19.8h, v28.8h, v2.h[4]\n" + "fmla v23.8h, v28.8h, v3.h[4]\n" + "fmla v27.8h, v28.8h, v4.h[4]\n" + "ldr q28, [x10, #0x150]\n" + "fmla v8.8h, v29.8h, v0.h[5]\n" + "fmla v12.8h, v29.8h, v1.h[5]\n" + "fmla v16.8h, v29.8h, v2.h[5]\n" + "fmla v20.8h, v29.8h, v3.h[5]\n" + "fmla v24.8h, v29.8h, v4.h[5]\n" + "ldr q29, [x10, #0x160]\n" + "fmla v9.8h, v28.8h, v0.h[5]\n" + "fmla v13.8h, v28.8h, v1.h[5]\n" + "fmla v17.8h, v28.8h, v2.h[5]\n" + "fmla v21.8h, v28.8h, v3.h[5]\n" + "fmla v25.8h, v28.8h, v4.h[5]\n" + "ldr q28, [x10, #0x170]\n" + "fmla v10.8h, v29.8h, v0.h[5]\n" + "fmla v14.8h, v29.8h, v1.h[5]\n" + "fmla v18.8h, v29.8h, v2.h[5]\n" + "fmla v22.8h, v29.8h, v3.h[5]\n" + "fmla v26.8h, v29.8h, v4.h[5]\n" + "ldr q29, [x10, #0x180]\n" + "fmla v11.8h, v28.8h, v0.h[5]\n" + "fmla v15.8h, v28.8h, v1.h[5]\n" + "fmla v19.8h, v28.8h, v2.h[5]\n" + "fmla v23.8h, v28.8h, v3.h[5]\n" + "fmla v27.8h, v28.8h, v4.h[5]\n" + "ldr q28, [x10, #0x190]\n" + "fmla v8.8h, v29.8h, v0.h[6]\n" + "fmla v12.8h, v29.8h, v1.h[6]\n" + "fmla v16.8h, v29.8h, v2.h[6]\n" + "fmla v20.8h, v29.8h, v3.h[6]\n" + "fmla v24.8h, v29.8h, v4.h[6]\n" + "ldr q29, [x10, #0x1a0]\n" + "fmla v9.8h, v28.8h, v0.h[6]\n" + "fmla v13.8h, v28.8h, v1.h[6]\n" + "fmla v17.8h, v28.8h, v2.h[6]\n" + "fmla v21.8h, v28.8h, v3.h[6]\n" + "fmla v25.8h, v28.8h, v4.h[6]\n" + "ldr q28, [x10, #0x1b0]\n" + "fmla v10.8h, v29.8h, v0.h[6]\n" + "fmla v14.8h, v29.8h, v1.h[6]\n" + "fmla v18.8h, v29.8h, v2.h[6]\n" + "fmla v22.8h, v29.8h, v3.h[6]\n" + "fmla v26.8h, v29.8h, v4.h[6]\n" + "ldr q29, [x10, #0x1c0]\n" + "fmla v11.8h, v28.8h, v0.h[6]\n" + "fmla v15.8h, v28.8h, v1.h[6]\n" + "fmla v19.8h, v28.8h, v2.h[6]\n" + "fmla v23.8h, v28.8h, v3.h[6]\n" + "fmla v27.8h, v28.8h, v4.h[6]\n" + "ldr q28, [x10, #0x1d0]\n" + "fmla v8.8h, v29.8h, v0.h[7]\n" + "fmla v12.8h, v29.8h, v1.h[7]\n" + "fmla v16.8h, v29.8h, v2.h[7]\n" + "fmla v20.8h, v29.8h, v3.h[7]\n" + "fmla v24.8h, v29.8h, v4.h[7]\n" + "ldr q29, [x10, #0x1e0]\n" + "fmla v9.8h, v28.8h, v0.h[7]\n" + "fmla v13.8h, v28.8h, v1.h[7]\n" + "fmla v17.8h, v28.8h, v2.h[7]\n" + "fmla v21.8h, v28.8h, v3.h[7]\n" + "fmla v25.8h, v28.8h, v4.h[7]\n" + "ldr q28, [x10, #0x1f0]\n" + "add x10, x10, #0x200\n" + "fmla v10.8h, v29.8h, v0.h[7]\n" + "fmla v14.8h, v29.8h, v1.h[7]\n" + "fmla v18.8h, v29.8h, v2.h[7]\n" + "fmla v22.8h, v29.8h, v3.h[7]\n" + "fmla v26.8h, v29.8h, v4.h[7]\n" + "ldr q6, [x10, #0x0]\n" + "fmla v11.8h, v28.8h, v0.h[7]\n" + "ldr q0, [x26, #0x0]\n" + "fmla v15.8h, v28.8h, v1.h[7]\n" + "ldr q1, [x25, #0x0]\n" + "fmla v19.8h, v28.8h, v2.h[7]\n" + "ldr q2, [x24, #0x0]\n" + "fmla v23.8h, v28.8h, v3.h[7]\n" + "ldr q3, [x23, #0x0]\n" + "fmla v27.8h, v28.8h, v4.h[7]\n" + "ldr q4, [x22, #0x0]\n" "ldr q7, [x10, #0x10]\n" + "bge 222b\n" + "223:" // Height 5: Multiply loop: Single iteration only + "fmla v8.8h, v6.8h, v0.h[0]\n" + "fmla v12.8h, v6.8h, v1.h[0]\n" + "add x26, x26, #0x10\n" + "add x25, x25, #0x10\n" "fmla v16.8h, v6.8h, v2.h[0]\n" "fmla v20.8h, v6.8h, v3.h[0]\n" + "add x24, x24, #0x10\n" + "add x23, x23, #0x10\n" "fmla v24.8h, v6.8h, v4.h[0]\n" - "ldr q6, [x10, #0x20]\n" + "ldr q29, [x10, #0x20]\n" "fmla v9.8h, v7.8h, v0.h[0]\n" + "add x22, x22, #0x10\n" "fmla v13.8h, v7.8h, v1.h[0]\n" "fmla v17.8h, v7.8h, v2.h[0]\n" + "sub x27, x27, #0x8\n" + "prfm pldl1keep, [x26, #0x80]\n" "fmla v21.8h, v7.8h, v3.h[0]\n" "fmla v25.8h, v7.8h, v4.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr q28, [x10, #0x30]\n" + "prfm pldl1keep, [x25, #0x80]\n" + "fmla v10.8h, v29.8h, v0.h[0]\n" + "fmla v14.8h, v29.8h, v1.h[0]\n" + "prfm pldl1keep, [x24, #0x80]\n" + "prfm pldl1keep, [x23, #0x80]\n" + "fmla v18.8h, v29.8h, v2.h[0]\n" + "fmla v22.8h, v29.8h, v3.h[0]\n" + "prfm pldl1keep, [x22, #0x80]\n" + "fmla v26.8h, v29.8h, v4.h[0]\n" + "ldr q29, [x10, #0x40]\n" + "fmla v11.8h, v28.8h, v0.h[0]\n" + "fmla v15.8h, v28.8h, v1.h[0]\n" + "fmla v19.8h, v28.8h, v2.h[0]\n" + "fmla v23.8h, v28.8h, v3.h[0]\n" + "fmla v27.8h, v28.8h, v4.h[0]\n" + "ldr q28, [x10, #0x50]\n" + "fmla v8.8h, v29.8h, v0.h[1]\n" + "fmla v12.8h, v29.8h, v1.h[1]\n" + "fmla v16.8h, v29.8h, v2.h[1]\n" + "fmla v20.8h, v29.8h, v3.h[1]\n" + "fmla v24.8h, v29.8h, v4.h[1]\n" + "ldr q29, [x10, #0x60]\n" + "fmla v9.8h, v28.8h, v0.h[1]\n" + "fmla v13.8h, v28.8h, v1.h[1]\n" + "fmla v17.8h, v28.8h, v2.h[1]\n" + "fmla v21.8h, v28.8h, v3.h[1]\n" + "fmla v25.8h, v28.8h, v4.h[1]\n" + "ldr q28, [x10, #0x70]\n" + "fmla v10.8h, v29.8h, v0.h[1]\n" + "fmla v14.8h, v29.8h, v1.h[1]\n" + "fmla v18.8h, v29.8h, v2.h[1]\n" + "fmla v22.8h, v29.8h, v3.h[1]\n" + "fmla v26.8h, v29.8h, v4.h[1]\n" + "ldr q29, [x10, #0x80]\n" + "fmla v11.8h, v28.8h, v0.h[1]\n" + "fmla v15.8h, v28.8h, v1.h[1]\n" + "fmla v19.8h, v28.8h, v2.h[1]\n" + "fmla v23.8h, v28.8h, v3.h[1]\n" + "fmla v27.8h, v28.8h, v4.h[1]\n" + "ldr q28, [x10, #0x90]\n" + "fmla v8.8h, v29.8h, v0.h[2]\n" + "fmla v12.8h, v29.8h, v1.h[2]\n" + "fmla v16.8h, v29.8h, v2.h[2]\n" + "fmla v20.8h, v29.8h, v3.h[2]\n" + "fmla v24.8h, v29.8h, v4.h[2]\n" + "ldr q29, [x10, #0xa0]\n" + "fmla v9.8h, v28.8h, v0.h[2]\n" + "fmla v13.8h, v28.8h, v1.h[2]\n" + "fmla v17.8h, v28.8h, v2.h[2]\n" + "fmla v21.8h, v28.8h, v3.h[2]\n" + "fmla v25.8h, v28.8h, v4.h[2]\n" + "ldr q28, [x10, #0xb0]\n" + "fmla v10.8h, v29.8h, v0.h[2]\n" + "fmla v14.8h, v29.8h, v1.h[2]\n" + "fmla v18.8h, v29.8h, v2.h[2]\n" + "fmla v22.8h, v29.8h, v3.h[2]\n" + "fmla v26.8h, v29.8h, v4.h[2]\n" + "ldr q29, [x10, #0xc0]\n" + "fmla v11.8h, v28.8h, v0.h[2]\n" + "fmla v15.8h, v28.8h, v1.h[2]\n" + "fmla v19.8h, v28.8h, v2.h[2]\n" + "fmla v23.8h, v28.8h, v3.h[2]\n" + "fmla v27.8h, v28.8h, v4.h[2]\n" + "ldr q28, [x10, #0xd0]\n" + "fmla v8.8h, v29.8h, v0.h[3]\n" + "fmla v12.8h, v29.8h, v1.h[3]\n" + "fmla v16.8h, v29.8h, v2.h[3]\n" + "fmla v20.8h, v29.8h, v3.h[3]\n" + "fmla v24.8h, v29.8h, v4.h[3]\n" + "ldr q29, [x10, #0xe0]\n" + "fmla v9.8h, v28.8h, v0.h[3]\n" + "fmla v13.8h, v28.8h, v1.h[3]\n" + "fmla v17.8h, v28.8h, v2.h[3]\n" + "fmla v21.8h, v28.8h, v3.h[3]\n" + "fmla v25.8h, v28.8h, v4.h[3]\n" + "ldr q28, [x10, #0xf0]\n" + "fmla v10.8h, v29.8h, v0.h[3]\n" + "fmla v14.8h, v29.8h, v1.h[3]\n" + "fmla v18.8h, v29.8h, v2.h[3]\n" + "fmla v22.8h, v29.8h, v3.h[3]\n" + "fmla v26.8h, v29.8h, v4.h[3]\n" + "ldr q29, [x10, #0x100]\n" + "fmla v11.8h, v28.8h, v0.h[3]\n" + "fmla v15.8h, v28.8h, v1.h[3]\n" + "fmla v19.8h, v28.8h, v2.h[3]\n" + "fmla v23.8h, v28.8h, v3.h[3]\n" + "fmla v27.8h, v28.8h, v4.h[3]\n" + "ldr q28, [x10, #0x110]\n" + "fmla v8.8h, v29.8h, v0.h[4]\n" + "fmla v12.8h, v29.8h, v1.h[4]\n" + "fmla v16.8h, v29.8h, v2.h[4]\n" + "fmla v20.8h, v29.8h, v3.h[4]\n" + "fmla v24.8h, v29.8h, v4.h[4]\n" + "ldr q29, [x10, #0x120]\n" + "fmla v9.8h, v28.8h, v0.h[4]\n" + "fmla v13.8h, v28.8h, v1.h[4]\n" + "fmla v17.8h, v28.8h, v2.h[4]\n" + "fmla v21.8h, v28.8h, v3.h[4]\n" + "fmla v25.8h, v28.8h, v4.h[4]\n" + "ldr q28, [x10, #0x130]\n" + "fmla v10.8h, v29.8h, v0.h[4]\n" + "fmla v14.8h, v29.8h, v1.h[4]\n" + "fmla v18.8h, v29.8h, v2.h[4]\n" + "fmla v22.8h, v29.8h, v3.h[4]\n" + "fmla v26.8h, v29.8h, v4.h[4]\n" + "ldr q29, [x10, #0x140]\n" + "fmla v11.8h, v28.8h, v0.h[4]\n" + "fmla v15.8h, v28.8h, v1.h[4]\n" + "fmla v19.8h, v28.8h, v2.h[4]\n" + "fmla v23.8h, v28.8h, v3.h[4]\n" + "fmla v27.8h, v28.8h, v4.h[4]\n" + "ldr q28, [x10, #0x150]\n" + "fmla v8.8h, v29.8h, v0.h[5]\n" + "fmla v12.8h, v29.8h, v1.h[5]\n" + "fmla v16.8h, v29.8h, v2.h[5]\n" + "fmla v20.8h, v29.8h, v3.h[5]\n" + "fmla v24.8h, v29.8h, v4.h[5]\n" + "ldr q29, [x10, #0x160]\n" + "fmla v9.8h, v28.8h, v0.h[5]\n" + "fmla v13.8h, v28.8h, v1.h[5]\n" + "fmla v17.8h, v28.8h, v2.h[5]\n" + "fmla v21.8h, v28.8h, v3.h[5]\n" + "fmla v25.8h, v28.8h, v4.h[5]\n" + "ldr q28, [x10, #0x170]\n" + "fmla v10.8h, v29.8h, v0.h[5]\n" + "fmla v14.8h, v29.8h, v1.h[5]\n" + "fmla v18.8h, v29.8h, v2.h[5]\n" + "fmla v22.8h, v29.8h, v3.h[5]\n" + "fmla v26.8h, v29.8h, v4.h[5]\n" + "ldr q29, [x10, #0x180]\n" + "fmla v11.8h, v28.8h, v0.h[5]\n" + "fmla v15.8h, v28.8h, v1.h[5]\n" + "fmla v19.8h, v28.8h, v2.h[5]\n" + "fmla v23.8h, v28.8h, v3.h[5]\n" + "fmla v27.8h, v28.8h, v4.h[5]\n" + "ldr q28, [x10, #0x190]\n" + "fmla v8.8h, v29.8h, v0.h[6]\n" + "fmla v12.8h, v29.8h, v1.h[6]\n" + "fmla v16.8h, v29.8h, v2.h[6]\n" + "fmla v20.8h, v29.8h, v3.h[6]\n" + "fmla v24.8h, v29.8h, v4.h[6]\n" + "ldr q29, [x10, #0x1a0]\n" + "fmla v9.8h, v28.8h, v0.h[6]\n" + "fmla v13.8h, v28.8h, v1.h[6]\n" + "fmla v17.8h, v28.8h, v2.h[6]\n" + "fmla v21.8h, v28.8h, v3.h[6]\n" + "fmla v25.8h, v28.8h, v4.h[6]\n" + "ldr q28, [x10, #0x1b0]\n" + "fmla v10.8h, v29.8h, v0.h[6]\n" + "fmla v14.8h, v29.8h, v1.h[6]\n" + "fmla v18.8h, v29.8h, v2.h[6]\n" + "fmla v22.8h, v29.8h, v3.h[6]\n" + "fmla v26.8h, v29.8h, v4.h[6]\n" + "ldr q29, [x10, #0x1c0]\n" + "fmla v11.8h, v28.8h, v0.h[6]\n" + "fmla v15.8h, v28.8h, v1.h[6]\n" + "fmla v19.8h, v28.8h, v2.h[6]\n" + "fmla v23.8h, v28.8h, v3.h[6]\n" + "fmla v27.8h, v28.8h, v4.h[6]\n" + "ldr q28, [x10, #0x1d0]\n" + "fmla v8.8h, v29.8h, v0.h[7]\n" + "fmla v12.8h, v29.8h, v1.h[7]\n" + "fmla v16.8h, v29.8h, v2.h[7]\n" + "fmla v20.8h, v29.8h, v3.h[7]\n" + "fmla v24.8h, v29.8h, v4.h[7]\n" + "ldr q29, [x10, #0x1e0]\n" + "fmla v9.8h, v28.8h, v0.h[7]\n" + "fmla v13.8h, v28.8h, v1.h[7]\n" + "fmla v17.8h, v28.8h, v2.h[7]\n" + "fmla v21.8h, v28.8h, v3.h[7]\n" + "fmla v25.8h, v28.8h, v4.h[7]\n" + "ldr q28, [x10, #0x1f0]\n" + "add x10, x10, #0x200\n" + "fmla v10.8h, v29.8h, v0.h[7]\n" + "fmla v14.8h, v29.8h, v1.h[7]\n" + "fmla v18.8h, v29.8h, v2.h[7]\n" + "fmla v22.8h, v29.8h, v3.h[7]\n" + "fmla v26.8h, v29.8h, v4.h[7]\n" + "fmla v11.8h, v28.8h, v0.h[7]\n" + "fmla v15.8h, v28.8h, v1.h[7]\n" + "fmla v19.8h, v28.8h, v2.h[7]\n" + "fmla v23.8h, v28.8h, v3.h[7]\n" + "fmla v27.8h, v28.8h, v4.h[7]\n" + "224:" // Height 5: Multiply loop: Main loop skip + "cbz x27, 226f\n" + "225:" // Height 5: Multiply loop: Odd block loop + "ldr h4, [x26], #0x2\n" + "ldr h3, [x25], #0x2\n" + "sub x27, x27, #0x1\n" + "ldr h2, [x24], #0x2\n" + "ldr h1, [x23], #0x2\n" + "ldr h0, [x22], #0x2\n" + "ldr q29, [x10, #0x0]\n" + "fmla v8.8h, v29.8h, v4.h[0]\n" + "fmla v12.8h, v29.8h, v3.h[0]\n" + "ldr q28, [x10, #0x10]\n" + "fmla v16.8h, v29.8h, v2.h[0]\n" + "fmla v20.8h, v29.8h, v1.h[0]\n" + "fmla v24.8h, v29.8h, v0.h[0]\n" + "ldr q29, [x10, #0x20]\n" + "fmla v9.8h, v28.8h, v4.h[0]\n" + "fmla v13.8h, v28.8h, v3.h[0]\n" + "fmla v17.8h, v28.8h, v2.h[0]\n" + "fmla v21.8h, v28.8h, v1.h[0]\n" + "fmla v25.8h, v28.8h, v0.h[0]\n" + "ldr q28, [x10, #0x30]\n" "add x10, x10, #0x40\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" - "fmla v18.8h, v6.8h, v2.h[0]\n" - "fmla v22.8h, v6.8h, v3.h[0]\n" - "fmla v26.8h, v6.8h, v4.h[0]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" - "fmla v19.8h, v7.8h, v2.h[0]\n" - "fmla v23.8h, v7.8h, v3.h[0]\n" - "fmla v27.8h, v7.8h, v4.h[0]\n" + "fmla v10.8h, v29.8h, v4.h[0]\n" + "fmla v14.8h, v29.8h, v3.h[0]\n" + "fmla v18.8h, v29.8h, v2.h[0]\n" + "fmla v22.8h, v29.8h, v1.h[0]\n" + "fmla v26.8h, v29.8h, v0.h[0]\n" + "fmla v11.8h, v28.8h, v4.h[0]\n" + "fmla v15.8h, v28.8h, v3.h[0]\n" + "fmla v19.8h, v28.8h, v2.h[0]\n" + "fmla v23.8h, v28.8h, v1.h[0]\n" + "fmla v27.8h, v28.8h, v0.h[0]\n" "cbnz x27, 225b\n" "226:" // Height 5: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -3719,49 +3719,49 @@ void a64_hybrid_fp16_mla_6x32 ( "prfm pstl1keep, [x22, #0x0]\n" "tbz %x[flags], #1, 227f\n" "add x20, %x[args_ptr], %[offset_max]\n" - "ld1r { v1.8h }, [x20]\n" + "ld1r { v29.8h }, [x20]\n" "add x20, %x[args_ptr], %[offset_min]\n" - "ld1r { v0.8h }, [x20]\n" - "fmin v8.8h, v8.8h, v1.8h\n" - "fmin v9.8h, v9.8h, v1.8h\n" - "fmin v10.8h, v10.8h, v1.8h\n" - "fmin v11.8h, v11.8h, v1.8h\n" - "fmin v12.8h, v12.8h, v1.8h\n" - "fmin v13.8h, v13.8h, v1.8h\n" - "fmin v14.8h, v14.8h, v1.8h\n" - "fmin v15.8h, v15.8h, v1.8h\n" - "fmin v16.8h, v16.8h, v1.8h\n" - "fmin v17.8h, v17.8h, v1.8h\n" - "fmin v18.8h, v18.8h, v1.8h\n" - "fmin v19.8h, v19.8h, v1.8h\n" - "fmin v20.8h, v20.8h, v1.8h\n" - "fmin v21.8h, v21.8h, v1.8h\n" - "fmin v22.8h, v22.8h, v1.8h\n" - "fmin v23.8h, v23.8h, v1.8h\n" - "fmin v24.8h, v24.8h, v1.8h\n" - "fmin v25.8h, v25.8h, v1.8h\n" - "fmin v26.8h, v26.8h, v1.8h\n" - "fmin v27.8h, v27.8h, v1.8h\n" - "fmax v8.8h, v8.8h, v0.8h\n" - "fmax v9.8h, v9.8h, v0.8h\n" - "fmax v10.8h, v10.8h, v0.8h\n" - "fmax v11.8h, v11.8h, v0.8h\n" - "fmax v12.8h, v12.8h, v0.8h\n" - "fmax v13.8h, v13.8h, v0.8h\n" - "fmax v14.8h, v14.8h, v0.8h\n" - "fmax v15.8h, v15.8h, v0.8h\n" - "fmax v16.8h, v16.8h, v0.8h\n" - "fmax v17.8h, v17.8h, v0.8h\n" - "fmax v18.8h, v18.8h, v0.8h\n" - "fmax v19.8h, v19.8h, v0.8h\n" - "fmax v20.8h, v20.8h, v0.8h\n" - "fmax v21.8h, v21.8h, v0.8h\n" - "fmax v22.8h, v22.8h, v0.8h\n" - "fmax v23.8h, v23.8h, v0.8h\n" - "fmax v24.8h, v24.8h, v0.8h\n" - "fmax v25.8h, v25.8h, v0.8h\n" - "fmax v26.8h, v26.8h, v0.8h\n" - "fmax v27.8h, v27.8h, v0.8h\n" + "ld1r { v28.8h }, [x20]\n" + "fmin v8.8h, v8.8h, v29.8h\n" + "fmin v9.8h, v9.8h, v29.8h\n" + "fmin v10.8h, v10.8h, v29.8h\n" + "fmin v11.8h, v11.8h, v29.8h\n" + "fmin v12.8h, v12.8h, v29.8h\n" + "fmin v13.8h, v13.8h, v29.8h\n" + "fmin v14.8h, v14.8h, v29.8h\n" + "fmin v15.8h, v15.8h, v29.8h\n" + "fmin v16.8h, v16.8h, v29.8h\n" + "fmin v17.8h, v17.8h, v29.8h\n" + "fmin v18.8h, v18.8h, v29.8h\n" + "fmin v19.8h, v19.8h, v29.8h\n" + "fmin v20.8h, v20.8h, v29.8h\n" + "fmin v21.8h, v21.8h, v29.8h\n" + "fmin v22.8h, v22.8h, v29.8h\n" + "fmin v23.8h, v23.8h, v29.8h\n" + "fmin v24.8h, v24.8h, v29.8h\n" + "fmin v25.8h, v25.8h, v29.8h\n" + "fmin v26.8h, v26.8h, v29.8h\n" + "fmin v27.8h, v27.8h, v29.8h\n" + "fmax v8.8h, v8.8h, v28.8h\n" + "fmax v9.8h, v9.8h, v28.8h\n" + "fmax v10.8h, v10.8h, v28.8h\n" + "fmax v11.8h, v11.8h, v28.8h\n" + "fmax v12.8h, v12.8h, v28.8h\n" + "fmax v13.8h, v13.8h, v28.8h\n" + "fmax v14.8h, v14.8h, v28.8h\n" + "fmax v15.8h, v15.8h, v28.8h\n" + "fmax v16.8h, v16.8h, v28.8h\n" + "fmax v17.8h, v17.8h, v28.8h\n" + "fmax v18.8h, v18.8h, v28.8h\n" + "fmax v19.8h, v19.8h, v28.8h\n" + "fmax v20.8h, v20.8h, v28.8h\n" + "fmax v21.8h, v21.8h, v28.8h\n" + "fmax v22.8h, v22.8h, v28.8h\n" + "fmax v23.8h, v23.8h, v28.8h\n" + "fmax v24.8h, v24.8h, v28.8h\n" + "fmax v25.8h, v25.8h, v28.8h\n" + "fmax v26.8h, v26.8h, v28.8h\n" + "fmax v27.8h, v27.8h, v28.8h\n" "227:" // Height 5: No activation "cmp x11, #0x20\n" "bge 244f\n" @@ -4386,16 +4386,16 @@ void a64_hybrid_fp16_mla_6x32 ( "268:" // 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, 269f\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, 270f\n" "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n" "add x26, x26, x20, LSL #1\n" @@ -4407,11 +4407,11 @@ void a64_hybrid_fp16_mla_6x32 ( "b 270f\n" "269:" // Height 6: setup direct input "mov x26, %x[input_ptr]\n" - "add x25, x26, x20, LSL #1\n" - "add x24, x25, x20, LSL #1\n" - "add x23, x24, x20, LSL #1\n" - "add x22, x23, x20, LSL #1\n" - "add x21, x22, x20, LSL #1\n" + "add x25, x26, x21, LSL #1\n" + "add x24, x25, x21, LSL #1\n" + "add x23, x24, x21, LSL #1\n" + "add x22, x23, x21, LSL #1\n" + "add x21, x22, x21, LSL #1\n" "270:" // Height 6: input setup done "cmp x27, #0x8\n" "blt 273f\n" @@ -4912,42 +4912,42 @@ void a64_hybrid_fp16_mla_6x32 ( "273:" // Height 6: Multiply loop: Main loop skip "cbz x27, 275f\n" "274:" // Height 6: Multiply loop: Odd block loop - "ldr h0, [x26], #0x2\n" - "ldr h1, [x25], #0x2\n" + "ldr h7, [x26], #0x2\n" + "ldr h6, [x25], #0x2\n" "sub x27, x27, #0x1\n" - "ldr h2, [x24], #0x2\n" - "ldr h3, [x23], #0x2\n" - "ldr h4, [x22], #0x2\n" - "ldr h5, [x21], #0x2\n" - "ldr q6, [x10, #0x0]\n" - "ldr q7, [x10, #0x10]\n" - "fmla v8.8h, v6.8h, v0.h[0]\n" - "fmla v12.8h, v6.8h, v1.h[0]\n" - "fmla v16.8h, v6.8h, v2.h[0]\n" - "fmla v20.8h, v6.8h, v3.h[0]\n" - "fmla v24.8h, v6.8h, v4.h[0]\n" - "fmla v28.8h, v6.8h, v5.h[0]\n" - "ldr q6, [x10, #0x20]\n" - "fmla v9.8h, v7.8h, v0.h[0]\n" - "fmla v13.8h, v7.8h, v1.h[0]\n" - "fmla v17.8h, v7.8h, v2.h[0]\n" - "fmla v21.8h, v7.8h, v3.h[0]\n" - "fmla v25.8h, v7.8h, v4.h[0]\n" - "fmla v29.8h, v7.8h, v5.h[0]\n" - "ldr q7, [x10, #0x30]\n" + "ldr h5, [x24], #0x2\n" + "ldr h4, [x23], #0x2\n" + "ldr h3, [x22], #0x2\n" + "ldr h2, [x21], #0x2\n" + "ldr q1, [x10, #0x0]\n" + "ldr q0, [x10, #0x10]\n" + "fmla v8.8h, v1.8h, v7.h[0]\n" + "fmla v12.8h, v1.8h, v6.h[0]\n" + "fmla v16.8h, v1.8h, v5.h[0]\n" + "fmla v20.8h, v1.8h, v4.h[0]\n" + "fmla v24.8h, v1.8h, v3.h[0]\n" + "fmla v28.8h, v1.8h, v2.h[0]\n" + "ldr q1, [x10, #0x20]\n" + "fmla v9.8h, v0.8h, v7.h[0]\n" + "fmla v13.8h, v0.8h, v6.h[0]\n" + "fmla v17.8h, v0.8h, v5.h[0]\n" + "fmla v21.8h, v0.8h, v4.h[0]\n" + "fmla v25.8h, v0.8h, v3.h[0]\n" + "fmla v29.8h, v0.8h, v2.h[0]\n" + "ldr q0, [x10, #0x30]\n" "add x10, x10, #0x40\n" - "fmla v10.8h, v6.8h, v0.h[0]\n" - "fmla v14.8h, v6.8h, v1.h[0]\n" - "fmla v18.8h, v6.8h, v2.h[0]\n" - "fmla v22.8h, v6.8h, v3.h[0]\n" - "fmla v26.8h, v6.8h, v4.h[0]\n" - "fmla v30.8h, v6.8h, v5.h[0]\n" - "fmla v11.8h, v7.8h, v0.h[0]\n" - "fmla v15.8h, v7.8h, v1.h[0]\n" - "fmla v19.8h, v7.8h, v2.h[0]\n" - "fmla v23.8h, v7.8h, v3.h[0]\n" - "fmla v27.8h, v7.8h, v4.h[0]\n" - "fmla v31.8h, v7.8h, v5.h[0]\n" + "fmla v10.8h, v1.8h, v7.h[0]\n" + "fmla v14.8h, v1.8h, v6.h[0]\n" + "fmla v18.8h, v1.8h, v5.h[0]\n" + "fmla v22.8h, v1.8h, v4.h[0]\n" + "fmla v26.8h, v1.8h, v3.h[0]\n" + "fmla v30.8h, v1.8h, v2.h[0]\n" + "fmla v11.8h, v0.8h, v7.h[0]\n" + "fmla v15.8h, v0.8h, v6.h[0]\n" + "fmla v19.8h, v0.8h, v5.h[0]\n" + "fmla v23.8h, v0.8h, v4.h[0]\n" + "fmla v27.8h, v0.8h, v3.h[0]\n" + "fmla v31.8h, v0.8h, v2.h[0]\n" "cbnz x27, 274b\n" "275:" // Height 6: Multiply loop: No odd multiplies "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n" @@ -5317,7 +5317,6 @@ void a64_hybrid_fp16_mla_6x32 ( "madd %x[input_ptr], x20, x21, %x[input_ptr]\n" "b 1b\n" "296:" // 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" -- cgit v1.2.1