Fix depthwise convolution assembly kernels

Resolves: COMPMID-4710 Change-Id: I35b964731aeed0e6f4f873f59341bee48e4a41fd Signed-off-by: Freddie Liardet <frederick.liardet@arm.com> Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6039 Tested-by: Arm Jenkins <bsgcomp@arm.com> Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com> Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com> Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
author: Freddie Liardet <frederick.liardet@arm.com> 2021-08-04 11:57:37 +0100
committer: Sheri Zhang <sheri.zhang@arm.com> 2021-08-04 15:54:00 +0000
commit: 6269f87f5caca3a83a4822e0556fbf284158615c (patch)
tree: 0f814b2f6356aa073529a8d00c94913e006d1f6d /src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
parent: 38ce29d5dce9a09769b55b2717d105f984247e31 (diff)
download: ComputeLibrary-6269f87f5caca3a83a4822e0556fbf284158615c.tar.gz
1 files changed, 1483 insertions, 979 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
index bf18469199..c403576cf1 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
@@ -87,1143 +87,1647 @@ void a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
   );
 
   __asm__ __volatile__(
-    "mov x4, #0x0\n"
     "mov x26, #0x0\n"
+    "mov x25, #0x0\n"
     "1:"  // Tile loop
-    "str x4, [%x[params_struct], %[offsetof_args_tile_i]]\n"
-    "mov x25, #0x4\n"
-    "str x26, [%x[params_struct], %[offsetof_args_tile_j]]\n"
+    "str x26, [%x[params_struct], %[offsetof_args_tile_i]]\n"
     "mov x24, #0x4\n"
-    "ldr x5, [%x[params_struct], %[offsetof_args_params]]\n"
-    "add x23, %x[params_struct], %[offsetof_args_min]\n"
-    "ldr x22, [%x[params_struct], %[offsetof_args_ld_input_row]]\n"
-    "add x21, %x[params_struct], %[offsetof_args_max]\n"
-    "ldr x6, [%x[params_struct], %[offsetof_args_ld_input_col]]\n"
-    "mov x7, #0x0\n"
+    "mov x22, #0x4\n"
+    "str x25, [%x[params_struct], %[offsetof_args_tile_j]]\n"
+    "ldr x23, [%x[params_struct], %[offsetof_args_ld_input_row]]\n"
+    "ldr x21, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
+    "mul x20, x26, x23\n"  // offset = tile_i * ld_input_row
+    "ldr x5, [%x[params_struct], %[offsetof_args_ld_input_col]]\n"
+    "ldr x6, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
+    "mul x19, x26, x21\n"  // offset = tile_i * ld_output_row
+    "mov x7, #0x10\n"  // cntb _, ALL, #1
+    "madd x20, x25, x5, x20\n"  // offset += tile_j * ld_input_col
     "ldr x8, [%x[params_struct], %[offsetof_args_inptr]]\n"
-    "mul x19, x4, x22\n" // offset = tile_i * ld_input_row
-    "ldr x20, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
-    "madd x19, x26, x6, x19\n" // offset += tile_j * ld_input_col
-    "ldr x17, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
-    "mul x19, x19, x25\n" // offset *= kernel_stride * output_size
-    "ldr x16, [%x[params_struct], %[offsetof_args_outptr]]\n"
-    "add x8, x8, x19, LSL #1\n" // inptr[0] += offset * sizeof(__fp16)
-    "ld1r { v15.8h }, [x23]\n"
-    "add x15, x8, x22, LSL #1\n"
-    "ld1r { v14.8h }, [x21]\n"
-    "add x14, x15, x22, LSL #1\n"
+    "lsl x5, x5, #0x1\n"
+    "ldr x17, [%x[params_struct], %[offsetof_args_outptr]]\n"
+    "madd x19, x25, x6, x19\n"  // offset += tile_j * ld_output_col
     "lsl x6, x6, #0x1\n"
-    "add x13, x14, x22, LSL #1\n"
-    "add x12, x13, x22, LSL #1\n"
-    "add x11, x12, x22, LSL #1\n"
-    "add x10, x6, x6\n"
-    "add x9, x10, x6\n"
-    "add x28, x9, x6\n"
-    "add x27, x28, x6\n"
-    "mul x19, x4, x20\n" // offset = tile_i * ld_output_row
-    "madd x19, x26, x17, x19\n" // offset += tile_j * ld_output_col
-    "mul x19, x19, x24\n" // offset *= output_tile_size
-    "add x16, x16, x19, LSL #1\n" // outptrs[0] += offset * sizeof(__fp16)
-    "add x26, x16, x20, LSL #1\n"
-    "add x25, x26, x20, LSL #1\n"
-    "add x24, x25, x20, LSL #1\n"
-    "lsl x17, x17, #0x1\n"
-    "add x23, x17, x17\n"
-    "add x22, x23, x17\n"
-    "mov x21, #0x10\n" // cntb _, ALL, #1
-    "sub x20, XZR, x21\n"
-    "lsr x19, %x[n_channels], #0x3\n"
-    "cbz x19, 4f\n"
-    "ldr q13, [x5, #0x0]\n"
-    "ldr q0, [x5, #0x10]\n"
-    "cmp x21, x19, LSL #4\n"
-    "ldr q1, [x5, #0x20]\n"
-    "ldr q2, [x5, #0x30]\n"
-    "ldr q3, [x5, #0x40]\n"
-    "ldr q4, [x5, #0x50]\n"
-    "ldr q5, [x5, #0x60]\n"
-    "ldr q6, [x5, #0x70]\n"
-    "ldr q7, [x5, #0x80]\n"
-    "ldr q8, [x5, #0x90]\n"
-    "add x5, x5, #0xa0\n"
-    "ldr q9, [x14, x10]\n"
+    "add x16, x5, x5\n"
+    "ldr x15, [%x[params_struct], %[offsetof_args_params]]\n"
+    "mul x20, x20, x24\n"  // offset *= kernel_stride * output_size
+    "add x8, x8, x20, LSL #1\n"  // inptr[0] += offset * sizeof(__fp16)
+    "add x14, x8, x23, LSL #1\n"
+    "mul x19, x19, x22\n"  // offset *= output_tile_size
+    "add x13, x14, x23, LSL #1\n"
+    "add x17, x17, x19, LSL #1\n"  // outptrs[0] += offset * sizeof(__fp16)
+    "lsr x12, %x[n_channels], #0x3\n"
+    "add x11, x13, x23, LSL #1\n"
+    "add x10, x16, x5\n"
+    "add x9, x17, x21, LSL #1\n"
+    "add x28, x11, x23, LSL #1\n"
+    "add x27, x10, x5\n"
+    "add x26, x9, x21, LSL #1\n"
+    "add x22, x6, x6\n"
+    "add x20, %x[params_struct], %[offsetof_args_min]\n"
+    "add x19, %x[params_struct], %[offsetof_args_max]\n"
+    "ld1r { v15.8h }, [x20]\n"
+    "ld1r { v14.8h }, [x19]\n"
+    "add x25, x28, x23, LSL #1\n"
+    "add x24, x27, x5\n"
+    "add x23, x26, x21, LSL #1\n"
+    "add x21, x22, x6\n"
+    "mov x20, #0x0\n"
+    "sub x19, XZR, x7\n"
+    "cbz x12, 4f\n"
+    "ldr q13, [x15, #0x0]\n"
+    "cmp x7, x12, LSL #4\n"
+    "ldr q0, [x15, #0x10]\n"
+    "ldr q1, [x15, #0x20]\n"
+    "ldr q2, [x15, #0x30]\n"
+    "ldr q3, [x15, #0x40]\n"
+    "ldr q4, [x15, #0x50]\n"
+    "ldr q5, [x15, #0x60]\n"
+    "ldr q6, [x15, #0x70]\n"
+    "ldr q7, [x15, #0x80]\n"
+    "ldr q8, [x15, #0x90]\n"
+    "ldr q9, [x13, x16]\n"
+    "add x15, x15, #0xa0\n"
     "ld1 { v10.8h }, [x8]\n"
-    "ldr q11, [x8, x27]\n"
-    "ldr q12, [x14, x9]\n"
+    "ldr q11, [x8, x24]\n"
+    "ldr q12, [x13, x10]\n"
     "bge 3f\n"
     "2:"  // Tile loop: Channel loop
-    "mov v31.16b, v13.16b\n fmla v31.8h, v8.8h, v9.8h\n"
-    "add x20, x20, #0x10\n"
-    "mov v30.16b, v13.16b\n fmla v30.8h, v7.8h, v9.8h\n"
+    "mov v21.16b, v13.16b\n fmla v21.8h, v4.8h, v9.8h\n"
+    "mov v16.16b, v13.16b\n fmla v16.8h, v8.8h, v9.8h\n"
     "add x7, x7, #0x10\n"
-    "mov v29.16b, v13.16b\n fmla v29.8h, v6.8h, v9.8h\n"
-    "add x21, x21, #0x10\n"
-    "mov v27.16b, v13.16b\n fmla v27.8h, v5.8h, v9.8h\n"
-    "cmp x21, x19, LSL #4\n"
-    "mov v26.16b, v13.16b\n fmla v26.8h, v4.8h, v9.8h\n"
-    "mov v25.16b, v13.16b\n fmla v25.8h, v3.8h, v9.8h\n"
-    "mov v23.16b, v13.16b\n fmla v23.8h, v2.8h, v9.8h\n"
-    "mov v22.16b, v13.16b\n fmla v22.8h, v1.8h, v9.8h\n"
-    "mov v21.16b, v13.16b\n fmla v21.8h, v0.8h, v9.8h\n"
-    "ldr q9, [x13, x10]\n"
-    "fmla v31.8h, v0.8h, v10.8h\n"
-    "ld1 { v10.8h }, [x11]\n"
-    "mov v28.16b, v13.16b\n fmla v28.8h, v2.8h, v11.8h\n"
-    "ldr q11, [x11, x27]\n"
-    "fmla v30.8h, v8.8h, v12.8h\n"
-    "fmla v29.8h, v7.8h, v12.8h\n"
-    "fmla v26.8h, v5.8h, v12.8h\n"
-    "fmla v28.8h, v6.8h, v12.8h\n"
-    "fmla v25.8h, v4.8h, v12.8h\n"
-    "mov v24.16b, v13.16b\n fmla v24.8h, v3.8h, v12.8h\n"
-    "fmla v22.8h, v2.8h, v12.8h\n"
-    "fmla v21.8h, v1.8h, v12.8h\n"
-    "mov v20.16b, v13.16b\n fmla v20.8h, v0.8h, v12.8h\n"
-    "ldr q12, [x8, x6]\n"
-    "mov v19.16b, v13.16b\n fmla v19.8h, v6.8h, v10.8h\n"
-    "ldr q10, [x13, x9]\n"
-    "mov v16.16b, v13.16b\n fmla v16.8h, v8.8h, v11.8h\n"
-    "ldr q11, [x8, x28]\n"
-    "fmla v27.8h, v8.8h, v9.8h\n"
-    "fmla v26.8h, v7.8h, v9.8h\n"
-    "fmla v25.8h, v6.8h, v9.8h\n"
-    "fmla v23.8h, v5.8h, v9.8h\n"
-    "fmla v22.8h, v4.8h, v9.8h\n"
-    "fmla v21.8h, v3.8h, v9.8h\n"
-    "fmla v19.8h, v2.8h, v9.8h\n"
-    "mov v18.16b, v13.16b\n fmla v18.8h, v1.8h, v9.8h\n"
-    "mov v17.16b, v13.16b\n fmla v17.8h, v0.8h, v9.8h\n"
-    "ld1 { v9.8h }, [x15]\n"
-    "fmla v31.8h, v1.8h, v12.8h\n"
-    "ldr q13, [x5, #0x0]\n"
-    "fmla v30.8h, v0.8h, v12.8h\n"
-    "ldr q12, [x15, x27]\n"
-    "fmla v29.8h, v2.8h, v11.8h\n"
-    "fmla v28.8h, v1.8h, v11.8h\n"
-    "ld1 { v11.8h }, [x12]\n"
-    "fmla v26.8h, v8.8h, v10.8h\n"
-    "fmla v25.8h, v7.8h, v10.8h\n"
-    "fmla v24.8h, v6.8h, v10.8h\n"
-    "fmla v22.8h, v5.8h, v10.8h\n"
-    "fmla v21.8h, v4.8h, v10.8h\n"
-    "fmla v20.8h, v3.8h, v10.8h\n"
-    "fmla v18.8h, v2.8h, v10.8h\n"
-    "fmla v17.8h, v1.8h, v10.8h\n"
+    "cmp x7, x12, LSL #4\n"
+    "mov v22.16b, v13.16b\n fmla v22.8h, v3.8h, v9.8h\n"
+    "mov v25.16b, v13.16b\n fmla v25.8h, v1.8h, v9.8h\n"
+    "add x19, x19, #0x10\n"
+    "add x20, x20, #0x10\n"
+    "mov v26.16b, v13.16b\n fmla v26.8h, v0.8h, v9.8h\n"
+    "fmla v21.8h, v5.8h, v12.8h\n"
+    "mov v17.16b, v13.16b\n fmla v17.8h, v7.8h, v9.8h\n"
+    "mov v18.16b, v13.16b\n fmla v18.8h, v6.8h, v9.8h\n"
+    "mov v20.16b, v13.16b\n fmla v20.8h, v5.8h, v9.8h\n"
+    "mov v24.16b, v13.16b\n fmla v24.8h, v2.8h, v9.8h\n"
+    "ldr q9, [x11, x16]\n"
     "fmla v16.8h, v0.8h, v10.8h\n"
-    "ldr q10, [x15, x10]\n"
-    "fmla v31.8h, v3.8h, v9.8h\n"
-    "fmla v27.8h, v0.8h, v9.8h\n"
-    "fmla v28.8h, v5.8h, v12.8h\n"
-    "fmla v24.8h, v2.8h, v12.8h\n"
-    "ldr q12, [x15, x9]\n"
-    "fmla v23.8h, v6.8h, v11.8h\n"
-    "fmla v19.8h, v3.8h, v11.8h\n"
-    "ldr q11, [x12, x27]\n"
-    "fmla v31.8h, v5.8h, v10.8h\n"
-    "fmla v30.8h, v4.8h, v10.8h\n"
-    "fmla v29.8h, v3.8h, v10.8h\n"
-    "fmla v27.8h, v2.8h, v10.8h\n"
-    "fmla v26.8h, v1.8h, v10.8h\n"
-    "fmla v25.8h, v0.8h, v10.8h\n"
-    "ldr q10, [x14, x6]\n"
-    "fmla v20.8h, v8.8h, v11.8h\n"
-    "fmla v16.8h, v5.8h, v11.8h\n"
-    "ldr q11, [x11, x6]\n"
-    "fmla v30.8h, v5.8h, v12.8h\n"
-    "fmla v29.8h, v4.8h, v12.8h\n"
-    "fmla v28.8h, v3.8h, v12.8h\n"
-    "fmla v26.8h, v2.8h, v12.8h\n"
-    "fmla v25.8h, v1.8h, v12.8h\n"
-    "fmla v24.8h, v0.8h, v12.8h\n"
-    "ldr q12, [x14, x28]\n"
-    "fmla v19.8h, v7.8h, v11.8h\n"
-    "fmla v18.8h, v6.8h, v11.8h\n"
-    "ldr q11, [x11, x28]\n"
-    "fmla v31.8h, v7.8h, v10.8h\n"
-    "fmla v30.8h, v6.8h, v10.8h\n"
-    "fmla v27.8h, v4.8h, v10.8h\n"
-    "fmla v26.8h, v3.8h, v10.8h\n"
-    "fmla v23.8h, v1.8h, v10.8h\n"
+    "mov v19.16b, v13.16b\n fmla v19.8h, v2.8h, v11.8h\n"
+    "ld1 { v10.8h }, [x25]\n"
+    "ldr q11, [x25, x24]\n"
+    "fmla v22.8h, v4.8h, v12.8h\n"
+    "fmla v25.8h, v2.8h, v12.8h\n"
+    "fmla v26.8h, v1.8h, v12.8h\n"
+    "mov v28.16b, v13.16b\n fmla v28.8h, v6.8h, v10.8h\n"
+    "ldr q10, [x11, x10]\n"
+    "fmla v21.8h, v7.8h, v9.8h\n"
+    "fmla v17.8h, v8.8h, v12.8h\n"
+    "fmla v18.8h, v7.8h, v12.8h\n"
+    "fmla v19.8h, v6.8h, v12.8h\n"
+    "mov v23.16b, v13.16b\n fmla v23.8h, v3.8h, v12.8h\n"
+    "mov v27.16b, v13.16b\n fmla v27.8h, v0.8h, v12.8h\n"
+    "ldr q12, [x8, x5]\n"
+    "mov v31.16b, v13.16b\n fmla v31.8h, v8.8h, v11.8h\n"
+    "fmla v22.8h, v6.8h, v9.8h\n"
+    "ldr q11, [x8, x27]\n"
+    "fmla v25.8h, v4.8h, v9.8h\n"
+    "fmla v26.8h, v3.8h, v9.8h\n"
+    "mov v29.16b, v13.16b\n fmla v29.8h, v1.8h, v9.8h\n"
+    "mov v30.16b, v13.16b\n fmla v30.8h, v0.8h, v9.8h\n"
+    "ldr q13, [x15, #0x0]\n"
+    "fmla v20.8h, v8.8h, v9.8h\n"
+    "fmla v24.8h, v5.8h, v9.8h\n"
+    "fmla v28.8h, v2.8h, v9.8h\n"
+    "fmla v21.8h, v8.8h, v10.8h\n"
+    "ld1 { v9.8h }, [x14]\n"
+    "fmla v16.8h, v1.8h, v12.8h\n"
+    "fmla v17.8h, v0.8h, v12.8h\n"
+    "ldr q12, [x14, x24]\n"
+    "fmla v18.8h, v2.8h, v11.8h\n"
+    "fmla v19.8h, v1.8h, v11.8h\n"
+    "ld1 { v11.8h }, [x28]\n"
+    "fmla v22.8h, v7.8h, v10.8h\n"
+    "fmla v23.8h, v6.8h, v10.8h\n"
+    "fmla v25.8h, v5.8h, v10.8h\n"
+    "fmla v26.8h, v4.8h, v10.8h\n"
+    "fmla v27.8h, v3.8h, v10.8h\n"
+    "fmla v29.8h, v2.8h, v10.8h\n"
+    "fmla v30.8h, v1.8h, v10.8h\n"
+    "fmla v31.8h, v0.8h, v10.8h\n"
+    "ldr q10, [x14, x16]\n"
+    "fmla v20.8h, v0.8h, v9.8h\n"
+    "fmla v24.8h, v6.8h, v11.8h\n"
+    "fmla v28.8h, v3.8h, v11.8h\n"
+    "fmla v21.8h, v1.8h, v10.8h\n"
+    "ldr q11, [x28, x24]\n"
+    "fmla v16.8h, v3.8h, v9.8h\n"
+    "fmla v19.8h, v5.8h, v12.8h\n"
+    "fmla v23.8h, v2.8h, v12.8h\n"
+    "fmla v17.8h, v4.8h, v10.8h\n"
+    "ldr q12, [x14, x10]\n"
+    "fmla v18.8h, v3.8h, v10.8h\n"
     "fmla v22.8h, v0.8h, v10.8h\n"
-    "ldr q10, [x8, x10]\n"
-    "fmla v17.8h, v8.8h, v11.8h\n"
-    "fmla v16.8h, v7.8h, v11.8h\n"
-    "ldr q11, [x13, x6]\n"
-    "fmla v29.8h, v8.8h, v12.8h\n"
-    "fmla v28.8h, v7.8h, v12.8h\n"
-    "fmla v25.8h, v5.8h, v12.8h\n"
-    "fmla v24.8h, v4.8h, v12.8h\n"
+    "fmla v27.8h, v8.8h, v11.8h\n"
+    "fmla v31.8h, v5.8h, v11.8h\n"
+    "ldr q11, [x25, x5]\n"
+    "fmla v20.8h, v2.8h, v10.8h\n"
     "fmla v21.8h, v2.8h, v12.8h\n"
-    "fmla v20.8h, v1.8h, v12.8h\n"
-    "ldr q12, [x8, x9]\n"
+    "fmla v16.8h, v5.8h, v10.8h\n"
+    "fmla v17.8h, v5.8h, v12.8h\n"
+    "ldr q10, [x13, x5]\n"
+    "fmla v18.8h, v4.8h, v12.8h\n"
+    "fmla v19.8h, v3.8h, v12.8h\n"
+    "fmla v22.8h, v1.8h, v12.8h\n"
+    "fmla v23.8h, v0.8h, v12.8h\n"
+    "ldr q12, [x13, x27]\n"
+    "fmla v28.8h, v7.8h, v11.8h\n"
+    "fmla v29.8h, v6.8h, v11.8h\n"
+    "ldr q11, [x25, x27]\n"
+    "fmla v20.8h, v4.8h, v10.8h\n"
+    "fmla v21.8h, v3.8h, v10.8h\n"
+    "fmla v24.8h, v1.8h, v10.8h\n"
+    "fmla v25.8h, v0.8h, v10.8h\n"
+    "fmla v16.8h, v7.8h, v10.8h\n"
+    "fmla v17.8h, v6.8h, v10.8h\n"
+    "ldr q10, [x8, x16]\n"
+    "fmla v30.8h, v8.8h, v11.8h\n"
+    "fmla v31.8h, v7.8h, v11.8h\n"
+    "ldr q11, [x11, x5]\n"
+    "fmla v18.8h, v8.8h, v12.8h\n"
+    "fmla v19.8h, v7.8h, v12.8h\n"
+    "fmla v22.8h, v5.8h, v12.8h\n"
+    "fmla v23.8h, v4.8h, v12.8h\n"
+    "fmla v26.8h, v2.8h, v12.8h\n"
+    "fmla v27.8h, v1.8h, v12.8h\n"
+    "ldr q12, [x8, x10]\n"
     "add x8, x8, #0x10\n"
-    "fmla v31.8h, v2.8h, v10.8h\n"
-    "fmla v30.8h, v1.8h, v10.8h\n"
-    "fmla v29.8h, v0.8h, v10.8h\n"
-    "ld1 { v10.8h }, [x14]\n"
-    "fmla v27.8h, v7.8h, v11.8h\n"
-    "fmla v26.8h, v6.8h, v11.8h\n"
-    "fmla v23.8h, v4.8h, v11.8h\n"
-    "fmla v22.8h, v3.8h, v11.8h\n"
-    "fmla v19.8h, v1.8h, v11.8h\n"
-    "fmla v18.8h, v0.8h, v11.8h\n"
-    "ldr q11, [x13, x28]\n"
-    "fmla v30.8h, v2.8h, v12.8h\n"
-    "fmla v29.8h, v1.8h, v12.8h\n"
-    "fmla v28.8h, v0.8h, v12.8h\n"
-    "ldr q12, [x14, x27]\n"
-    "add x14, x14, #0x10\n"
-    "fmla v31.8h, v6.8h, v10.8h\n"
-    "ldr q9, [x14, x10]\n"
-    "fmla v27.8h, v3.8h, v10.8h\n"
-    "fmla v23.8h, v0.8h, v10.8h\n"
+    "fmla v20.8h, v7.8h, v11.8h\n"
+    "fmla v21.8h, v6.8h, v11.8h\n"
+    "fmla v24.8h, v4.8h, v11.8h\n"
+    "fmla v25.8h, v3.8h, v11.8h\n"
+    "fmla v28.8h, v1.8h, v11.8h\n"
+    "fmla v29.8h, v0.8h, v11.8h\n"
+    "ldr q11, [x11, x27]\n"
+    "fmla v16.8h, v2.8h, v10.8h\n"
+    "fmla v17.8h, v1.8h, v10.8h\n"
+    "fmla v18.8h, v0.8h, v10.8h\n"
     "ld1 { v10.8h }, [x13]\n"
-    "fmla v25.8h, v8.8h, v11.8h\n"
-    "fmla v24.8h, v7.8h, v11.8h\n"
-    "fmla v21.8h, v5.8h, v11.8h\n"
-    "fmla v20.8h, v4.8h, v11.8h\n"
-    "fmla v17.8h, v2.8h, v11.8h\n"
-    "fmla v16.8h, v1.8h, v11.8h\n"
-    "ldr q11, [x12, x10]\n"
-    "fmla v28.8h, v8.8h, v12.8h\n"
-    "fmla v24.8h, v5.8h, v12.8h\n"
-    "fmla v20.8h, v2.8h, v12.8h\n"
-    "ldr q12, [x13, x27]\n"
+    "fmla v30.8h, v2.8h, v11.8h\n"
+    "fmla v19.8h, v0.8h, v12.8h\n"
+    "fmla v20.8h, v3.8h, v10.8h\n"
+    "fmla v24.8h, v0.8h, v10.8h\n"
+    "fmla v22.8h, v8.8h, v11.8h\n"
+    "fmla v23.8h, v7.8h, v11.8h\n"
+    "fmla v26.8h, v5.8h, v11.8h\n"
+    "fmla v27.8h, v4.8h, v11.8h\n"
+    "fmla v31.8h, v1.8h, v11.8h\n"
+    "ldr q11, [x28, x16]\n"
+    "fmla v17.8h, v2.8h, v12.8h\n"
+    "fmla v18.8h, v1.8h, v12.8h\n"
+    "ldr q12, [x13, x24]\n"
     "add x13, x13, #0x10\n"
-    "fmla v27.8h, v6.8h, v10.8h\n"
-    "fmla v23.8h, v3.8h, v10.8h\n"
-    "fmla v19.8h, v0.8h, v10.8h\n"
-    "ldr q10, [x11, x10]\n"
-    "fmla v22.8h, v7.8h, v11.8h\n"
-    "fmla v21.8h, v6.8h, v11.8h\n"
-    "fmla v23.8h, v8.8h, v11.8h\n"
-    "fmla v19.8h, v5.8h, v11.8h\n"
-    "fmla v18.8h, v4.8h, v11.8h\n"
-    "fmla v17.8h, v3.8h, v11.8h\n"
-    "ldr q11, [x12, x9]\n"
-    "fmla v24.8h, v8.8h, v12.8h\n"
-    "fmla v20.8h, v5.8h, v12.8h\n"
-    "fmla v16.8h, v2.8h, v12.8h\n"
-    "ldr q12, [x11, x9]\n"
+    "fmla v16.8h, v6.8h, v10.8h\n"
+    "ld1 { v10.8h }, [x11]\n"
+    "fmla v29.8h, v4.8h, v11.8h\n"
+    "ldr q9, [x13, x16]\n"
+    "fmla v30.8h, v3.8h, v11.8h\n"
+    "fmla v19.8h, v8.8h, v12.8h\n"
+    "fmla v23.8h, v5.8h, v12.8h\n"
+    "fmla v27.8h, v2.8h, v12.8h\n"
+    "ldr q12, [x11, x24]\n"
     "add x11, x11, #0x10\n"
-    "fmla v19.8h, v8.8h, v10.8h\n"
-    "fmla v18.8h, v7.8h, v10.8h\n"
-    "fmla v17.8h, v6.8h, v10.8h\n"
-    "ldr q10, [x15, x6]\n"
-    "fmla v22.8h, v8.8h, v11.8h\n"
-    "fmla v21.8h, v7.8h, v11.8h\n"
-    "fmla v20.8h, v6.8h, v11.8h\n"
-    "fmla v18.8h, v5.8h, v11.8h\n"
-    "fmla v17.8h, v4.8h, v11.8h\n"
-    "fmla v16.8h, v3.8h, v11.8h\n"
-    "ldr q11, [x15, x28]\n"
-    "add x15, x15, #0x10\n"
-    "fmla v18.8h, v8.8h, v12.8h\n"
-    "fmla v31.8h, v4.8h, v10.8h\n"
-    "fmla v17.8h, v7.8h, v12.8h\n"
-    "fmla v16.8h, v6.8h, v12.8h\n"
-    "ldr q12, [x12, x6]\n"
-    "fmla v30.8h, v3.8h, v10.8h\n"
-    "fmla v27.8h, v1.8h, v10.8h\n"
-    "fmla v26.8h, v0.8h, v10.8h\n"
-    "ldr q10, [x12, x28]\n"
-    "add x12, x12, #0x10\n"
+    "fmla v20.8h, v6.8h, v10.8h\n"
+    "fmla v24.8h, v3.8h, v10.8h\n"
+    "fmla v28.8h, v0.8h, v10.8h\n"
+    "ldr q10, [x25, x16]\n"
+    "fmla v31.8h, v2.8h, v12.8h\n"
+    "fmla v29.8h, v7.8h, v10.8h\n"
+    "fmla v30.8h, v6.8h, v10.8h\n"
+    "fmla v24.8h, v8.8h, v11.8h\n"
+    "fmla v25.8h, v7.8h, v11.8h\n"
+    "fmla v26.8h, v6.8h, v11.8h\n"
+    "fmla v28.8h, v5.8h, v11.8h\n"
+    "ldr q11, [x28, x10]\n"
+    "fmla v27.8h, v5.8h, v12.8h\n"
     "fmla v29.8h, v5.8h, v11.8h\n"
-    "ldr q0, [x5, #0x10]\n"
-    "fmla v28.8h, v4.8h, v11.8h\n"
-    "fmla v25.8h, v2.8h, v11.8h\n"
-    "ldr q2, [x5, #0x30]\n"
-    "fmla v24.8h, v1.8h, v11.8h\n"
-    "ldr q11, [x8, x27]\n"
-    "fmla v23.8h, v7.8h, v12.8h\n"
-    "ldr q1, [x5, #0x20]\n"
-    "fmla v22.8h, v6.8h, v12.8h\n"
-    "ldr q6, [x5, #0x70]\n"
-    "fmla v19.8h, v4.8h, v12.8h\n"
-    "fmla v18.8h, v3.8h, v12.8h\n"
-    "ldr q12, [x14, x9]\n"
-    "fmla v21.8h, v8.8h, v10.8h\n"
-    "ldr q3, [x5, #0x40]\n"
-    "fmla v20.8h, v7.8h, v10.8h\n"
-    "ldr q7, [x5, #0x80]\n"
-    "fmla v17.8h, v5.8h, v10.8h\n"
-    "ldr q5, [x5, #0x60]\n"
+    "fmla v30.8h, v4.8h, v11.8h\n"
+    "fmla v31.8h, v3.8h, v11.8h\n"
+    "fmla v23.8h, v8.8h, v12.8h\n"
+    "ldr q12, [x25, x10]\n"
+    "fmla v28.8h, v8.8h, v10.8h\n"
+    "ldr q10, [x14, x5]\n"
+    "fmla v25.8h, v8.8h, v11.8h\n"
+    "fmla v26.8h, v7.8h, v11.8h\n"
+    "add x25, x25, #0x10\n"
+    "fmla v27.8h, v6.8h, v11.8h\n"
+    "fmla v29.8h, v8.8h, v12.8h\n"
+    "ldr q11, [x14, x27]\n"
+    "add x14, x14, #0x10\n"
+    "fmla v30.8h, v7.8h, v12.8h\n"
+    "fmla v31.8h, v6.8h, v12.8h\n"
+    "ldr q12, [x28, x5]\n"
     "fmla v16.8h, v4.8h, v10.8h\n"
+    "fmla v17.8h, v3.8h, v10.8h\n"
+    "fmax v16.8h, v16.8h, v15.8h\n"
+    "fmla v20.8h, v1.8h, v10.8h\n"
+    "fmla v21.8h, v0.8h, v10.8h\n"
+    "ldr q10, [x28, x27]\n"
+    "fmax v17.8h, v17.8h, v15.8h\n"
+    "fmla v18.8h, v5.8h, v11.8h\n"
+    "fmla v19.8h, v4.8h, v11.8h\n"
+    "fmax v18.8h, v18.8h, v15.8h\n"
+    "add x28, x28, #0x10\n"
+    "fmla v22.8h, v2.8h, v11.8h\n"
+    "fmla v23.8h, v1.8h, v11.8h\n"
+    "fmax v19.8h, v19.8h, v15.8h\n"
+    "ldr q11, [x8, x24]\n"
+    "fmla v24.8h, v7.8h, v12.8h\n"
+    "fmla v25.8h, v6.8h, v12.8h\n"
+    "fmax v20.8h, v20.8h, v15.8h\n"
+    "ldr q0, [x15, #0x10]\n"
+    "fmla v28.8h, v4.8h, v12.8h\n"
+    "fmla v29.8h, v3.8h, v12.8h\n"
+    "fmax v21.8h, v21.8h, v15.8h\n"
+    "ldr q12, [x13, x10]\n"
+    "fmla v26.8h, v8.8h, v10.8h\n"
+    "fmla v27.8h, v7.8h, v10.8h\n"
+    "fmax v22.8h, v22.8h, v15.8h\n"
+    "ldr q1, [x15, #0x20]\n"
+    "fmla v30.8h, v5.8h, v10.8h\n"
+    "fmla v31.8h, v4.8h, v10.8h\n"
+    "fmax v23.8h, v23.8h, v15.8h\n"
     "ld1 { v10.8h }, [x8]\n"
-    "fmax v31.8h, v31.8h, v15.8h\n"
-    "ldr q4, [x5, #0x50]\n"
-    "fmax v30.8h, v30.8h, v15.8h\n"
-    "ldr q8, [x5, #0x90]\n"
-    "add x5, x5, #0xa0\n"
-    "fmin v31.8h, v31.8h, v14.8h\n"
-    "st1 { v31.8h }, [x16]\n"
-    "fmin v30.8h, v30.8h, v14.8h\n"
-    "fmax v29.8h, v29.8h, v15.8h\n"
-    "str q30, [x16, x17]\n"
-    "fmin v29.8h, v29.8h, v14.8h\n"
-    "fmax v28.8h, v28.8h, v15.8h\n"
-    "str q29, [x16, x23]\n"
-    "fmin v28.8h, v28.8h, v14.8h\n"
-    "fmax v27.8h, v27.8h, v15.8h\n"
-    "str q28, [x16, x22]\n"
-    "fmin v27.8h, v27.8h, v14.8h\n"
-    "add x16, x16, #0x10\n"
-    "fmax v26.8h, v26.8h, v15.8h\n"
-    "st1 { v27.8h }, [x26]\n"
-    "fmax v25.8h, v25.8h, v15.8h\n"
     "fmax v24.8h, v24.8h, v15.8h\n"
-    "fmin v26.8h, v26.8h, v14.8h\n"
-    "str q26, [x26, x17]\n"
-    "fmin v25.8h, v25.8h, v14.8h\n"
+    "fmax v25.8h, v25.8h, v15.8h\n"
+    "ldr q2, [x15, #0x30]\n"
+    "ldr q3, [x15, #0x40]\n"
+    "fmax v26.8h, v26.8h, v15.8h\n"
+    "fmax v27.8h, v27.8h, v15.8h\n"
+    "ldr q4, [x15, #0x50]\n"
+    "ldr q5, [x15, #0x60]\n"
+    "fmax v28.8h, v28.8h, v15.8h\n"
+    "fmax v29.8h, v29.8h, v15.8h\n"
+    "ldr q6, [x15, #0x70]\n"
+    "ldr q7, [x15, #0x80]\n"
+    "fmax v30.8h, v30.8h, v15.8h\n"
+    "fmax v31.8h, v31.8h, v15.8h\n"
+    "ldr q8, [x15, #0x90]\n"
+    "add x15, x15, #0xa0\n"
+    "fmin v16.8h, v16.8h, v14.8h\n"
+    "fmin v17.8h, v17.8h, v14.8h\n"
+    "st1 { v16.8h }, [x17]\n"
+    "fmin v18.8h, v18.8h, v14.8h\n"
+    "fmin v19.8h, v19.8h, v14.8h\n"
+    "str q17, [x17, x6]\n"
+    "fmin v20.8h, v20.8h, v14.8h\n"
+    "fmin v21.8h, v21.8h, v14.8h\n"
+    "str q18, [x17, x22]\n"
+    "fmin v22.8h, v22.8h, v14.8h\n"
+    "fmin v23.8h, v23.8h, v14.8h\n"
+    "str q19, [x17, x21]\n"
+    "add x17, x17, #0x10\n"
     "fmin v24.8h, v24.8h, v14.8h\n"
-    "str q25, [x26, x23]\n"
-    "fmax v23.8h, v23.8h, v15.8h\n"
-    "fmax v22.8h, v22.8h, v15.8h\n"
-    "str q24, [x26, x22]\n"
+    "fmin v25.8h, v25.8h, v14.8h\n"
+    "st1 { v20.8h }, [x9]\n"
+    "fmin v26.8h, v26.8h, v14.8h\n"
+    "fmin v27.8h, v27.8h, v14.8h\n"
+    "str q21, [x9, x6]\n"
+    "fmin v28.8h, v28.8h, v14.8h\n"
+    "fmin v29.8h, v29.8h, v14.8h\n"
+    "str q22, [x9, x22]\n"
+    "fmin v30.8h, v30.8h, v14.8h\n"
+    "fmin v31.8h, v31.8h, v14.8h\n"
+    "str q23, [x9, x21]\n"
+    "add x9, x9, #0x10\n"
+    "st1 { v24.8h }, [x26]\n"
+    "str q25, [x26, x6]\n"
+    "str q26, [x26, x22]\n"
+    "str q27, [x26, x21]\n"
     "add x26, x26, #0x10\n"
-    "fmax v21.8h, v21.8h, v15.8h\n"
-    "fmax v20.8h, v20.8h, v15.8h\n"
-    "fmin v23.8h, v23.8h, v14.8h\n"
-    "st1 { v23.8h }, [x25]\n"
-    "fmin v22.8h, v22.8h, v14.8h\n"
-    "fmin v21.8h, v21.8h, v14.8h\n"
-    "str q22, [x25, x17]\n"
-    "fmin v20.8h, v20.8h, v14.8h\n"
-    "fmax v19.8h, v19.8h, v15.8h\n"
-    "str q21, [x25, x23]\n"
-    "fmax v18.8h, v18.8h, v15.8h\n"
-    "str q20, [x25, x22]\n"
-    "fmin v19.8h, v19.8h, v14.8h\n"
-    "add x25, x25, #0x10\n"
-    "fmin v18.8h, v18.8h, v14.8h\n"
-    "st1 { v19.8h }, [x24]\n"
-    "fmax v17.8h, v17.8h, v15.8h\n"
-    "fmax v16.8h, v16.8h, v15.8h\n"
-    "str q18, [x24, x17]\n"
-    "fmin v17.8h, v17.8h, v14.8h\n"
-    "str q17, [x24, x23]\n"
-    "fmin v16.8h, v16.8h, v14.8h\n"
-    "str q16, [x24, x22]\n"
-    "add x24, x24, #0x10\n"
+    "st1 { v28.8h }, [x23]\n"
+    "str q29, [x23, x6]\n"
+    "str q30, [x23, x22]\n"
+    "str q31, [x23, x21]\n"
+    "add x23, x23, #0x10\n"
     "blt 2b\n"
     "3:"  // Tile loop: Channel tail
-    "mov v31.16b, v13.16b\n fmla v31.8h, v8.8h, v9.8h\n"
-    "mov v30.16b, v13.16b\n fmla v30.8h, v7.8h, v9.8h\n"
-    "mov v29.16b, v13.16b\n fmla v29.8h, v6.8h, v9.8h\n"
-    "mov v27.16b, v13.16b\n fmla v27.8h, v5.8h, v9.8h\n"
-    "mov v26.16b, v13.16b\n fmla v26.8h, v4.8h, v9.8h\n"
-    "mov v25.16b, v13.16b\n fmla v25.8h, v3.8h, v9.8h\n"
-    "mov v23.16b, v13.16b\n fmla v23.8h, v2.8h, v9.8h\n"
-    "mov v22.16b, v13.16b\n fmla v22.8h, v1.8h, v9.8h\n"
-    "mov v21.16b, v13.16b\n fmla v21.8h, v0.8h, v9.8h\n"
-    "ldr q9, [x13, x10]\n"
-    "fmla v31.8h, v0.8h, v10.8h\n"
-    "ld1 { v10.8h }, [x11]\n"
-    "mov v28.16b, v13.16b\n fmla v28.8h, v2.8h, v11.8h\n"
-    "ldr q11, [x11, x27]\n"
-    "fmla v30.8h, v8.8h, v12.8h\n"
-    "fmla v29.8h, v7.8h, v12.8h\n"
-    "fmla v26.8h, v5.8h, v12.8h\n"
-    "fmla v28.8h, v6.8h, v12.8h\n"
-    "fmla v25.8h, v4.8h, v12.8h\n"
-    "mov v24.16b, v13.16b\n fmla v24.8h, v3.8h, v12.8h\n"
-    "fmla v22.8h, v2.8h, v12.8h\n"
-    "fmla v21.8h, v1.8h, v12.8h\n"
-    "mov v20.16b, v13.16b\n fmla v20.8h, v0.8h, v12.8h\n"
-    "ldr q12, [x8, x6]\n"
-    "mov v19.16b, v13.16b\n fmla v19.8h, v6.8h, v10.8h\n"
-    "ldr q10, [x13, x9]\n"
-    "mov v16.16b, v13.16b\n fmla v16.8h, v8.8h, v11.8h\n"
-    "ldr q11, [x8, x28]\n"
-    "fmla v27.8h, v8.8h, v9.8h\n"
-    "fmla v26.8h, v7.8h, v9.8h\n"
-    "fmla v25.8h, v6.8h, v9.8h\n"
-    "fmla v23.8h, v5.8h, v9.8h\n"
-    "fmla v22.8h, v4.8h, v9.8h\n"
-    "fmla v21.8h, v3.8h, v9.8h\n"
-    "fmla v19.8h, v2.8h, v9.8h\n"
-    "mov v18.16b, v13.16b\n fmla v18.8h, v1.8h, v9.8h\n"
-    "mov v17.16b, v13.16b\n fmla v17.8h, v0.8h, v9.8h\n"
-    "ld1 { v9.8h }, [x15]\n"
-    "fmla v31.8h, v1.8h, v12.8h\n"
-    "fmla v30.8h, v0.8h, v12.8h\n"
-    "ldr q12, [x15, x27]\n"
-    "fmla v29.8h, v2.8h, v11.8h\n"
-    "fmla v28.8h, v1.8h, v11.8h\n"
-    "ld1 { v11.8h }, [x12]\n"
-    "fmla v26.8h, v8.8h, v10.8h\n"
-    "fmla v25.8h, v7.8h, v10.8h\n"
-    "fmla v24.8h, v6.8h, v10.8h\n"
-    "fmla v22.8h, v5.8h, v10.8h\n"
-    "fmla v21.8h, v4.8h, v10.8h\n"
-    "fmla v20.8h, v3.8h, v10.8h\n"
-    "fmla v18.8h, v2.8h, v10.8h\n"
-    "fmla v17.8h, v1.8h, v10.8h\n"
+    "mov v21.16b, v13.16b\n fmla v21.8h, v4.8h, v9.8h\n"
+    "mov v16.16b, v13.16b\n fmla v16.8h, v8.8h, v9.8h\n"
+    "mov v22.16b, v13.16b\n fmla v22.8h, v3.8h, v9.8h\n"
+    "mov v25.16b, v13.16b\n fmla v25.8h, v1.8h, v9.8h\n"
+    "mov v26.16b, v13.16b\n fmla v26.8h, v0.8h, v9.8h\n"
+    "fmla v21.8h, v5.8h, v12.8h\n"
+    "mov v17.16b, v13.16b\n fmla v17.8h, v7.8h, v9.8h\n"
+    "mov v18.16b, v13.16b\n fmla v18.8h, v6.8h, v9.8h\n"
+    "mov v20.16b, v13.16b\n fmla v20.8h, v5.8h, v9.8h\n"
+    "mov v24.16b, v13.16b\n fmla v24.8h, v2.8h, v9.8h\n"
+    "ldr q9, [x11, x16]\n"
     "fmla v16.8h, v0.8h, v10.8h\n"
-    "ldr q10, [x15, x10]\n"
-    "fmla v31.8h, v3.8h, v9.8h\n"
-    "fmla v27.8h, v0.8h, v9.8h\n"
-    "fmla v28.8h, v5.8h, v12.8h\n"
-    "fmla v24.8h, v2.8h, v12.8h\n"
-    "ldr q12, [x15, x9]\n"
-    "fmla v23.8h, v6.8h, v11.8h\n"
-    "fmla v19.8h, v3.8h, v11.8h\n"
-    "ldr q11, [x12, x27]\n"
-    "fmla v31.8h, v5.8h, v10.8h\n"
-    "fmla v30.8h, v4.8h, v10.8h\n"
-    "fmla v29.8h, v3.8h, v10.8h\n"
-    "fmla v27.8h, v2.8h, v10.8h\n"
-    "fmla v26.8h, v1.8h, v10.8h\n"
-    "fmla v25.8h, v0.8h, v10.8h\n"
-    "ldr q10, [x14, x6]\n"
-    "fmla v20.8h, v8.8h, v11.8h\n"
-    "fmla v16.8h, v5.8h, v11.8h\n"
-    "ldr q11, [x11, x6]\n"
-    "fmla v30.8h, v5.8h, v12.8h\n"
-    "fmla v29.8h, v4.8h, v12.8h\n"
-    "fmla v28.8h, v3.8h, v12.8h\n"
-    "fmla v26.8h, v2.8h, v12.8h\n"
-    "fmla v25.8h, v1.8h, v12.8h\n"
-    "fmla v24.8h, v0.8h, v12.8h\n"
-    "ldr q12, [x14, x28]\n"
-    "fmla v19.8h, v7.8h, v11.8h\n"
-    "fmla v18.8h, v6.8h, v11.8h\n"
-    "ldr q11, [x11, x28]\n"
-    "fmla v31.8h, v7.8h, v10.8h\n"
-    "fmla v30.8h, v6.8h, v10.8h\n"
-    "fmla v27.8h, v4.8h, v10.8h\n"
-    "fmla v26.8h, v3.8h, v10.8h\n"
-    "fmla v23.8h, v1.8h, v10.8h\n"
+    "mov v19.16b, v13.16b\n fmla v19.8h, v2.8h, v11.8h\n"
+    "ld1 { v10.8h }, [x25]\n"
+    "ldr q11, [x25, x24]\n"
+    "fmla v22.8h, v4.8h, v12.8h\n"
+    "fmla v25.8h, v2.8h, v12.8h\n"
+    "fmla v26.8h, v1.8h, v12.8h\n"
+    "mov v28.16b, v13.16b\n fmla v28.8h, v6.8h, v10.8h\n"
+    "ldr q10, [x11, x10]\n"
+    "fmla v21.8h, v7.8h, v9.8h\n"
+    "fmla v17.8h, v8.8h, v12.8h\n"
+    "fmla v18.8h, v7.8h, v12.8h\n"
+    "fmla v19.8h, v6.8h, v12.8h\n"
+    "mov v23.16b, v13.16b\n fmla v23.8h, v3.8h, v12.8h\n"
+    "mov v27.16b, v13.16b\n fmla v27.8h, v0.8h, v12.8h\n"
+    "ldr q12, [x8, x5]\n"
+    "mov v31.16b, v13.16b\n fmla v31.8h, v8.8h, v11.8h\n"
+    "fmla v22.8h, v6.8h, v9.8h\n"
+    "ldr q11, [x8, x27]\n"
+    "fmla v25.8h, v4.8h, v9.8h\n"
+    "fmla v26.8h, v3.8h, v9.8h\n"
+    "mov v29.16b, v13.16b\n fmla v29.8h, v1.8h, v9.8h\n"
+    "mov v30.16b, v13.16b\n fmla v30.8h, v0.8h, v9.8h\n"
+    "fmla v20.8h, v8.8h, v9.8h\n"
+    "fmla v24.8h, v5.8h, v9.8h\n"
+    "fmla v28.8h, v2.8h, v9.8h\n"
+    "fmla v21.8h, v8.8h, v10.8h\n"
+    "ld1 { v9.8h }, [x14]\n"
+    "fmla v16.8h, v1.8h, v12.8h\n"
+    "fmla v17.8h, v0.8h, v12.8h\n"
+    "ldr q12, [x14, x24]\n"
+    "fmla v18.8h, v2.8h, v11.8h\n"
+    "fmla v19.8h, v1.8h, v11.8h\n"
+    "ld1 { v11.8h }, [x28]\n"
+    "fmla v22.8h, v7.8h, v10.8h\n"
+    "fmla v23.8h, v6.8h, v10.8h\n"
+    "fmla v25.8h, v5.8h, v10.8h\n"
+    "fmla v26.8h, v4.8h, v10.8h\n"
+    "fmla v27.8h, v3.8h, v10.8h\n"
+    "fmla v29.8h, v2.8h, v10.8h\n"
+    "fmla v30.8h, v1.8h, v10.8h\n"
+    "fmla v31.8h, v0.8h, v10.8h\n"
+    "ldr q10, [x14, x16]\n"
+    "fmla v20.8h, v0.8h, v9.8h\n"
+    "fmla v24.8h, v6.8h, v11.8h\n"
+    "fmla v28.8h, v3.8h, v11.8h\n"
+    "fmla v21.8h, v1.8h, v10.8h\n"
+    "ldr q11, [x28, x24]\n"
+    "fmla v16.8h, v3.8h, v9.8h\n"
+    "fmla v19.8h, v5.8h, v12.8h\n"
+    "fmla v23.8h, v2.8h, v12.8h\n"
+    "fmla v17.8h, v4.8h, v10.8h\n"
+    "ldr q12, [x14, x10]\n"
+    "fmla v18.8h, v3.8h, v10.8h\n"
     "fmla v22.8h, v0.8h, v10.8h\n"
-    "ldr q10, [x8, x10]\n"
-    "fmla v17.8h, v8.8h, v11.8h\n"
-    "fmla v16.8h, v7.8h, v11.8h\n"
-    "ldr q11, [x13, x6]\n"
-    "fmla v29.8h, v8.8h, v12.8h\n"
-    "fmla v28.8h, v7.8h, v12.8h\n"
-    "fmla v25.8h, v5.8h, v12.8h\n"
-    "fmla v24.8h, v4.8h, v12.8h\n"
+    "fmla v27.8h, v8.8h, v11.8h\n"
+    "fmla v31.8h, v5.8h, v11.8h\n"
+    "ldr q11, [x25, x5]\n"
+    "fmla v20.8h, v2.8h, v10.8h\n"
     "fmla v21.8h, v2.8h, v12.8h\n"
-    "fmla v20.8h, v1.8h, v12.8h\n"
-    "ldr q12, [x8, x9]\n"
+    "fmla v16.8h, v5.8h, v10.8h\n"
+    "fmla v17.8h, v5.8h, v12.8h\n"
+    "ldr q10, [x13, x5]\n"
+    "fmla v18.8h, v4.8h, v12.8h\n"
+    "fmla v19.8h, v3.8h, v12.8h\n"
+    "fmla v22.8h, v1.8h, v12.8h\n"
+    "fmla v23.8h, v0.8h, v12.8h\n"
+    "ldr q12, [x13, x27]\n"
+    "fmla v28.8h, v7.8h, v11.8h\n"
+    "fmla v29.8h, v6.8h, v11.8h\n"
+    "ldr q11, [x25, x27]\n"
+    "fmla v20.8h, v4.8h, v10.8h\n"
+    "fmla v21.8h, v3.8h, v10.8h\n"
+    "fmla v24.8h, v1.8h, v10.8h\n"
+    "fmla v25.8h, v0.8h, v10.8h\n"
+    "fmla v16.8h, v7.8h, v10.8h\n"
+    "fmla v17.8h, v6.8h, v10.8h\n"
+    "ldr q10, [x8, x16]\n"
+    "fmla v30.8h, v8.8h, v11.8h\n"
+    "fmla v31.8h, v7.8h, v11.8h\n"
+    "ldr q11, [x11, x5]\n"
+    "fmla v18.8h, v8.8h, v12.8h\n"
+    "fmla v19.8h, v7.8h, v12.8h\n"
+    "fmla v22.8h, v5.8h, v12.8h\n"
+    "fmla v23.8h, v4.8h, v12.8h\n"
+    "fmla v26.8h, v2.8h, v12.8h\n"
+    "fmla v27.8h, v1.8h, v12.8h\n"
+    "ldr q12, [x8, x10]\n"
     "add x8, x8, #0x10\n"
-    "fmla v31.8h, v2.8h, v10.8h\n"
-    "fmla v30.8h, v1.8h, v10.8h\n"
-    "fmla v29.8h, v0.8h, v10.8h\n"
-    "ld1 { v10.8h }, [x14]\n"
-    "fmla v27.8h, v7.8h, v11.8h\n"
-    "fmla v26.8h, v6.8h, v11.8h\n"
-    "fmla v23.8h, v4.8h, v11.8h\n"
-    "fmla v22.8h, v3.8h, v11.8h\n"
-    "fmla v19.8h, v1.8h, v11.8h\n"
-    "fmla v18.8h, v0.8h, v11.8h\n"
-    "ldr q11, [x13, x28]\n"
-    "fmla v30.8h, v2.8h, v12.8h\n"
-    "fmla v29.8h, v1.8h, v12.8h\n"
-    "fmla v28.8h, v0.8h, v12.8h\n"
-    "ldr q12, [x14, x27]\n"
-    "add x14, x14, #0x10\n"
-    "fmla v31.8h, v6.8h, v10.8h\n"
-    "fmla v27.8h, v3.8h, v10.8h\n"
-    "fmla v23.8h, v0.8h, v10.8h\n"
+    "fmla v20.8h, v7.8h, v11.8h\n"
+    "fmla v21.8h, v6.8h, v11.8h\n"
+    "fmla v24.8h, v4.8h, v11.8h\n"
+    "fmla v25.8h, v3.8h, v11.8h\n"
+    "fmla v28.8h, v1.8h, v11.8h\n"
+    "fmla v29.8h, v0.8h, v11.8h\n"
+    "ldr q11, [x11, x27]\n"
+    "fmla v16.8h, v2.8h, v10.8h\n"
+    "fmla v17.8h, v1.8h, v10.8h\n"
+    "fmla v18.8h, v0.8h, v10.8h\n"
     "ld1 { v10.8h }, [x13]\n"
-    "fmla v25.8h, v8.8h, v11.8h\n"
-    "fmla v24.8h, v7.8h, v11.8h\n"
-    "fmla v21.8h, v5.8h, v11.8h\n"
-    "fmla v20.8h, v4.8h, v11.8h\n"
-    "fmla v17.8h, v2.8h, v11.8h\n"
-    "fmla v16.8h, v1.8h, v11.8h\n"
-    "ldr q11, [x12, x10]\n"
-    "fmla v28.8h, v8.8h, v12.8h\n"
-    "fmla v24.8h, v5.8h, v12.8h\n"
-    "fmla v20.8h, v2.8h, v12.8h\n"
-    "ldr q12, [x13, x27]\n"
+    "fmla v30.8h, v2.8h, v11.8h\n"
+    "fmla v19.8h, v0.8h, v12.8h\n"
+    "fmla v20.8h, v3.8h, v10.8h\n"
+    "fmla v24.8h, v0.8h, v10.8h\n"
+    "fmla v22.8h, v8.8h, v11.8h\n"
+    "fmla v23.8h, v7.8h, v11.8h\n"
+    "fmla v26.8h, v5.8h, v11.8h\n"
+    "fmla v27.8h, v4.8h, v11.8h\n"
+    "fmla v31.8h, v1.8h, v11.8h\n"
+    "ldr q11, [x28, x16]\n"
+    "fmla v17.8h, v2.8h, v12.8h\n"
+    "fmla v18.8h, v1.8h, v12.8h\n"
+    "ldr q12, [x13, x24]\n"
     "add x13, x13, #0x10\n"
-    "fmla v27.8h, v6.8h, v10.8h\n"
-    "fmla v23.8h, v3.8h, v10.8h\n"
-    "fmla v19.8h, v0.8h, v10.8h\n"
-    "ldr q10, [x11, x10]\n"
-    "fmla v22.8h, v7.8h, v11.8h\n"
-    "fmla v21.8h, v6.8h, v11.8h\n"
-    "fmla v23.8h, v8.8h, v11.8h\n"
-    "fmla v19.8h, v5.8h, v11.8h\n"
-    "fmla v18.8h, v4.8h, v11.8h\n"
-    "fmla v17.8h, v3.8h, v11.8h\n"
-    "ldr q11, [x12, x9]\n"
-    "fmla v24.8h, v8.8h, v12.8h\n"
-    "fmla v20.8h, v5.8h, v12.8h\n"
-    "fmla v16.8h, v2.8h, v12.8h\n"
-    "ldr q12, [x11, x9]\n"
+    "fmla v16.8h, v6.8h, v10.8h\n"
+    "ld1 { v10.8h }, [x11]\n"
+    "fmla v29.8h, v4.8h, v11.8h\n"
+    "fmla v30.8h, v3.8h, v11.8h\n"
+    "fmla v19.8h, v8.8h, v12.8h\n"
+    "fmla v23.8h, v5.8h, v12.8h\n"
+    "fmla v27.8h, v2.8h, v12.8h\n"
+    "ldr q12, [x11, x24]\n"
     "add x11, x11, #0x10\n"
-    "fmla v19.8h, v8.8h, v10.8h\n"
-    "fmla v18.8h, v7.8h, v10.8h\n"
-    "fmla v17.8h, v6.8h, v10.8h\n"
-    "ldr q10, [x15, x6]\n"
-    "fmla v22.8h, v8.8h, v11.8h\n"
-    "fmla v21.8h, v7.8h, v11.8h\n"
-    "fmla v20.8h, v6.8h, v11.8h\n"
-    "fmla v18.8h, v5.8h, v11.8h\n"
-    "fmla v17.8h, v4.8h, v11.8h\n"
-    "fmla v16.8h, v3.8h, v11.8h\n"
-    "ldr q11, [x15, x28]\n"
-    "add x15, x15, #0x10\n"
-    "fmla v18.8h, v8.8h, v12.8h\n"
-    "fmla v31.8h, v4.8h, v10.8h\n"
-    "fmla v17.8h, v7.8h, v12.8h\n"
-    "fmla v16.8h, v6.8h, v12.8h\n"
-    "ldr q12, [x12, x6]\n"
-    "fmla v30.8h, v3.8h, v10.8h\n"
-    "fmla v27.8h, v1.8h, v10.8h\n"
-    "fmla v26.8h, v0.8h, v10.8h\n"
-    "ldr q10, [x12, x28]\n"
-    "add x12, x12, #0x10\n"
+    "fmla v20.8h, v6.8h, v10.8h\n"
+    "fmla v24.8h, v3.8h, v10.8h\n"
+    "fmla v28.8h, v0.8h, v10.8h\n"
+    "ldr q10, [x25, x16]\n"
+    "fmla v31.8h, v2.8h, v12.8h\n"
+    "fmla v29.8h, v7.8h, v10.8h\n"
+    "fmla v30.8h, v6.8h, v10.8h\n"
+    "fmla v24.8h, v8.8h, v11.8h\n"
+    "fmla v25.8h, v7.8h, v11.8h\n"
+    "fmla v26.8h, v6.8h, v11.8h\n"
+    "fmla v28.8h, v5.8h, v11.8h\n"
+    "ldr q11, [x28, x10]\n"
+    "fmla v27.8h, v5.8h, v12.8h\n"
     "fmla v29.8h, v5.8h, v11.8h\n"
-    "fmla v28.8h, v4.8h, v11.8h\n"
-    "fmla v25.8h, v2.8h, v11.8h\n"
-    "fmla v24.8h, v1.8h, v11.8h\n"
-    "fmla v23.8h, v7.8h, v12.8h\n"
-    "fmla v22.8h, v6.8h, v12.8h\n"
-    "fmla v19.8h, v4.8h, v12.8h\n"
-    "fmla v18.8h, v3.8h, v12.8h\n"
-    "fmla v21.8h, v8.8h, v10.8h\n"
-    "fmla v20.8h, v7.8h, v10.8h\n"
-    "fmla v17.8h, v5.8h, v10.8h\n"
+    "fmla v30.8h, v4.8h, v11.8h\n"
+    "fmla v31.8h, v3.8h, v11.8h\n"
+    "fmla v23.8h, v8.8h, v12.8h\n"
+    "ldr q12, [x25, x10]\n"
+    "fmla v28.8h, v8.8h, v10.8h\n"
+    "ldr q10, [x14, x5]\n"
+    "fmla v25.8h, v8.8h, v11.8h\n"
+    "fmla v26.8h, v7.8h, v11.8h\n"
+    "add x25, x25, #0x10\n"
+    "fmla v27.8h, v6.8h, v11.8h\n"
+    "fmla v29.8h, v8.8h, v12.8h\n"
+    "ldr q11, [x14, x27]\n"
+    "add x14, x14, #0x10\n"
+    "fmla v30.8h, v7.8h, v12.8h\n"
+    "fmla v31.8h, v6.8h, v12.8h\n"
+    "ldr q12, [x28, x5]\n"
     "fmla v16.8h, v4.8h, v10.8h\n"
-    "fmax v31.8h, v31.8h, v15.8h\n"
-    "fmax v30.8h, v30.8h, v15.8h\n"
-    "fmax v29.8h, v29.8h, v15.8h\n"
-    "fmin v31.8h, v31.8h, v14.8h\n"
-    "st1 { v31.8h }, [x16]\n"
-    "fmin v30.8h, v30.8h, v14.8h\n"
-    "fmin v29.8h, v29.8h, v14.8h\n"
-    "str q30, [x16, x17]\n"
-    "fmax v28.8h, v28.8h, v15.8h\n"
-    "fmax v27.8h, v27.8h, v15.8h\n"
-    "str q29, [x16, x23]\n"
-    "fmax v26.8h, v26.8h, v15.8h\n"
-    "fmax v25.8h, v25.8h, v15.8h\n"
-    "fmin v28.8h, v28.8h, v14.8h\n"
-    "str q28, [x16, x22]\n"
-    "fmin v27.8h, v27.8h, v14.8h\n"
-    "add x16, x16, #0x10\n"
-    "fmin v26.8h, v26.8h, v14.8h\n"
-    "st1 { v27.8h }, [x26]\n"
-    "fmin v25.8h, v25.8h, v14.8h\n"
-    "fmax v24.8h, v24.8h, v15.8h\n"
-    "str q26, [x26, x17]\n"
-    "fmax v23.8h, v23.8h, v15.8h\n"
-    "str q25, [x26, x23]\n"
-    "fmin v24.8h, v24.8h, v14.8h\n"
+    "fmla v17.8h, v3.8h, v10.8h\n"
+    "fmax v16.8h, v16.8h, v15.8h\n"
+    "fmla v20.8h, v1.8h, v10.8h\n"
+    "fmla v21.8h, v0.8h, v10.8h\n"
+    "ldr q10, [x28, x27]\n"
+    "fmax v17.8h, v17.8h, v15.8h\n"
+    "fmla v18.8h, v5.8h, v11.8h\n"
+    "fmla v19.8h, v4.8h, v11.8h\n"
+    "fmax v18.8h, v18.8h, v15.8h\n"
+    "add x28, x28, #0x10\n"
+    "fmla v22.8h, v2.8h, v11.8h\n"
+    "fmla v23.8h, v1.8h, v11.8h\n"
+    "fmax v19.8h, v19.8h, v15.8h\n"
+    "fmla v24.8h, v7.8h, v12.8h\n"
+    "fmla v25.8h, v6.8h, v12.8h\n"
+    "fmax v20.8h, v20.8h, v15.8h\n"
+    "fmla v28.8h, v4.8h, v12.8h\n"
+    "fmla v29.8h, v3.8h, v12.8h\n"
+    "fmax v21.8h, v21.8h, v15.8h\n"
+    "fmla v26.8h, v8.8h, v10.8h\n"
+    "fmla v27.8h, v7.8h, v10.8h\n"
     "fmax v22.8h, v22.8h, v15.8h\n"
-    "str q24, [x26, x22]\n"
+    "fmla v30.8h, v5.8h, v10.8h\n"
+    "fmla v31.8h, v4.8h, v10.8h\n"
+    "fmax v23.8h, v23.8h, v15.8h\n"
+    "fmax v24.8h, v24.8h, v15.8h\n"
+    "fmax v25.8h, v25.8h, v15.8h\n"
+    "fmax v26.8h, v26.8h, v15.8h\n"
+    "fmax v27.8h, v27.8h, v15.8h\n"
+    "fmax v28.8h, v28.8h, v15.8h\n"
+    "fmax v29.8h, v29.8h, v15.8h\n"
+    "fmax v30.8h, v30.8h, v15.8h\n"
+    "fmax v31.8h, v31.8h, v15.8h\n"
+    "fmin v16.8h, v16.8h, v14.8h\n"
+    "fmin v17.8h, v17.8h, v14.8h\n"
+    "st1 { v16.8h }, [x17]\n"
+    "fmin v18.8h, v18.8h, v14.8h\n"
+    "fmin v19.8h, v19.8h, v14.8h\n"
+    "str q17, [x17, x6]\n"
+    "fmin v20.8h, v20.8h, v14.8h\n"
+    "fmin v21.8h, v21.8h, v14.8h\n"
+    "str q18, [x17, x22]\n"
+    "fmin v22.8h, v22.8h, v14.8h\n"
     "fmin v23.8h, v23.8h, v14.8h\n"
+    "str q19, [x17, x21]\n"
+    "add x17, x17, #0x10\n"
+    "fmin v24.8h, v24.8h, v14.8h\n"
+    "fmin v25.8h, v25.8h, v14.8h\n"
+    "st1 { v20.8h }, [x9]\n"
+    "fmin v26.8h, v26.8h, v14.8h\n"
+    "fmin v27.8h, v27.8h, v14.8h\n"
+    "str q21, [x9, x6]\n"
+    "fmin v28.8h, v28.8h, v14.8h\n"
+    "fmin v29.8h, v29.8h, v14.8h\n"
+    "str q22, [x9, x22]\n"
+    "fmin v30.8h, v30.8h, v14.8h\n"
+    "fmin v31.8h, v31.8h, v14.8h\n"
+    "str q23, [x9, x21]\n"
+    "add x9, x9, #0x10\n"
+    "st1 { v24.8h }, [x26]\n"
+    "str q25, [x26, x6]\n"
+    "str q26, [x26, x22]\n"
+    "str q27, [x26, x21]\n"
     "add x26, x26, #0x10\n"
-    "fmin v22.8h, v22.8h, v14.8h\n"
-    "st1 { v23.8h }, [x25]\n"
-    "fmax v21.8h, v21.8h, v15.8h\n"
-    "fmax v20.8h, v20.8h, v15.8h\n"
-    "str q22, [x25, x17]\n"
-    "fmax v19.8h, v19.8h, v15.8h\n"
-    "fmax v18.8h, v18.8h, v15.8h\n"
-    "fmin v21.8h, v21.8h, v14.8h\n"
-    "str q21, [x25, x23]\n"
-    "fmin v20.8h, v20.8h, v14.8h\n"
-    "fmin v19.8h, v19.8h, v14.8h\n"
-    "str q20, [x25, x22]\n"
-    "fmin v18.8h, v18.8h, v14.8h\n"
-    "add x25, x25, #0x10\n"
-    "fmax v17.8h, v17.8h, v15.8h\n"
-    "st1 { v19.8h }, [x24]\n"
-    "fmax v16.8h, v16.8h, v15.8h\n"
-    "str q18, [x24, x17]\n"
-    "fmin v17.8h, v17.8h, v14.8h\n"
-    "str q17, [x24, x23]\n"
-    "fmin v16.8h, v16.8h, v14.8h\n"
-    "str q16, [x24, x22]\n"
-    "add x24, x24, #0x10\n"
+    "st1 { v28.8h }, [x23]\n"
+    "str q29, [x23, x6]\n"
+    "str q30, [x23, x22]\n"
+    "str q31, [x23, x21]\n"
+    "add x23, x23, #0x10\n"
     "4:"  // Tile loop: Oddments
-    "tst %x[n_channels], #0x1\n"
-    "beq 73f\n"
-    "ldr q13, [x5, #0x0]\n"
-    "ldr q0, [x5, #0x10]\n"
-    "add x22, x14, x10\n"
-    "ldr q1, [x5, #0x20]\n"
+    "tst %x[n_channels], #0x7\n"
+    "beq 141f\n"
+    "ldr q13, [x15, #0x0]\n"
+    "ldr q0, [x15, #0x10]\n"
+    "ldr q1, [x15, #0x20]\n"
+    "ldr q2, [x15, #0x30]\n"
+    "add x22, x13, x16\n"
     "add x21, x8, XZR\n"
-    "ldr q2, [x5, #0x30]\n"
-    "add x20, x8, x27\n"
-    "ldr q3, [x5, #0x40]\n"
-    "add x19, x14, x9\n"
-    "ldr q4, [x5, #0x50]\n"
-    "ldr q5, [x5, #0x60]\n"
-    "ldr q6, [x5, #0x70]\n"
-    "ldr q7, [x5, #0x80]\n"
-    "ldr q8, [x5, #0x90]\n"
+    "ldr q3, [x15, #0x40]\n"
+    "ldr q4, [x15, #0x50]\n"
+    "add x20, x8, x24\n"
+    "add x19, x13, x10\n"
+    "ldr q5, [x15, #0x60]\n"
+    "ldr q6, [x15, #0x70]\n"
+    "ldr q7, [x15, #0x80]\n"
+    "ldr q8, [x15, #0x90]\n"
+    "tbz %x[n_channels], #2, 6f\n"
+    "ldr d9, [x22], #0x8\n"
+    "ldr d10, [x21], #0x8\n"
+    "ldr d11, [x20], #0x8\n"
+    "ldr d12, [x19], #0x8\n"
     "tbz %x[n_channels], #1, 5f\n"
+    "ld1 { v9.s }[2], [x22], #0x4\n"
+    "ld1 { v10.s }[2], [x21], #0x4\n"
+    "ld1 { v11.s }[2], [x20], #0x4\n"
+    "ld1 { v12.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 8f\n"
+    "ld1 { v9.h }[6], [x22]\n"
+    "ld1 { v10.h }[6], [x21]\n"
+    "ld1 { v11.h }[6], [x20]\n"
+    "ld1 { v12.h }[6], [x19]\n"
+    "b 8f\n"
+    "5:"  // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 5), (2, 3): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 8f\n"
+    "ld1 { v9.h }[4], [x22]\n"
+    "ld1 { v10.h }[4], [x21]\n"
+    "ld1 { v11.h }[4], [x20]\n"
+    "ld1 { v12.h }[4], [x19]\n"
+    "b 8f\n"
+    "6:"  // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 5), (2, 3): Bit 2: Unset
+    "tbz %x[n_channels], #1, 7f\n"
     "ldr s9, [x22], #0x4\n"
     "ldr s10, [x21], #0x4\n"
     "ldr s11, [x20], #0x4\n"
     "ldr s12, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 6f\n"
+    "tbz %x[n_channels], #0, 8f\n"
     "ld1 { v9.h }[2], [x22]\n"
     "ld1 { v10.h }[2], [x21]\n"
     "ld1 { v11.h }[2], [x20]\n"
     "ld1 { v12.h }[2], [x19]\n"
-    "b 6f\n"
-    "5:"  // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 5), (2, 3): Bit 1: Unset
+    "b 8f\n"
+    "7:"  // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 5), (2, 3): Bit 2: Unset: Bit 1: Unset
     "ldr h9, [x22, #0x0]\n"
     "ldr h10, [x21, #0x0]\n"
     "ldr h11, [x20, #0x0]\n"
     "ldr h12, [x19, #0x0]\n"
-    "6:"  // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 5), (2, 3): Bit 1: End
-    "mov v31.16b, v13.16b\n fmla v31.8h, v8.8h, v9.8h\n"
-    "add x19, x11, XZR\n"
-    "mov v30.16b, v13.16b\n fmla v30.8h, v7.8h, v9.8h\n"
-    "mov v29.16b, v13.16b\n fmla v29.8h, v6.8h, v9.8h\n"
-    "mov v27.16b, v13.16b\n fmla v27.8h, v5.8h, v9.8h\n"
-    "mov v26.16b, v13.16b\n fmla v26.8h, v4.8h, v9.8h\n"
-    "mov v25.16b, v13.16b\n fmla v25.8h, v3.8h, v9.8h\n"
-    "mov v23.16b, v13.16b\n fmla v23.8h, v2.8h, v9.8h\n"
-    "mov v22.16b, v13.16b\n fmla v22.8h, v1.8h, v9.8h\n"
-    "mov v21.16b, v13.16b\n fmla v21.8h, v0.8h, v9.8h\n"
-    "fmla v31.8h, v0.8h, v10.8h\n"
-    "mov v28.16b, v13.16b\n fmla v28.8h, v2.8h, v11.8h\n"
-    "fmla v30.8h, v8.8h, v12.8h\n"
-    "fmla v29.8h, v7.8h, v12.8h\n"
-    "fmla v26.8h, v5.8h, v12.8h\n"
-    "fmla v28.8h, v6.8h, v12.8h\n"
-    "fmla v25.8h, v4.8h, v12.8h\n"
-    "mov v24.16b, v13.16b\n fmla v24.8h, v3.8h, v12.8h\n"
-    "fmla v22.8h, v2.8h, v12.8h\n"
-    "fmla v21.8h, v1.8h, v12.8h\n"
-    "mov v20.16b, v13.16b\n fmla v20.8h, v0.8h, v12.8h\n"
-    "tbz %x[n_channels], #1, 7f\n"
-    "ldr s10, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 8f\n"
-    "ld1 { v10.h }[2], [x19]\n"
-    "b 8f\n"
-    "7:"  // Tile loop: Oddments: Load inputs: (5, 0): Bit 1: Unset
-    "ldr h10, [x19, #0x0]\n"
-    "8:"  // Tile loop: Oddments: Load inputs: (5, 0): Bit 1: End
-    "mov v19.16b, v13.16b\n fmla v19.8h, v6.8h, v10.8h\n"
-    "add x19, x11, x27\n"
+    "8:"  // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 5), (2, 3): Bit 2: End
+    "mov v16.16b, v13.16b\n fmla v16.8h, v8.8h, v9.8h\n"
+    "mov v17.16b, v13.16b\n fmla v17.8h, v7.8h, v9.8h\n"
+    "add x19, x25, XZR\n"
+    "mov v18.16b, v13.16b\n fmla v18.8h, v6.8h, v9.8h\n"
+    "mov v21.16b, v13.16b\n fmla v21.8h, v4.8h, v9.8h\n"
+    "mov v22.16b, v13.16b\n fmla v22.8h, v3.8h, v9.8h\n"
+    "mov v25.16b, v13.16b\n fmla v25.8h, v1.8h, v9.8h\n"
+    "mov v26.16b, v13.16b\n fmla v26.8h, v0.8h, v9.8h\n"
+    "mov v19.16b, v13.16b\n fmla v19.8h, v2.8h, v11.8h\n"
+    "mov v20.16b, v13.16b\n fmla v20.8h, v5.8h, v9.8h\n"
+    "mov v24.16b, v13.16b\n fmla v24.8h, v2.8h, v9.8h\n"
+    "fmla v16.8h, v0.8h, v10.8h\n"
+    "fmla v17.8h, v8.8h, v12.8h\n"
+    "fmla v18.8h, v7.8h, v12.8h\n"
+    "fmla v19.8h, v6.8h, v12.8h\n"
+    "fmla v21.8h, v5.8h, v12.8h\n"
+    "fmla v22.8h, v4.8h, v12.8h\n"
+    "mov v23.16b, v13.16b\n fmla v23.8h, v3.8h, v12.8h\n"
+    "fmla v25.8h, v2.8h, v12.8h\n"
+    "fmla v26.8h, v1.8h, v12.8h\n"
+    "mov v27.16b, v13.16b\n fmla v27.8h, v0.8h, v12.8h\n"
+    "tbz %x[n_channels], #2, 10f\n"
+    "ldr d10, [x19], #0x8\n"
     "tbz %x[n_channels], #1, 9f\n"
-    "ldr s11, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 10f\n"
-    "ld1 { v11.h }[2], [x19]\n"
-    "b 10f\n"
-    "9:"  // Tile loop: Oddments: Load inputs: (5, 5): Bit 1: Unset
-    "ldr h11, [x19, #0x0]\n"
-    "10:"  // Tile loop: Oddments: Load inputs: (5, 5): Bit 1: End
-    "mov v16.16b, v13.16b\n fmla v16.8h, v8.8h, v11.8h\n"
-    "add x19, x13, x10\n"
+    "ld1 { v10.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 12f\n"
+    "ld1 { v10.h }[6], [x19]\n"
+    "b 12f\n"
+    "9:"  // Tile loop: Oddments: Load inputs: (5, 0): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 12f\n"
+    "ld1 { v10.h }[4], [x19]\n"
+    "b 12f\n"
+    "10:"  // Tile loop: Oddments: Load inputs: (5, 0): Bit 2: Unset
     "tbz %x[n_channels], #1, 11f\n"
-    "ldr s9, [x19], #0x4\n"
+    "ldr s10, [x19], #0x4\n"
     "tbz %x[n_channels], #0, 12f\n"
-    "ld1 { v9.h }[2], [x19]\n"
+    "ld1 { v10.h }[2], [x19]\n"
     "b 12f\n"
-    "11:"  // Tile loop: Oddments: Load inputs: (3, 2): Bit 1: Unset
-    "ldr h9, [x19, #0x0]\n"
-    "12:"  // Tile loop: Oddments: Load inputs: (3, 2): Bit 1: End
-    "fmla v27.8h, v8.8h, v9.8h\n"
-    "add x19, x8, x6\n"
-    "fmla v26.8h, v7.8h, v9.8h\n"
-    "fmla v25.8h, v6.8h, v9.8h\n"
-    "fmla v23.8h, v5.8h, v9.8h\n"
-    "fmla v22.8h, v4.8h, v9.8h\n"
-    "fmla v21.8h, v3.8h, v9.8h\n"
-    "fmla v19.8h, v2.8h, v9.8h\n"
-    "mov v18.16b, v13.16b\n fmla v18.8h, v1.8h, v9.8h\n"
-    "mov v17.16b, v13.16b\n fmla v17.8h, v0.8h, v9.8h\n"
+    "11:"  // Tile loop: Oddments: Load inputs: (5, 0): Bit 2: Unset: Bit 1: Unset
+    "ldr h10, [x19, #0x0]\n"
+    "12:"  // Tile loop: Oddments: Load inputs: (5, 0): Bit 2: End
+    "mov v28.16b, v13.16b\n fmla v28.8h, v6.8h, v10.8h\n"
+    "add x19, x25, x24\n"
+    "tbz %x[n_channels], #2, 14f\n"
+    "ldr d11, [x19], #0x8\n"
     "tbz %x[n_channels], #1, 13f\n"
-    "ldr s12, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 14f\n"
-    "ld1 { v12.h }[2], [x19]\n"
-    "b 14f\n"
-    "13:"  // Tile loop: Oddments: Load inputs: (0, 1): Bit 1: Unset
-    "ldr h12, [x19, #0x0]\n"
-    "14:"  // Tile loop: Oddments: Load inputs: (0, 1): Bit 1: End
-    "fmla v31.8h, v1.8h, v12.8h\n"
-    "add x19, x8, x28\n"
-    "fmla v30.8h, v0.8h, v12.8h\n"
+    "ld1 { v11.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 16f\n"
+    "ld1 { v11.h }[6], [x19]\n"
+    "b 16f\n"
+    "13:"  // Tile loop: Oddments: Load inputs: (5, 5): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 16f\n"
+    "ld1 { v11.h }[4], [x19]\n"
+    "b 16f\n"
+    "14:"  // Tile loop: Oddments: Load inputs: (5, 5): Bit 2: Unset
     "tbz %x[n_channels], #1, 15f\n"
     "ldr s11, [x19], #0x4\n"
     "tbz %x[n_channels], #0, 16f\n"
     "ld1 { v11.h }[2], [x19]\n"
     "b 16f\n"
-    "15:"  // Tile loop: Oddments: Load inputs: (0, 4): Bit 1: Unset
+    "15:"  // Tile loop: Oddments: Load inputs: (5, 5): Bit 2: Unset: Bit 1: Unset
     "ldr h11, [x19, #0x0]\n"
-    "16:"  // Tile loop: Oddments: Load inputs: (0, 4): Bit 1: End
-    "fmla v29.8h, v2.8h, v11.8h\n"
-    "add x19, x13, x9\n"
-    "fmla v28.8h, v1.8h, v11.8h\n"
+    "16:"  // Tile loop: Oddments: Load inputs: (5, 5): Bit 2: End
+    "mov v31.16b, v13.16b\n fmla v31.8h, v8.8h, v11.8h\n"
+    "add x19, x11, x16\n"
+    "tbz %x[n_channels], #2, 18f\n"
+    "ldr d9, [x19], #0x8\n"
     "tbz %x[n_channels], #1, 17f\n"
-    "ldr s10, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 18f\n"
-    "ld1 { v10.h }[2], [x19]\n"
-    "b 18f\n"
-    "17:"  // Tile loop: Oddments: Load inputs: (3, 3): Bit 1: Unset
-    "ldr h10, [x19, #0x0]\n"
-    "18:"  // Tile loop: Oddments: Load inputs: (3, 3): Bit 1: End
-    "fmla v26.8h, v8.8h, v10.8h\n"
-    "add x19, x15, XZR\n"
-    "fmla v25.8h, v7.8h, v10.8h\n"
-    "fmla v24.8h, v6.8h, v10.8h\n"
-    "fmla v22.8h, v5.8h, v10.8h\n"
-    "fmla v21.8h, v4.8h, v10.8h\n"
-    "fmla v20.8h, v3.8h, v10.8h\n"
-    "fmla v18.8h, v2.8h, v10.8h\n"
-    "fmla v17.8h, v1.8h, v10.8h\n"
-    "fmla v16.8h, v0.8h, v10.8h\n"
+    "ld1 { v9.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 20f\n"
+    "ld1 { v9.h }[6], [x19]\n"
+    "b 20f\n"
+    "17:"  // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 20f\n"
+    "ld1 { v9.h }[4], [x19]\n"
+    "b 20f\n"
+    "18:"  // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: Unset
     "tbz %x[n_channels], #1, 19f\n"
     "ldr s9, [x19], #0x4\n"
     "tbz %x[n_channels], #0, 20f\n"
     "ld1 { v9.h }[2], [x19]\n"
     "b 20f\n"
-    "19:"  // Tile loop: Oddments: Load inputs: (1, 0): Bit 1: Unset
+    "19:"  // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: Unset: Bit 1: Unset
     "ldr h9, [x19, #0x0]\n"
-    "20:"  // Tile loop: Oddments: Load inputs: (1, 0): Bit 1: End
-    "fmla v31.8h, v3.8h, v9.8h\n"
-    "add x19, x15, x27\n"
-    "fmla v27.8h, v0.8h, v9.8h\n"
+    "20:"  // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: End
+    "fmla v20.8h, v8.8h, v9.8h\n"
+    "fmla v21.8h, v7.8h, v9.8h\n"
+    "add x19, x8, x5\n"
+    "fmla v22.8h, v6.8h, v9.8h\n"
+    "fmla v24.8h, v5.8h, v9.8h\n"
+    "fmla v25.8h, v4.8h, v9.8h\n"
+    "fmla v26.8h, v3.8h, v9.8h\n"
+    "fmla v28.8h, v2.8h, v9.8h\n"
+    "mov v29.16b, v13.16b\n fmla v29.8h, v1.8h, v9.8h\n"
+    "mov v30.16b, v13.16b\n fmla v30.8h, v0.8h, v9.8h\n"
+    "tbz %x[n_channels], #2, 22f\n"
+    "ldr d12, [x19], #0x8\n"
     "tbz %x[n_channels], #1, 21f\n"
-    "ldr s12, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 22f\n"
-    "ld1 { v12.h }[2], [x19]\n"
-    "b 22f\n"
-    "21:"  // Tile loop: Oddments: Load inputs: (1, 5): Bit 1: Unset
-    "ldr h12, [x19, #0x0]\n"
-    "22:"  // Tile loop: Oddments: Load inputs: (1, 5): Bit 1: End
-    "fmla v28.8h, v5.8h, v12.8h\n"
-    "add x19, x12, XZR\n"
-    "fmla v24.8h, v2.8h, v12.8h\n"
+    "ld1 { v12.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 24f\n"
+    "ld1 { v12.h }[6], [x19]\n"
+    "b 24f\n"
+    "21:"  // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 24f\n"
+    "ld1 { v12.h }[4], [x19]\n"
+    "b 24f\n"
+    "22:"  // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: Unset
     "tbz %x[n_channels], #1, 23f\n"
-    "ldr s11, [x19], #0x4\n"
+    "ldr s12, [x19], #0x4\n"
     "tbz %x[n_channels], #0, 24f\n"
-    "ld1 { v11.h }[2], [x19]\n"
+    "ld1 { v12.h }[2], [x19]\n"
     "b 24f\n"
-    "23:"  // Tile loop: Oddments: Load inputs: (4, 0): Bit 1: Unset
-    "ldr h11, [x19, #0x0]\n"
-    "24:"  // Tile loop: Oddments: Load inputs: (4, 0): Bit 1: End
-    "fmla v23.8h, v6.8h, v11.8h\n"
-    "add x19, x15, x10\n"
-    "fmla v19.8h, v3.8h, v11.8h\n"
+    "23:"  // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: Unset: Bit 1: Unset
+    "ldr h12, [x19, #0x0]\n"
+    "24:"  // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: End
+    "fmla v16.8h, v1.8h, v12.8h\n"
+    "fmla v17.8h, v0.8h, v12.8h\n"
+    "add x19, x8, x27\n"
+    "tbz %x[n_channels], #2, 26f\n"
+    "ldr d11, [x19], #0x8\n"
     "tbz %x[n_channels], #1, 25f\n"
-    "ldr s10, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 26f\n"
-    "ld1 { v10.h }[2], [x19]\n"
-    "b 26f\n"
-    "25:"  // Tile loop: Oddments: Load inputs: (1, 2): Bit 1: Unset
-    "ldr h10, [x19, #0x0]\n"
-    "26:"  // Tile loop: Oddments: Load inputs: (1, 2): Bit 1: End
-    "fmla v31.8h, v5.8h, v10.8h\n"
-    "add x19, x12, x27\n"
-    "fmla v30.8h, v4.8h, v10.8h\n"
-    "fmla v29.8h, v3.8h, v10.8h\n"
-    "fmla v27.8h, v2.8h, v10.8h\n"
-    "fmla v26.8h, v1.8h, v10.8h\n"
-    "fmla v25.8h, v0.8h, v10.8h\n"
+    "ld1 { v11.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 28f\n"
+    "ld1 { v11.h }[6], [x19]\n"
+    "b 28f\n"
+    "25:"  // Tile loop: Oddments: Load inputs: (0, 4): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 28f\n"
+    "ld1 { v11.h }[4], [x19]\n"
+    "b 28f\n"
+    "26:"  // Tile loop: Oddments: Load inputs: (0, 4): Bit 2: Unset
     "tbz %x[n_channels], #1, 27f\n"
     "ldr s11, [x19], #0x4\n"
     "tbz %x[n_channels], #0, 28f\n"
     "ld1 { v11.h }[2], [x19]\n"
     "b 28f\n"
-    "27:"  // Tile loop: Oddments: Load inputs: (4, 5): Bit 1: Unset
+    "27:"  // Tile loop: Oddments: Load inputs: (0, 4): Bit 2: Unset: Bit 1: Unset
     "ldr h11, [x19, #0x0]\n"
-    "28:"  // Tile loop: Oddments: Load inputs: (4, 5): Bit 1: End
-    "fmla v20.8h, v8.8h, v11.8h\n"
-    "add x19, x15, x9\n"
-    "fmla v16.8h, v5.8h, v11.8h\n"
+    "28:"  // Tile loop: Oddments: Load inputs: (0, 4): Bit 2: End
+    "fmla v18.8h, v2.8h, v11.8h\n"
+    "fmla v19.8h, v1.8h, v11.8h\n"
+    "add x19, x11, x10\n"
+    "tbz %x[n_channels], #2, 30f\n"
+    "ldr d10, [x19], #0x8\n"
     "tbz %x[n_channels], #1, 29f\n"
+    "ld1 { v10.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 32f\n"
+    "ld1 { v10.h }[6], [x19]\n"
+    "b 32f\n"
+    "29:"  // Tile loop: Oddments: Load inputs: (3, 3): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 32f\n"
+    "ld1 { v10.h }[4], [x19]\n"
+    "b 32f\n"
+    "30:"  // Tile loop: Oddments: Load inputs: (3, 3): Bit 2: Unset
+    "tbz %x[n_channels], #1, 31f\n"
+    "ldr s10, [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 32f\n"
+    "ld1 { v10.h }[2], [x19]\n"
+    "b 32f\n"
+    "31:"  // Tile loop: Oddments: Load inputs: (3, 3): Bit 2: Unset: Bit 1: Unset
+    "ldr h10, [x19, #0x0]\n"
+    "32:"  // Tile loop: Oddments: Load inputs: (3, 3): Bit 2: End
+    "fmla v21.8h, v8.8h, v10.8h\n"
+    "fmla v22.8h, v7.8h, v10.8h\n"
+    "add x19, x14, XZR\n"
+    "fmla v23.8h, v6.8h, v10.8h\n"
+    "fmla v25.8h, v5.8h, v10.8h\n"
+    "fmla v26.8h, v4.8h, v10.8h\n"
+    "fmla v27.8h, v3.8h, v10.8h\n"
+    "fmla v29.8h, v2.8h, v10.8h\n"
+    "fmla v30.8h, v1.8h, v10.8h\n"
+    "fmla v31.8h, v0.8h, v10.8h\n"
+    "tbz %x[n_channels], #2, 34f\n"
+    "ldr d9, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 33f\n"
+    "ld1 { v9.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 36f\n"
+    "ld1 { v9.h }[6], [x19]\n"
+    "b 36f\n"
+    "33:"  // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 36f\n"
+    "ld1 { v9.h }[4], [x19]\n"
+    "b 36f\n"
+    "34:"  // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: Unset
+    "tbz %x[n_channels], #1, 35f\n"
+    "ldr s9, [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 36f\n"
+    "ld1 { v9.h }[2], [x19]\n"
+    "b 36f\n"
+    "35:"  // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: Unset: Bit 1: Unset
+    "ldr h9, [x19, #0x0]\n"
+    "36:"  // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: End
+    "fmla v16.8h, v3.8h, v9.8h\n"
+    "fmla v20.8h, v0.8h, v9.8h\n"
+    "add x19, x14, x24\n"
+    "tbz %x[n_channels], #2, 38f\n"
+    "ldr d12, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 37f\n"
+    "ld1 { v12.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 40f\n"
+    "ld1 { v12.h }[6], [x19]\n"
+    "b 40f\n"
+    "37:"  // Tile loop: Oddments: Load inputs: (1, 5): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 40f\n"
+    "ld1 { v12.h }[4], [x19]\n"
+    "b 40f\n"
+    "38:"  // Tile loop: Oddments: Load inputs: (1, 5): Bit 2: Unset
+    "tbz %x[n_channels], #1, 39f\n"
     "ldr s12, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 30f\n"
+    "tbz %x[n_channels], #0, 40f\n"
     "ld1 { v12.h }[2], [x19]\n"
-    "b 30f\n"
-    "29:"  // Tile loop: Oddments: Load inputs: (1, 3): Bit 1: Unset
+    "b 40f\n"
+    "39:"  // Tile loop: Oddments: Load inputs: (1, 5): Bit 2: Unset: Bit 1: Unset
     "ldr h12, [x19, #0x0]\n"
-    "30:"  // Tile loop: Oddments: Load inputs: (1, 3): Bit 1: End
-    "fmla v30.8h, v5.8h, v12.8h\n"
-    "add x19, x11, x6\n"
-    "fmla v29.8h, v4.8h, v12.8h\n"
-    "fmla v28.8h, v3.8h, v12.8h\n"
-    "fmla v26.8h, v2.8h, v12.8h\n"
-    "fmla v25.8h, v1.8h, v12.8h\n"
-    "fmla v24.8h, v0.8h, v12.8h\n"
-    "tbz %x[n_channels], #1, 31f\n"
+    "40:"  // Tile loop: Oddments: Load inputs: (1, 5): Bit 2: End
+    "fmla v19.8h, v5.8h, v12.8h\n"
+    "fmla v23.8h, v2.8h, v12.8h\n"
+    "add x19, x28, XZR\n"
+    "tbz %x[n_channels], #2, 42f\n"
+    "ldr d11, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 41f\n"
+    "ld1 { v11.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 44f\n"
+    "ld1 { v11.h }[6], [x19]\n"
+    "b 44f\n"
+    "41:"  // Tile loop: Oddments: Load inputs: (4, 0): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 44f\n"
+    "ld1 { v11.h }[4], [x19]\n"
+    "b 44f\n"
+    "42:"  // Tile loop: Oddments: Load inputs: (4, 0): Bit 2: Unset
+    "tbz %x[n_channels], #1, 43f\n"
     "ldr s11, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 32f\n"
+    "tbz %x[n_channels], #0, 44f\n"
     "ld1 { v11.h }[2], [x19]\n"
-    "b 32f\n"
-    "31:"  // Tile loop: Oddments: Load inputs: (5, 1): Bit 1: Unset
+    "b 44f\n"
+    "43:"  // Tile loop: Oddments: Load inputs: (4, 0): Bit 2: Unset: Bit 1: Unset
     "ldr h11, [x19, #0x0]\n"
-    "32:"  // Tile loop: Oddments: Load inputs: (5, 1): Bit 1: End
-    "fmla v19.8h, v7.8h, v11.8h\n"
-    "add x19, x14, x6\n"
-    "fmla v18.8h, v6.8h, v11.8h\n"
-    "tbz %x[n_channels], #1, 33f\n"
+    "44:"  // Tile loop: Oddments: Load inputs: (4, 0): Bit 2: End
+    "fmla v24.8h, v6.8h, v11.8h\n"
+    "fmla v28.8h, v3.8h, v11.8h\n"
+    "add x19, x14, x16\n"
+    "tbz %x[n_channels], #2, 46f\n"
+    "ldr d10, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 45f\n"
+    "ld1 { v10.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 48f\n"
+    "ld1 { v10.h }[6], [x19]\n"
+    "b 48f\n"
+    "45:"  // Tile loop: Oddments: Load inputs: (1, 2): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 48f\n"
+    "ld1 { v10.h }[4], [x19]\n"
+    "b 48f\n"
+    "46:"  // Tile loop: Oddments: Load inputs: (1, 2): Bit 2: Unset
+    "tbz %x[n_channels], #1, 47f\n"
     "ldr s10, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 34f\n"
+    "tbz %x[n_channels], #0, 48f\n"
     "ld1 { v10.h }[2], [x19]\n"
-    "b 34f\n"
-    "33:"  // Tile loop: Oddments: Load inputs: (2, 1): Bit 1: Unset
+    "b 48f\n"
+    "47:"  // Tile loop: Oddments: Load inputs: (1, 2): Bit 2: Unset: Bit 1: Unset
     "ldr h10, [x19, #0x0]\n"
-    "34:"  // Tile loop: Oddments: Load inputs: (2, 1): Bit 1: End
-    "fmla v31.8h, v7.8h, v10.8h\n"
-    "add x19, x11, x28\n"
-    "fmla v30.8h, v6.8h, v10.8h\n"
-    "fmla v27.8h, v4.8h, v10.8h\n"
-    "fmla v26.8h, v3.8h, v10.8h\n"
-    "fmla v23.8h, v1.8h, v10.8h\n"
+    "48:"  // Tile loop: Oddments: Load inputs: (1, 2): Bit 2: End
+    "fmla v16.8h, v5.8h, v10.8h\n"
+    "fmla v17.8h, v4.8h, v10.8h\n"
+    "add x19, x28, x24\n"
+    "fmla v18.8h, v3.8h, v10.8h\n"
+    "fmla v20.8h, v2.8h, v10.8h\n"
+    "fmla v21.8h, v1.8h, v10.8h\n"
     "fmla v22.8h, v0.8h, v10.8h\n"
-    "tbz %x[n_channels], #1, 35f\n"
+    "tbz %x[n_channels], #2, 50f\n"
+    "ldr d11, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 49f\n"
+    "ld1 { v11.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 52f\n"
+    "ld1 { v11.h }[6], [x19]\n"
+    "b 52f\n"
+    "49:"  // Tile loop: Oddments: Load inputs: (4, 5): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 52f\n"
+    "ld1 { v11.h }[4], [x19]\n"
+    "b 52f\n"
+    "50:"  // Tile loop: Oddments: Load inputs: (4, 5): Bit 2: Unset
+    "tbz %x[n_channels], #1, 51f\n"
     "ldr s11, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 36f\n"
+    "tbz %x[n_channels], #0, 52f\n"
     "ld1 { v11.h }[2], [x19]\n"
-    "b 36f\n"
-    "35:"  // Tile loop: Oddments: Load inputs: (5, 4): Bit 1: Unset
+    "b 52f\n"
+    "51:"  // Tile loop: Oddments: Load inputs: (4, 5): Bit 2: Unset: Bit 1: Unset
     "ldr h11, [x19, #0x0]\n"
-    "36:"  // Tile loop: Oddments: Load inputs: (5, 4): Bit 1: End
-    "fmla v17.8h, v8.8h, v11.8h\n"
-    "add x19, x14, x28\n"
-    "fmla v16.8h, v7.8h, v11.8h\n"
-    "tbz %x[n_channels], #1, 37f\n"
+    "52:"  // Tile loop: Oddments: Load inputs: (4, 5): Bit 2: End
+    "fmla v27.8h, v8.8h, v11.8h\n"
+    "fmla v31.8h, v5.8h, v11.8h\n"
+    "add x19, x14, x10\n"
+    "tbz %x[n_channels], #2, 54f\n"
+    "ldr d12, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 53f\n"
+    "ld1 { v12.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 56f\n"
+    "ld1 { v12.h }[6], [x19]\n"
+    "b 56f\n"
+    "53:"  // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 56f\n"
+    "ld1 { v12.h }[4], [x19]\n"
+    "b 56f\n"
+    "54:"  // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: Unset
+    "tbz %x[n_channels], #1, 55f\n"
     "ldr s12, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 38f\n"
+    "tbz %x[n_channels], #0, 56f\n"
     "ld1 { v12.h }[2], [x19]\n"
-    "b 38f\n"
-    "37:"  // Tile loop: Oddments: Load inputs: (2, 4): Bit 1: Unset
+    "b 56f\n"
+    "55:"  // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: Unset: Bit 1: Unset
     "ldr h12, [x19, #0x0]\n"
-    "38:"  // Tile loop: Oddments: Load inputs: (2, 4): Bit 1: End
-    "fmla v29.8h, v8.8h, v12.8h\n"
-    "add x19, x8, x10\n"
-    "fmla v28.8h, v7.8h, v12.8h\n"
-    "fmla v25.8h, v5.8h, v12.8h\n"
-    "fmla v24.8h, v4.8h, v12.8h\n"
+    "56:"  // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: End
+    "fmla v17.8h, v5.8h, v12.8h\n"
+    "fmla v18.8h, v4.8h, v12.8h\n"
+    "add x19, x25, x5\n"
+    "fmla v19.8h, v3.8h, v12.8h\n"
     "fmla v21.8h, v2.8h, v12.8h\n"
-    "fmla v20.8h, v1.8h, v12.8h\n"
-    "tbz %x[n_channels], #1, 39f\n"
+    "fmla v22.8h, v1.8h, v12.8h\n"
+    "fmla v23.8h, v0.8h, v12.8h\n"
+    "tbz %x[n_channels], #2, 58f\n"
+    "ldr d11, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 57f\n"
+    "ld1 { v11.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 60f\n"
+    "ld1 { v11.h }[6], [x19]\n"
+    "b 60f\n"
+    "57:"  // Tile loop: Oddments: Load inputs: (5, 1): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 60f\n"
+    "ld1 { v11.h }[4], [x19]\n"
+    "b 60f\n"
+    "58:"  // Tile loop: Oddments: Load inputs: (5, 1): Bit 2: Unset
+    "tbz %x[n_channels], #1, 59f\n"
+    "ldr s11, [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 60f\n"
+    "ld1 { v11.h }[2], [x19]\n"
+    "b 60f\n"
+    "59:"  // Tile loop: Oddments: Load inputs: (5, 1): Bit 2: Unset: Bit 1: Unset
+    "ldr h11, [x19, #0x0]\n"
+    "60:"  // Tile loop: Oddments: Load inputs: (5, 1): Bit 2: End
+    "fmla v28.8h, v7.8h, v11.8h\n"
+    "fmla v29.8h, v6.8h, v11.8h\n"
+    "add x19, x13, x5\n"
+    "tbz %x[n_channels], #2, 62f\n"
+    "ldr d10, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 61f\n"
+    "ld1 { v10.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 64f\n"
+    "ld1 { v10.h }[6], [x19]\n"
+    "b 64f\n"
+    "61:"  // Tile loop: Oddments: Load inputs: (2, 1): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 64f\n"
+    "ld1 { v10.h }[4], [x19]\n"
+    "b 64f\n"
+    "62:"  // Tile loop: Oddments: Load inputs: (2, 1): Bit 2: Unset
+    "tbz %x[n_channels], #1, 63f\n"
     "ldr s10, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 40f\n"
+    "tbz %x[n_channels], #0, 64f\n"
     "ld1 { v10.h }[2], [x19]\n"
-    "b 40f\n"
-    "39:"  // Tile loop: Oddments: Load inputs: (0, 2): Bit 1: Unset
+    "b 64f\n"
+    "63:"  // Tile loop: Oddments: Load inputs: (2, 1): Bit 2: Unset: Bit 1: Unset
     "ldr h10, [x19, #0x0]\n"
-    "40:"  // Tile loop: Oddments: Load inputs: (0, 2): Bit 1: End
-    "fmla v31.8h, v2.8h, v10.8h\n"
-    "add x19, x13, x6\n"
-    "fmla v30.8h, v1.8h, v10.8h\n"
-    "fmla v29.8h, v0.8h, v10.8h\n"
-    "tbz %x[n_channels], #1, 41f\n"
+    "64:"  // Tile loop: Oddments: Load inputs: (2, 1): Bit 2: End
+    "fmla v16.8h, v7.8h, v10.8h\n"
+    "fmla v17.8h, v6.8h, v10.8h\n"
+    "add x19, x25, x27\n"
+    "fmla v20.8h, v4.8h, v10.8h\n"
+    "fmla v21.8h, v3.8h, v10.8h\n"
+    "fmla v24.8h, v1.8h, v10.8h\n"
+    "fmla v25.8h, v0.8h, v10.8h\n"
+    "tbz %x[n_channels], #2, 66f\n"
+    "ldr d11, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 65f\n"
+    "ld1 { v11.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 68f\n"
+    "ld1 { v11.h }[6], [x19]\n"
+    "b 68f\n"
+    "65:"  // Tile loop: Oddments: Load inputs: (5, 4): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 68f\n"
+    "ld1 { v11.h }[4], [x19]\n"
+    "b 68f\n"
+    "66:"  // Tile loop: Oddments: Load inputs: (5, 4): Bit 2: Unset
+    "tbz %x[n_channels], #1, 67f\n"
     "ldr s11, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 42f\n"
+    "tbz %x[n_channels], #0, 68f\n"
     "ld1 { v11.h }[2], [x19]\n"
-    "b 42f\n"
-    "41:"  // Tile loop: Oddments: Load inputs: (3, 1): Bit 1: Unset
+    "b 68f\n"
+    "67:"  // Tile loop: Oddments: Load inputs: (5, 4): Bit 2: Unset: Bit 1: Unset
     "ldr h11, [x19, #0x0]\n"
-    "42:"  // Tile loop: Oddments: Load inputs: (3, 1): Bit 1: End
-    "fmla v27.8h, v7.8h, v11.8h\n"
-    "add x19, x8, x9\n"
-    "fmla v26.8h, v6.8h, v11.8h\n"
-    "fmla v23.8h, v4.8h, v11.8h\n"
-    "fmla v22.8h, v3.8h, v11.8h\n"
-    "fmla v19.8h, v1.8h, v11.8h\n"
-    "fmla v18.8h, v0.8h, v11.8h\n"
-    "tbz %x[n_channels], #1, 43f\n"
+    "68:"  // Tile loop: Oddments: Load inputs: (5, 4): Bit 2: End
+    "fmla v30.8h, v8.8h, v11.8h\n"
+    "fmla v31.8h, v7.8h, v11.8h\n"
+    "add x19, x13, x27\n"
+    "tbz %x[n_channels], #2, 70f\n"
+    "ldr d12, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 69f\n"
+    "ld1 { v12.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 72f\n"
+    "ld1 { v12.h }[6], [x19]\n"
+    "b 72f\n"
+    "69:"  // Tile loop: Oddments: Load inputs: (2, 4): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 72f\n"
+    "ld1 { v12.h }[4], [x19]\n"
+    "b 72f\n"
+    "70:"  // Tile loop: Oddments: Load inputs: (2, 4): Bit 2: Unset
+    "tbz %x[n_channels], #1, 71f\n"
     "ldr s12, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 44f\n"
+    "tbz %x[n_channels], #0, 72f\n"
     "ld1 { v12.h }[2], [x19]\n"
-    "b 44f\n"
-    "43:"  // Tile loop: Oddments: Load inputs: (0, 3): Bit 1: Unset
+    "b 72f\n"
+    "71:"  // Tile loop: Oddments: Load inputs: (2, 4): Bit 2: Unset: Bit 1: Unset
     "ldr h12, [x19, #0x0]\n"
-    "44:"  // Tile loop: Oddments: Load inputs: (0, 3): Bit 1: End
-    "fmla v30.8h, v2.8h, v12.8h\n"
-    "add x19, x14, XZR\n"
-    "fmla v29.8h, v1.8h, v12.8h\n"
-    "fmla v28.8h, v0.8h, v12.8h\n"
-    "tbz %x[n_channels], #1, 45f\n"
+    "72:"  // Tile loop: Oddments: Load inputs: (2, 4): Bit 2: End
+    "fmla v18.8h, v8.8h, v12.8h\n"
+    "fmla v19.8h, v7.8h, v12.8h\n"
+    "add x19, x8, x16\n"
+    "fmla v22.8h, v5.8h, v12.8h\n"
+    "fmla v23.8h, v4.8h, v12.8h\n"
+    "fmla v26.8h, v2.8h, v12.8h\n"
+    "fmla v27.8h, v1.8h, v12.8h\n"
+    "tbz %x[n_channels], #2, 74f\n"
+    "ldr d10, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 73f\n"
+    "ld1 { v10.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 76f\n"
+    "ld1 { v10.h }[6], [x19]\n"
+    "b 76f\n"
+    "73:"  // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 76f\n"
+    "ld1 { v10.h }[4], [x19]\n"
+    "b 76f\n"
+    "74:"  // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: Unset
+    "tbz %x[n_channels], #1, 75f\n"
     "ldr s10, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 46f\n"
+    "tbz %x[n_channels], #0, 76f\n"
     "ld1 { v10.h }[2], [x19]\n"
-    "b 46f\n"
-    "45:"  // Tile loop: Oddments: Load inputs: (2, 0): Bit 1: Unset
+    "b 76f\n"
+    "75:"  // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: Unset: Bit 1: Unset
     "ldr h10, [x19, #0x0]\n"
-    "46:"  // Tile loop: Oddments: Load inputs: (2, 0): Bit 1: End
-    "fmla v31.8h, v6.8h, v10.8h\n"
-    "add x19, x13, x28\n"
-    "fmla v27.8h, v3.8h, v10.8h\n"
-    "fmla v23.8h, v0.8h, v10.8h\n"
-    "tbz %x[n_channels], #1, 47f\n"
+    "76:"  // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: End
+    "fmla v16.8h, v2.8h, v10.8h\n"
+    "fmla v17.8h, v1.8h, v10.8h\n"
+    "add x19, x11, x5\n"
+    "fmla v18.8h, v0.8h, v10.8h\n"
+    "tbz %x[n_channels], #2, 78f\n"
+    "ldr d11, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 77f\n"
+    "ld1 { v11.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 80f\n"
+    "ld1 { v11.h }[6], [x19]\n"
+    "b 80f\n"
+    "77:"  // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 80f\n"
+    "ld1 { v11.h }[4], [x19]\n"
+    "b 80f\n"
+    "78:"  // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: Unset
+    "tbz %x[n_channels], #1, 79f\n"
     "ldr s11, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 48f\n"
+    "tbz %x[n_channels], #0, 80f\n"
     "ld1 { v11.h }[2], [x19]\n"
-    "b 48f\n"
-    "47:"  // Tile loop: Oddments: Load inputs: (3, 4): Bit 1: Unset
+    "b 80f\n"
+    "79:"  // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: Unset: Bit 1: Unset
     "ldr h11, [x19, #0x0]\n"
-    "48:"  // Tile loop: Oddments: Load inputs: (3, 4): Bit 1: End
-    "fmla v25.8h, v8.8h, v11.8h\n"
-    "add x19, x14, x27\n"
-    "fmla v24.8h, v7.8h, v11.8h\n"
-    "fmla v21.8h, v5.8h, v11.8h\n"
-    "fmla v20.8h, v4.8h, v11.8h\n"
-    "fmla v17.8h, v2.8h, v11.8h\n"
-    "fmla v16.8h, v1.8h, v11.8h\n"
-    "tbz %x[n_channels], #1, 49f\n"
+    "80:"  // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: End
+    "fmla v20.8h, v7.8h, v11.8h\n"
+    "fmla v21.8h, v6.8h, v11.8h\n"
+    "add x19, x8, x10\n"
+    "fmla v24.8h, v4.8h, v11.8h\n"
+    "fmla v25.8h, v3.8h, v11.8h\n"
+    "fmla v28.8h, v1.8h, v11.8h\n"
+    "fmla v29.8h, v0.8h, v11.8h\n"
+    "tbz %x[n_channels], #2, 82f\n"
+    "ldr d12, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 81f\n"
+    "ld1 { v12.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 84f\n"
+    "ld1 { v12.h }[6], [x19]\n"
+    "b 84f\n"
+    "81:"  // Tile loop: Oddments: Load inputs: (0, 3): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 84f\n"
+    "ld1 { v12.h }[4], [x19]\n"
+    "b 84f\n"
+    "82:"  // Tile loop: Oddments: Load inputs: (0, 3): Bit 2: Unset
+    "tbz %x[n_channels], #1, 83f\n"
     "ldr s12, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 50f\n"
+    "tbz %x[n_channels], #0, 84f\n"
     "ld1 { v12.h }[2], [x19]\n"
-    "b 50f\n"
-    "49:"  // Tile loop: Oddments: Load inputs: (2, 5): Bit 1: Unset
+    "b 84f\n"
+    "83:"  // Tile loop: Oddments: Load inputs: (0, 3): Bit 2: Unset: Bit 1: Unset
     "ldr h12, [x19, #0x0]\n"
-    "50:"  // Tile loop: Oddments: Load inputs: (2, 5): Bit 1: End
-    "fmla v28.8h, v8.8h, v12.8h\n"
+    "84:"  // Tile loop: Oddments: Load inputs: (0, 3): Bit 2: End
+    "fmla v17.8h, v2.8h, v12.8h\n"
+    "fmla v18.8h, v1.8h, v12.8h\n"
     "add x19, x13, XZR\n"
-    "fmla v24.8h, v5.8h, v12.8h\n"
-    "fmla v20.8h, v2.8h, v12.8h\n"
-    "tbz %x[n_channels], #1, 51f\n"
+    "fmla v19.8h, v0.8h, v12.8h\n"
+    "tbz %x[n_channels], #2, 86f\n"
+    "ldr d10, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 85f\n"
+    "ld1 { v10.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 88f\n"
+    "ld1 { v10.h }[6], [x19]\n"
+    "b 88f\n"
+    "85:"  // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 88f\n"
+    "ld1 { v10.h }[4], [x19]\n"
+    "b 88f\n"
+    "86:"  // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: Unset
+    "tbz %x[n_channels], #1, 87f\n"
     "ldr s10, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 52f\n"
+    "tbz %x[n_channels], #0, 88f\n"
     "ld1 { v10.h }[2], [x19]\n"
-    "b 52f\n"
-    "51:"  // Tile loop: Oddments: Load inputs: (3, 0): Bit 1: Unset
+    "b 88f\n"
+    "87:"  // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: Unset: Bit 1: Unset
     "ldr h10, [x19, #0x0]\n"
-    "52:"  // Tile loop: Oddments: Load inputs: (3, 0): Bit 1: End
-    "fmla v27.8h, v6.8h, v10.8h\n"
-    "add x19, x12, x10\n"
-    "fmla v23.8h, v3.8h, v10.8h\n"
-    "fmla v19.8h, v0.8h, v10.8h\n"
-    "tbz %x[n_channels], #1, 53f\n"
+    "88:"  // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: End
+    "fmla v16.8h, v6.8h, v10.8h\n"
+    "fmla v20.8h, v3.8h, v10.8h\n"
+    "add x19, x11, x27\n"
+    "fmla v24.8h, v0.8h, v10.8h\n"
+    "tbz %x[n_channels], #2, 90f\n"
+    "ldr d11, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 89f\n"
+    "ld1 { v11.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 92f\n"
+    "ld1 { v11.h }[6], [x19]\n"
+    "b 92f\n"
+    "89:"  // Tile loop: Oddments: Load inputs: (3, 4): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 92f\n"
+    "ld1 { v11.h }[4], [x19]\n"
+    "b 92f\n"
+    "90:"  // Tile loop: Oddments: Load inputs: (3, 4): Bit 2: Unset
+    "tbz %x[n_channels], #1, 91f\n"
     "ldr s11, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 54f\n"
+    "tbz %x[n_channels], #0, 92f\n"
     "ld1 { v11.h }[2], [x19]\n"
-    "b 54f\n"
-    "53:"  // Tile loop: Oddments: Load inputs: (4, 2): Bit 1: Unset
+    "b 92f\n"
+    "91:"  // Tile loop: Oddments: Load inputs: (3, 4): Bit 2: Unset: Bit 1: Unset
     "ldr h11, [x19, #0x0]\n"
-    "54:"  // Tile loop: Oddments: Load inputs: (4, 2): Bit 1: End
-    "fmla v23.8h, v8.8h, v11.8h\n"
-    "add x19, x13, x27\n"
-    "fmla v22.8h, v7.8h, v11.8h\n"
-    "fmla v21.8h, v6.8h, v11.8h\n"
-    "fmla v19.8h, v5.8h, v11.8h\n"
-    "fmla v18.8h, v4.8h, v11.8h\n"
-    "fmla v17.8h, v3.8h, v11.8h\n"
-    "tbz %x[n_channels], #1, 55f\n"
+    "92:"  // Tile loop: Oddments: Load inputs: (3, 4): Bit 2: End
+    "fmla v22.8h, v8.8h, v11.8h\n"
+    "fmla v23.8h, v7.8h, v11.8h\n"
+    "add x19, x13, x24\n"
+    "fmla v26.8h, v5.8h, v11.8h\n"
+    "fmla v27.8h, v4.8h, v11.8h\n"
+    "fmla v30.8h, v2.8h, v11.8h\n"
+    "fmla v31.8h, v1.8h, v11.8h\n"
+    "tbz %x[n_channels], #2, 94f\n"
+    "ldr d12, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 93f\n"
+    "ld1 { v12.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 96f\n"
+    "ld1 { v12.h }[6], [x19]\n"
+    "b 96f\n"
+    "93:"  // Tile loop: Oddments: Load inputs: (2, 5): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 96f\n"
+    "ld1 { v12.h }[4], [x19]\n"
+    "b 96f\n"
+    "94:"  // Tile loop: Oddments: Load inputs: (2, 5): Bit 2: Unset
+    "tbz %x[n_channels], #1, 95f\n"
     "ldr s12, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 56f\n"
+    "tbz %x[n_channels], #0, 96f\n"
     "ld1 { v12.h }[2], [x19]\n"
-    "b 56f\n"
-    "55:"  // Tile loop: Oddments: Load inputs: (3, 5): Bit 1: Unset
+    "b 96f\n"
+    "95:"  // Tile loop: Oddments: Load inputs: (2, 5): Bit 2: Unset: Bit 1: Unset
     "ldr h12, [x19, #0x0]\n"
-    "56:"  // Tile loop: Oddments: Load inputs: (3, 5): Bit 1: End
-    "fmla v24.8h, v8.8h, v12.8h\n"
-    "add x19, x11, x10\n"
-    "fmla v20.8h, v5.8h, v12.8h\n"
-    "fmla v16.8h, v2.8h, v12.8h\n"
-    "tbz %x[n_channels], #1, 57f\n"
+    "96:"  // Tile loop: Oddments: Load inputs: (2, 5): Bit 2: End
+    "fmla v19.8h, v8.8h, v12.8h\n"
+    "fmla v23.8h, v5.8h, v12.8h\n"
+    "add x19, x11, XZR\n"
+    "fmla v27.8h, v2.8h, v12.8h\n"
+    "tbz %x[n_channels], #2, 98f\n"
+    "ldr d10, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 97f\n"
+    "ld1 { v10.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 100f\n"
+    "ld1 { v10.h }[6], [x19]\n"
+    "b 100f\n"
+    "97:"  // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 100f\n"
+    "ld1 { v10.h }[4], [x19]\n"
+    "b 100f\n"
+    "98:"  // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: Unset
+    "tbz %x[n_channels], #1, 99f\n"
     "ldr s10, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 58f\n"
+    "tbz %x[n_channels], #0, 100f\n"
     "ld1 { v10.h }[2], [x19]\n"
-    "b 58f\n"
-    "57:"  // Tile loop: Oddments: Load inputs: (5, 2): Bit 1: Unset
+    "b 100f\n"
+    "99:"  // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: Unset: Bit 1: Unset
     "ldr h10, [x19, #0x0]\n"
-    "58:"  // Tile loop: Oddments: Load inputs: (5, 2): Bit 1: End
-    "fmla v19.8h, v8.8h, v10.8h\n"
-    "add x19, x12, x9\n"
-    "fmla v18.8h, v7.8h, v10.8h\n"
-    "fmla v17.8h, v6.8h, v10.8h\n"
-    "tbz %x[n_channels], #1, 59f\n"
+    "100:"  // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: End
+    "fmla v20.8h, v6.8h, v10.8h\n"
+    "fmla v24.8h, v3.8h, v10.8h\n"
+    "add x19, x28, x16\n"
+    "fmla v28.8h, v0.8h, v10.8h\n"
+    "tbz %x[n_channels], #2, 102f\n"
+    "ldr d11, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 101f\n"
+    "ld1 { v11.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 104f\n"
+    "ld1 { v11.h }[6], [x19]\n"
+    "b 104f\n"
+    "101:"  // Tile loop: Oddments: Load inputs: (4, 2): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 104f\n"
+    "ld1 { v11.h }[4], [x19]\n"
+    "b 104f\n"
+    "102:"  // Tile loop: Oddments: Load inputs: (4, 2): Bit 2: Unset
+    "tbz %x[n_channels], #1, 103f\n"
     "ldr s11, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 60f\n"
+    "tbz %x[n_channels], #0, 104f\n"
     "ld1 { v11.h }[2], [x19]\n"
-    "b 60f\n"
-    "59:"  // Tile loop: Oddments: Load inputs: (4, 3): Bit 1: Unset
+    "b 104f\n"
+    "103:"  // Tile loop: Oddments: Load inputs: (4, 2): Bit 2: Unset: Bit 1: Unset
     "ldr h11, [x19, #0x0]\n"
-    "60:"  // Tile loop: Oddments: Load inputs: (4, 3): Bit 1: End
-    "fmla v22.8h, v8.8h, v11.8h\n"
-    "add x19, x11, x9\n"
-    "fmla v21.8h, v7.8h, v11.8h\n"
-    "fmla v20.8h, v6.8h, v11.8h\n"
-    "fmla v18.8h, v5.8h, v11.8h\n"
-    "fmla v17.8h, v4.8h, v11.8h\n"
-    "fmla v16.8h, v3.8h, v11.8h\n"
-    "tbz %x[n_channels], #1, 61f\n"
+    "104:"  // Tile loop: Oddments: Load inputs: (4, 2): Bit 2: End
+    "fmla v24.8h, v8.8h, v11.8h\n"
+    "fmla v25.8h, v7.8h, v11.8h\n"
+    "add x19, x11, x24\n"
+    "fmla v26.8h, v6.8h, v11.8h\n"
+    "fmla v28.8h, v5.8h, v11.8h\n"
+    "fmla v29.8h, v4.8h, v11.8h\n"
+    "fmla v30.8h, v3.8h, v11.8h\n"
+    "tbz %x[n_channels], #2, 106f\n"
+    "ldr d12, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 105f\n"
+    "ld1 { v12.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 108f\n"
+    "ld1 { v12.h }[6], [x19]\n"
+    "b 108f\n"
+    "105:"  // Tile loop: Oddments: Load inputs: (3, 5): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 108f\n"
+    "ld1 { v12.h }[4], [x19]\n"
+    "b 108f\n"
+    "106:"  // Tile loop: Oddments: Load inputs: (3, 5): Bit 2: Unset
+    "tbz %x[n_channels], #1, 107f\n"
     "ldr s12, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 62f\n"
+    "tbz %x[n_channels], #0, 108f\n"
     "ld1 { v12.h }[2], [x19]\n"
-    "b 62f\n"
-    "61:"  // Tile loop: Oddments: Load inputs: (5, 3): Bit 1: Unset
+    "b 108f\n"
+    "107:"  // Tile loop: Oddments: Load inputs: (3, 5): Bit 2: Unset: Bit 1: Unset
     "ldr h12, [x19, #0x0]\n"
-    "62:"  // Tile loop: Oddments: Load inputs: (5, 3): Bit 1: End
-    "fmla v18.8h, v8.8h, v12.8h\n"
-    "add x19, x15, x6\n"
-    "fmla v17.8h, v7.8h, v12.8h\n"
-    "fmla v16.8h, v6.8h, v12.8h\n"
-    "tbz %x[n_channels], #1, 63f\n"
+    "108:"  // Tile loop: Oddments: Load inputs: (3, 5): Bit 2: End
+    "fmla v23.8h, v8.8h, v12.8h\n"
+    "fmla v27.8h, v5.8h, v12.8h\n"
+    "add x19, x25, x16\n"
+    "fmla v31.8h, v2.8h, v12.8h\n"
+    "tbz %x[n_channels], #2, 110f\n"
+    "ldr d10, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 109f\n"
+    "ld1 { v10.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 112f\n"
+    "ld1 { v10.h }[6], [x19]\n"
+    "b 112f\n"
+    "109:"  // Tile loop: Oddments: Load inputs: (5, 2): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 112f\n"
+    "ld1 { v10.h }[4], [x19]\n"
+    "b 112f\n"
+    "110:"  // Tile loop: Oddments: Load inputs: (5, 2): Bit 2: Unset
+    "tbz %x[n_channels], #1, 111f\n"
     "ldr s10, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 64f\n"
+    "tbz %x[n_channels], #0, 112f\n"
     "ld1 { v10.h }[2], [x19]\n"
-    "b 64f\n"
-    "63:"  // Tile loop: Oddments: Load inputs: (1, 1): Bit 1: Unset
+    "b 112f\n"
+    "111:"  // Tile loop: Oddments: Load inputs: (5, 2): Bit 2: Unset: Bit 1: Unset
     "ldr h10, [x19, #0x0]\n"
-    "64:"  // Tile loop: Oddments: Load inputs: (1, 1): Bit 1: End
-    "fmla v31.8h, v4.8h, v10.8h\n"
-    "add x19, x15, x28\n"
-    "fmla v30.8h, v3.8h, v10.8h\n"
-    "fmla v27.8h, v1.8h, v10.8h\n"
-    "fmla v26.8h, v0.8h, v10.8h\n"
-    "tbz %x[n_channels], #1, 65f\n"
+    "112:"  // Tile loop: Oddments: Load inputs: (5, 2): Bit 2: End
+    "fmla v28.8h, v8.8h, v10.8h\n"
+    "fmla v29.8h, v7.8h, v10.8h\n"
+    "add x19, x28, x10\n"
+    "fmla v30.8h, v6.8h, v10.8h\n"
+    "tbz %x[n_channels], #2, 114f\n"
+    "ldr d11, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 113f\n"
+    "ld1 { v11.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 116f\n"
+    "ld1 { v11.h }[6], [x19]\n"
+    "b 116f\n"
+    "113:"  // Tile loop: Oddments: Load inputs: (4, 3): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 116f\n"
+    "ld1 { v11.h }[4], [x19]\n"
+    "b 116f\n"
+    "114:"  // Tile loop: Oddments: Load inputs: (4, 3): Bit 2: Unset
+    "tbz %x[n_channels], #1, 115f\n"
     "ldr s11, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 66f\n"
+    "tbz %x[n_channels], #0, 116f\n"
     "ld1 { v11.h }[2], [x19]\n"
-    "b 66f\n"
-    "65:"  // Tile loop: Oddments: Load inputs: (1, 4): Bit 1: Unset
+    "b 116f\n"
+    "115:"  // Tile loop: Oddments: Load inputs: (4, 3): Bit 2: Unset: Bit 1: Unset
     "ldr h11, [x19, #0x0]\n"
-    "66:"  // Tile loop: Oddments: Load inputs: (1, 4): Bit 1: End
+    "116:"  // Tile loop: Oddments: Load inputs: (4, 3): Bit 2: End
+    "fmla v25.8h, v8.8h, v11.8h\n"
+    "fmla v26.8h, v7.8h, v11.8h\n"
+    "add x19, x25, x10\n"
+    "fmla v27.8h, v6.8h, v11.8h\n"
     "fmla v29.8h, v5.8h, v11.8h\n"
-    "add x19, x12, x6\n"
-    "fmla v28.8h, v4.8h, v11.8h\n"
-    "fmla v25.8h, v2.8h, v11.8h\n"
-    "fmla v24.8h, v1.8h, v11.8h\n"
-    "tbz %x[n_channels], #1, 67f\n"
+    "fmla v30.8h, v4.8h, v11.8h\n"
+    "fmla v31.8h, v3.8h, v11.8h\n"
+    "tbz %x[n_channels], #2, 118f\n"
+    "ldr d12, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 117f\n"
+    "ld1 { v12.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 120f\n"
+    "ld1 { v12.h }[6], [x19]\n"
+    "b 120f\n"
+    "117:"  // Tile loop: Oddments: Load inputs: (5, 3): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 120f\n"
+    "ld1 { v12.h }[4], [x19]\n"
+    "b 120f\n"
+    "118:"  // Tile loop: Oddments: Load inputs: (5, 3): Bit 2: Unset
+    "tbz %x[n_channels], #1, 119f\n"
     "ldr s12, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 68f\n"
+    "tbz %x[n_channels], #0, 120f\n"
     "ld1 { v12.h }[2], [x19]\n"
-    "b 68f\n"
-    "67:"  // Tile loop: Oddments: Load inputs: (4, 1): Bit 1: Unset
+    "b 120f\n"
+    "119:"  // Tile loop: Oddments: Load inputs: (5, 3): Bit 2: Unset: Bit 1: Unset
     "ldr h12, [x19, #0x0]\n"
-    "68:"  // Tile loop: Oddments: Load inputs: (4, 1): Bit 1: End
-    "fmla v23.8h, v7.8h, v12.8h\n"
-    "add x19, x12, x28\n"
-    "fmla v22.8h, v6.8h, v12.8h\n"
-    "fmla v19.8h, v4.8h, v12.8h\n"
-    "fmla v18.8h, v3.8h, v12.8h\n"
-    "tbz %x[n_channels], #1, 69f\n"
+    "120:"  // Tile loop: Oddments: Load inputs: (5, 3): Bit 2: End
+    "fmla v29.8h, v8.8h, v12.8h\n"
+    "fmla v30.8h, v7.8h, v12.8h\n"
+    "add x19, x14, x5\n"
+    "fmla v31.8h, v6.8h, v12.8h\n"
+    "tbz %x[n_channels], #2, 122f\n"
+    "ldr d10, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 121f\n"
+    "ld1 { v10.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 124f\n"
+    "ld1 { v10.h }[6], [x19]\n"
+    "b 124f\n"
+    "121:"  // Tile loop: Oddments: Load inputs: (1, 1): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 124f\n"
+    "ld1 { v10.h }[4], [x19]\n"
+    "b 124f\n"
+    "122:"  // Tile loop: Oddments: Load inputs: (1, 1): Bit 2: Unset
+    "tbz %x[n_channels], #1, 123f\n"
     "ldr s10, [x19], #0x4\n"
-    "tbz %x[n_channels], #0, 70f\n"
+    "tbz %x[n_channels], #0, 124f\n"
     "ld1 { v10.h }[2], [x19]\n"
-    "b 70f\n"
-    "69:"  // Tile loop: Oddments: Load inputs: (4, 4): Bit 1: Unset
+    "b 124f\n"
+    "123:"  // Tile loop: Oddments: Load inputs: (1, 1): Bit 2: Unset: Bit 1: Unset
     "ldr h10, [x19, #0x0]\n"
-    "70:"  // Tile loop: Oddments: Load inputs: (4, 4): Bit 1: End
-    "fmla v21.8h, v8.8h, v10.8h\n"
-    "fmla v20.8h, v7.8h, v10.8h\n"
-    "fmla v17.8h, v5.8h, v10.8h\n"
+    "124:"  // Tile loop: Oddments: Load inputs: (1, 1): Bit 2: End
     "fmla v16.8h, v4.8h, v10.8h\n"
-    "fmax v31.8h, v31.8h, v15.8h\n"
-    "fmax v30.8h, v30.8h, v15.8h\n"
-    "fmax v29.8h, v29.8h, v15.8h\n"
-    "fmin v31.8h, v31.8h, v14.8h\n"
-    "fmin v30.8h, v30.8h, v14.8h\n"
-    "fmin v29.8h, v29.8h, v14.8h\n"
-    "fmax v28.8h, v28.8h, v15.8h\n"
-    "fmax v27.8h, v27.8h, v15.8h\n"
-    "fmax v26.8h, v26.8h, v15.8h\n"
-    "fmin v28.8h, v28.8h, v14.8h\n"
-    "fmin v27.8h, v27.8h, v14.8h\n"
-    "fmin v26.8h, v26.8h, v14.8h\n"
-    "fmax v25.8h, v25.8h, v15.8h\n"
-    "fmax v24.8h, v24.8h, v15.8h\n"
-    "fmax v23.8h, v23.8h, v15.8h\n"
-    "fmin v25.8h, v25.8h, v14.8h\n"
-    "fmin v24.8h, v24.8h, v14.8h\n"
-    "fmin v23.8h, v23.8h, v14.8h\n"
-    "fmax v22.8h, v22.8h, v15.8h\n"
-    "fmax v21.8h, v21.8h, v15.8h\n"
-    "fmax v20.8h, v20.8h, v15.8h\n"
-    "fmin v22.8h, v22.8h, v14.8h\n"
-    "fmin v21.8h, v21.8h, v14.8h\n"
-    "fmin v20.8h, v20.8h, v14.8h\n"
-    "fmax v19.8h, v19.8h, v15.8h\n"
-    "fmax v18.8h, v18.8h, v15.8h\n"
-    "fmax v17.8h, v17.8h, v15.8h\n"
-    "fmin v19.8h, v19.8h, v14.8h\n"
-    "fmin v18.8h, v18.8h, v14.8h\n"
-    "fmin v17.8h, v17.8h, v14.8h\n"
+    "fmla v17.8h, v3.8h, v10.8h\n"
+    "add x19, x14, x27\n"
+    "fmla v20.8h, v1.8h, v10.8h\n"
+    "fmla v21.8h, v0.8h, v10.8h\n"
+    "tbz %x[n_channels], #2, 126f\n"
+    "ldr d11, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 125f\n"
+    "ld1 { v11.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 128f\n"
+    "ld1 { v11.h }[6], [x19]\n"
+    "b 128f\n"
+    "125:"  // Tile loop: Oddments: Load inputs: (1, 4): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 128f\n"
+    "ld1 { v11.h }[4], [x19]\n"
+    "b 128f\n"
+    "126:"  // Tile loop: Oddments: Load inputs: (1, 4): Bit 2: Unset
+    "tbz %x[n_channels], #1, 127f\n"
+    "ldr s11, [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 128f\n"
+    "ld1 { v11.h }[2], [x19]\n"
+    "b 128f\n"
+    "127:"  // Tile loop: Oddments: Load inputs: (1, 4): Bit 2: Unset: Bit 1: Unset
+    "ldr h11, [x19, #0x0]\n"
+    "128:"  // Tile loop: Oddments: Load inputs: (1, 4): Bit 2: End
+    "fmla v18.8h, v5.8h, v11.8h\n"
+    "fmla v19.8h, v4.8h, v11.8h\n"
+    "add x19, x28, x5\n"
+    "fmla v22.8h, v2.8h, v11.8h\n"
+    "fmla v23.8h, v1.8h, v11.8h\n"
+    "tbz %x[n_channels], #2, 130f\n"
+    "ldr d12, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 129f\n"
+    "ld1 { v12.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 132f\n"
+    "ld1 { v12.h }[6], [x19]\n"
+    "b 132f\n"
+    "129:"  // Tile loop: Oddments: Load inputs: (4, 1): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 132f\n"
+    "ld1 { v12.h }[4], [x19]\n"
+    "b 132f\n"
+    "130:"  // Tile loop: Oddments: Load inputs: (4, 1): Bit 2: Unset
+    "tbz %x[n_channels], #1, 131f\n"
+    "ldr s12, [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 132f\n"
+    "ld1 { v12.h }[2], [x19]\n"
+    "b 132f\n"
+    "131:"  // Tile loop: Oddments: Load inputs: (4, 1): Bit 2: Unset: Bit 1: Unset
+    "ldr h12, [x19, #0x0]\n"
+    "132:"  // Tile loop: Oddments: Load inputs: (4, 1): Bit 2: End
+    "fmla v24.8h, v7.8h, v12.8h\n"
+    "fmla v25.8h, v6.8h, v12.8h\n"
+    "add x19, x28, x27\n"
+    "fmla v28.8h, v4.8h, v12.8h\n"
+    "fmla v29.8h, v3.8h, v12.8h\n"
+    "tbz %x[n_channels], #2, 134f\n"
+    "ldr d10, [x19], #0x8\n"
+    "tbz %x[n_channels], #1, 133f\n"
+    "ld1 { v10.s }[2], [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 136f\n"
+    "ld1 { v10.h }[6], [x19]\n"
+    "b 136f\n"
+    "133:"  // Tile loop: Oddments: Load inputs: (4, 4): Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 136f\n"
+    "ld1 { v10.h }[4], [x19]\n"
+    "b 136f\n"
+    "134:"  // Tile loop: Oddments: Load inputs: (4, 4): Bit 2: Unset
+    "tbz %x[n_channels], #1, 135f\n"
+    "ldr s10, [x19], #0x4\n"
+    "tbz %x[n_channels], #0, 136f\n"
+    "ld1 { v10.h }[2], [x19]\n"
+    "b 136f\n"
+    "135:"  // Tile loop: Oddments: Load inputs: (4, 4): Bit 2: Unset: Bit 1: Unset
+    "ldr h10, [x19, #0x0]\n"
+    "136:"  // Tile loop: Oddments: Load inputs: (4, 4): Bit 2: End
+    "fmla v26.8h, v8.8h, v10.8h\n"
+    "fmla v27.8h, v7.8h, v10.8h\n"
     "fmax v16.8h, v16.8h, v15.8h\n"
+    "fmla v30.8h, v5.8h, v10.8h\n"
+    "fmla v31.8h, v4.8h, v10.8h\n"
+    "fmax v17.8h, v17.8h, v15.8h\n"
+    "fmax v18.8h, v18.8h, v15.8h\n"
+    "fmax v19.8h, v19.8h, v15.8h\n"
+    "fmax v20.8h, v20.8h, v15.8h\n"
+    "fmax v21.8h, v21.8h, v15.8h\n"
+    "fmax v22.8h, v22.8h, v15.8h\n"
+    "fmax v23.8h, v23.8h, v15.8h\n"
+    "fmax v24.8h, v24.8h, v15.8h\n"
+    "fmax v25.8h, v25.8h, v15.8h\n"
+    "fmax v26.8h, v26.8h, v15.8h\n"
+    "fmax v27.8h, v27.8h, v15.8h\n"
+    "fmax v28.8h, v28.8h, v15.8h\n"
+    "fmax v29.8h, v29.8h, v15.8h\n"
+    "fmax v30.8h, v30.8h, v15.8h\n"
+    "fmax v31.8h, v31.8h, v15.8h\n"
     "fmin v16.8h, v16.8h, v14.8h\n"
-    "tbz %x[n_channels], #1, 71f\n"
-    "mov x19, x16\n"
-    "st1 { v31.s }[0], [x19], x17\n"
-    "add x16, x16, #0x4\n"
-    "st1 { v30.s }[0], [x19], x17\n"
-    "mov x21, x26\n"
-    "st1 { v29.s }[0], [x19], x17\n"
-    "st1 { v27.s }[0], [x21], x17\n"
+    "fmin v17.8h, v17.8h, v14.8h\n"
+    "fmin v18.8h, v18.8h, v14.8h\n"
+    "fmin v19.8h, v19.8h, v14.8h\n"
+    "fmin v20.8h, v20.8h, v14.8h\n"
+    "fmin v21.8h, v21.8h, v14.8h\n"
+    "fmin v22.8h, v22.8h, v14.8h\n"
+    "fmin v23.8h, v23.8h, v14.8h\n"
+    "fmin v24.8h, v24.8h, v14.8h\n"
+    "fmin v25.8h, v25.8h, v14.8h\n"
+    "fmin v26.8h, v26.8h, v14.8h\n"
+    "fmin v27.8h, v27.8h, v14.8h\n"
+    "fmin v28.8h, v28.8h, v14.8h\n"
+    "fmin v29.8h, v29.8h, v14.8h\n"
+    "fmin v30.8h, v30.8h, v14.8h\n"
+    "fmin v31.8h, v31.8h, v14.8h\n"
+    "tbz %x[n_channels], #2, 138f\n"
+    "mov x22, x17\n"
+    "mov x21, x9\n"
+    "mov x20, x26\n"
+    "st1 { v16.d }[0], [x22], x6\n"
+    "mov x19, x23\n"
+    "st1 { v20.d }[0], [x21], x6\n"
+    "add x17, x17, #0x8\n"
+    "st1 { v24.d }[0], [x20], x6\n"
+    "add x9, x9, #0x8\n"
+    "add x26, x26, #0x8\n"
+    "st1 { v28.d }[0], [x19], x6\n"
+    "add x23, x23, #0x8\n"
+    "st1 { v17.d }[0], [x22], x6\n"
+    "st1 { v21.d }[0], [x21], x6\n"
+    "st1 { v25.d }[0], [x20], x6\n"
+    "st1 { v29.d }[0], [x19], x6\n"
+    "st1 { v18.d }[0], [x22], x6\n"
+    "st1 { v22.d }[0], [x21], x6\n"
+    "st1 { v26.d }[0], [x20], x6\n"
+    "st1 { v30.d }[0], [x19], x6\n"
+    "st1 { v19.d }[0], [x22]\n"
+    "st1 { v23.d }[0], [x21]\n"
+    "st1 { v27.d }[0], [x20]\n"
+    "st1 { v31.d }[0], [x19]\n"
+    "tbz %x[n_channels], #1, 137f\n"
+    "mov x22, x17\n"
+    "mov x21, x9\n"
+    "mov x20, x26\n"
+    "mov x19, x23\n"
+    "st1 { v16.s }[2], [x22], x6\n"
+    "st1 { v20.s }[2], [x21], x6\n"
+    "add x17, x17, #0x4\n"
+    "add x9, x9, #0x4\n"
+    "st1 { v24.s }[2], [x20], x6\n"
     "add x26, x26, #0x4\n"
-    "st1 { v28.s }[0], [x19]\n"
-    "mov x20, x25\n"
-    "st1 { v26.s }[0], [x21], x17\n"
-    "add x25, x25, #0x4\n"
-    "st1 { v25.s }[0], [x21], x17\n"
-    "mov x19, x24\n"
-    "st1 { v24.s }[0], [x21]\n"
-    "add x24, x24, #0x4\n"
-    "st1 { v23.s }[0], [x20], x17\n"
-    "st1 { v22.s }[0], [x20], x17\n"
-    "st1 { v21.s }[0], [x20], x17\n"
-    "st1 { v20.s }[0], [x20]\n"
-    "st1 { v19.s }[0], [x19], x17\n"
-    "st1 { v18.s }[0], [x19], x17\n"
-    "st1 { v17.s }[0], [x19], x17\n"
-    "st1 { v16.s }[0], [x19]\n"
-    "tbz %x[n_channels], #0, 72f\n"
-    "mov x22, x16\n"
-    "st1 { v31.h }[2], [x22], x17\n"
-    "mov x21, x26\n"
-    "st1 { v30.h }[2], [x22], x17\n"
-    "st1 { v27.h }[2], [x21], x17\n"
-    "mov x20, x25\n"
-    "st1 { v29.h }[2], [x22], x17\n"
-    "mov x19, x24\n"
-    "st1 { v28.h }[2], [x22]\n"
-    "st1 { v26.h }[2], [x21], x17\n"
-    "st1 { v25.h }[2], [x21], x17\n"
-    "st1 { v24.h }[2], [x21]\n"
-    "st1 { v23.h }[2], [x20], x17\n"
-    "st1 { v22.h }[2], [x20], x17\n"
-    "st1 { v21.h }[2], [x20], x17\n"
-    "st1 { v20.h }[2], [x20]\n"
-    "st1 { v19.h }[2], [x19], x17\n"
-    "st1 { v18.h }[2], [x19], x17\n"
-    "st1 { v17.h }[2], [x19], x17\n"
-    "st1 { v16.h }[2], [x19]\n"
-    "b 72f\n"
-    "71:"  // Tile loop: Oddments: Store: Bit 1: Unset
-    "mov x22, x16\n"
-    "st1 { v31.h }[0], [x22], x17\n"
-    "mov x21, x26\n"
-    "mov x20, x25\n"
-    "st1 { v30.h }[0], [x22], x17\n"
-    "st1 { v27.h }[0], [x21], x17\n"
-    "mov x19, x24\n"
-    "st1 { v29.h }[0], [x22], x17\n"
-    "st1 { v28.h }[0], [x22]\n"
-    "st1 { v26.h }[0], [x21], x17\n"
-    "st1 { v25.h }[0], [x21], x17\n"
-    "st1 { v24.h }[0], [x21]\n"
-    "st1 { v23.h }[0], [x20], x17\n"
-    "st1 { v22.h }[0], [x20], x17\n"
-    "st1 { v21.h }[0], [x20], x17\n"
-    "st1 { v20.h }[0], [x20]\n"
-    "st1 { v19.h }[0], [x19], x17\n"
-    "st1 { v18.h }[0], [x19], x17\n"
-    "st1 { v17.h }[0], [x19], x17\n"
-    "st1 { v16.h }[0], [x19]\n"
-    "72:"  // Tile loop: Oddments: Store: Bit 1: End
+    "add x23, x23, #0x4\n"
+    "st1 { v28.s }[2], [x19], x6\n"
+    "st1 { v17.s }[2], [x22], x6\n"
+    "st1 { v21.s }[2], [x21], x6\n"
+    "st1 { v25.s }[2], [x20], x6\n"
+    "st1 { v29.s }[2], [x19], x6\n"
+    "st1 { v18.s }[2], [x22], x6\n"
+    "st1 { v22.s }[2], [x21], x6\n"
+    "st1 { v26.s }[2], [x20], x6\n"
+    "st1 { v30.s }[2], [x19], x6\n"
+    "st1 { v19.s }[2], [x22]\n"
+    "st1 { v23.s }[2], [x21]\n"
+    "st1 { v27.s }[2], [x20]\n"
+    "st1 { v31.s }[2], [x19]\n"
+    "tbz %x[n_channels], #0, 140f\n"
+    "mov x22, x17\n"
+    "mov x21, x9\n"
+    "mov x20, x26\n"
+    "mov x19, x23\n"
+    "st1 { v16.h }[6], [x22], x6\n"
+    "st1 { v20.h }[6], [x21], x6\n"
+    "st1 { v24.h }[6], [x20], x6\n"
+    "st1 { v28.h }[6], [x19], x6\n"
+    "st1 { v17.h }[6], [x22], x6\n"
+    "st1 { v21.h }[6], [x21], x6\n"
+    "st1 { v25.h }[6], [x20], x6\n"
+    "st1 { v29.h }[6], [x19], x6\n"
+    "st1 { v18.h }[6], [x22], x6\n"
+    "st1 { v22.h }[6], [x21], x6\n"
+    "st1 { v26.h }[6], [x20], x6\n"
+    "st1 { v30.h }[6], [x19], x6\n"
+    "st1 { v19.h }[6], [x22]\n"
+    "st1 { v23.h }[6], [x21]\n"
+    "st1 { v27.h }[6], [x20]\n"
+    "st1 { v31.h }[6], [x19]\n"
+    "b 140f\n"
+    "137:"  // Tile loop: Oddments: Store: Bit 2: Bit 1: Unset
+    "tbz %x[n_channels], #0, 140f\n"
+    "mov x22, x17\n"
+    "mov x21, x9\n"
+    "st1 { v16.h }[4], [x22], x6\n"
+    "mov x20, x26\n"
+    "mov x19, x23\n"
+    "st1 { v20.h }[4], [x21], x6\n"
+    "st1 { v24.h }[4], [x20], x6\n"
+    "st1 { v28.h }[4], [x19], x6\n"
+    "st1 { v17.h }[4], [x22], x6\n"
+    "st1 { v21.h }[4], [x21], x6\n"
+    "st1 { v25.h }[4], [x20], x6\n"
+    "st1 { v29.h }[4], [x19], x6\n"
+    "st1 { v18.h }[4], [x22], x6\n"
+    "st1 { v22.h }[4], [x21], x6\n"
+    "st1 { v26.h }[4], [x20], x6\n"
+    "st1 { v30.h }[4], [x19], x6\n"
+    "st1 { v19.h }[4], [x22]\n"
+    "st1 { v23.h }[4], [x21]\n"
+    "st1 { v27.h }[4], [x20]\n"
+    "st1 { v31.h }[4], [x19]\n"
+    "b 140f\n"
+    "138:"  // Tile loop: Oddments: Store: Bit 2: Unset
+    "tbz %x[n_channels], #1, 139f\n"
+    "mov x22, x17\n"
+    "mov x21, x9\n"
+    "st1 { v16.s }[0], [x22], x6\n"
+    "mov x20, x26\n"
+    "mov x19, x23\n"
+    "st1 { v20.s }[0], [x21], x6\n"
+    "st1 { v24.s }[0], [x20], x6\n"
+    "add x17, x17, #0x4\n"
+    "add x9, x9, #0x4\n"
+    "st1 { v28.s }[0], [x19], x6\n"
+    "add x26, x26, #0x4\n"
+    "add x23, x23, #0x4\n"
+    "st1 { v17.s }[0], [x22], x6\n"
+    "st1 { v21.s }[0], [x21], x6\n"
+    "st1 { v25.s }[0], [x20], x6\n"
+    "st1 { v29.s }[0], [x19], x6\n"
+    "st1 { v18.s }[0], [x22], x6\n"
+    "st1 { v22.s }[0], [x21], x6\n"
+    "st1 { v26.s }[0], [x20], x6\n"
+    "st1 { v30.s }[0], [x19], x6\n"
+    "st1 { v19.s }[0], [x22]\n"
+    "st1 { v23.s }[0], [x21]\n"
+    "st1 { v27.s }[0], [x20]\n"
+    "st1 { v31.s }[0], [x19]\n"
+    "tbz %x[n_channels], #0, 140f\n"
+    "mov x22, x17\n"
+    "mov x21, x9\n"
+    "mov x20, x26\n"
+    "mov x19, x23\n"
+    "st1 { v16.h }[2], [x22], x6\n"
+    "st1 { v20.h }[2], [x21], x6\n"
+    "st1 { v24.h }[2], [x20], x6\n"
+    "st1 { v28.h }[2], [x19], x6\n"
+    "st1 { v17.h }[2], [x22], x6\n"
+    "st1 { v21.h }[2], [x21], x6\n"
+    "st1 { v25.h }[2], [x20], x6\n"
+    "st1 { v29.h }[2], [x19], x6\n"
+    "st1 { v18.h }[2], [x22], x6\n"
+    "st1 { v22.h }[2], [x21], x6\n"
+    "st1 { v26.h }[2], [x20], x6\n"
+    "st1 { v30.h }[2], [x19], x6\n"
+    "st1 { v19.h }[2], [x22]\n"
+    "st1 { v23.h }[2], [x21]\n"
+    "st1 { v27.h }[2], [x20]\n"
+    "st1 { v31.h }[2], [x19]\n"
+    "b 140f\n"
+    "139:"  // Tile loop: Oddments: Store: Bit 2: Unset: Bit 1: Unset
+    "mov x22, x17\n"
+    "mov x21, x9\n"
+    "st1 { v16.h }[0], [x22], x6\n"
+    "mov x20, x26\n"
+    "mov x19, x23\n"
+    "st1 { v20.h }[0], [x21], x6\n"
+    "st1 { v24.h }[0], [x20], x6\n"
+    "st1 { v28.h }[0], [x19], x6\n"
+    "st1 { v17.h }[0], [x22], x6\n"
+    "st1 { v21.h }[0], [x21], x6\n"
+    "st1 { v25.h }[0], [x20], x6\n"
+    "st1 { v29.h }[0], [x19], x6\n"
+    "st1 { v18.h }[0], [x22], x6\n"
+    "st1 { v22.h }[0], [x21], x6\n"
+    "st1 { v26.h }[0], [x20], x6\n"
+    "st1 { v30.h }[0], [x19], x6\n"
+    "st1 { v19.h }[0], [x22]\n"
+    "st1 { v23.h }[0], [x21]\n"
+    "st1 { v27.h }[0], [x20]\n"
+    "st1 { v31.h }[0], [x19]\n"
+    "140:"  // Tile loop: Oddments: Store: Bit 2: End
 
-    "73:"  // Tile loop: End
-    "ldr x4, [%x[params_struct], %[offsetof_args_tile_i]]\n"
-    "add x21, x4, #0x1\n"
-    "ldr x26, [%x[params_struct], %[offsetof_args_tile_j]]\n"
-    "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
-    "add x26, x26, #0x1\n"
+    "141:"  // Tile loop: End
+    "ldr x25, [%x[params_struct], %[offsetof_args_tile_j]]\n"
+    "ldr x26, [%x[params_struct], %[offsetof_args_tile_i]]\n"
+    "add x25, x25, #0x1\n"
+    "add x20, x26, #0x1\n"
     "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
+    "cmp x25, x19\n"
+    "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
+    "csel x26, x26, x20, LT\n"
+    "csel x25, x25, XZR, LT\n"
     "cmp x26, x19\n"
-    "csel x26, x26, XZR, LT\n"
-    "csel x4, x4, x21, LT\n"
-    "cmp x4, x20\n"
     "blt 1b\n"
     :
     : [n_channels] "r" ((unsigned long) n_channels), [offsetof_args_inptr] "I" (offsetof(Args, inptr)), [offsetof_args_ld_input_col] "I" (offsetof(Args, ld_input_col)), [offsetof_args_ld_input_row] "I" (offsetof(Args, ld_input_row)), [offsetof_args_ld_output_col] "I" (offsetof(Args, ld_output_col)), [offsetof_args_ld_output_row] "I" (offsetof(Args, ld_output_row)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_n_tile_cols] "I" (offsetof(Args, n_tile_cols)), [offsetof_args_n_tile_rows] "I" (offsetof(Args, n_tile_rows)), [offsetof_args_outptr] "I" (offsetof(Args, outptr)), [offsetof_args_params] "I" (offsetof(Args, params)), [offsetof_args_tile_i] "I" (offsetof(Args, tile_i)), [offsetof_args_tile_j] "I" (offsetof(Args, tile_j)), [params_struct] "r" (&params_struct)
-    : "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", "x4", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+    : "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", "x5", "x6", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
   );
 }
author	Freddie Liardet <frederick.liardet@arm.com>	2021-08-04 11:57:37 +0100
committer	Sheri Zhang <sheri.zhang@arm.com>	2021-08-04 15:54:00 +0000
commit	6269f87f5caca3a83a4822e0556fbf284158615c (patch)
tree	0f814b2f6356aa073529a8d00c94913e006d1f6d /src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
parent	38ce29d5dce9a09769b55b2717d105f984247e31 (diff)
download	ComputeLibrary-6269f87f5caca3a83a4822e0556fbf284158615c.tar.gz