From 6269f87f5caca3a83a4822e0556fbf284158615c Mon Sep 17 00:00:00 2001 From: Freddie Liardet Date: Wed, 4 Aug 2021 11:57:37 +0100 Subject: Fix depthwise convolution assembly kernels Resolves: COMPMID-4710 Change-Id: I35b964731aeed0e6f4f873f59341bee48e4a41fd Signed-off-by: Freddie Liardet Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6039 Tested-by: Arm Jenkins Reviewed-by: Gian Marco Iodice Reviewed-by: Michele Di Giorgio Comments-Addressed: Arm Jenkins --- .../generic_direct.cpp | 886 +++++++++++++-------- 1 file changed, 541 insertions(+), 345 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp index 99f46015aa..143070cc7e 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst/generic_direct.cpp @@ -87,438 +87,634 @@ void a64_fp16_nhwc_3x3_s1_output2x2_mla_depthfirst_direct_impl( ); __asm__ __volatile__( - "mov x17, #0x0\n" - "mov x16, #0x0\n" + "mov x22, #0x0\n" + "mov x21, #0x0\n" "1:" // Tile loop - "str x17, [%x[params_struct], %[offsetof_args_tile_i]]\n" + "str x22, [%x[params_struct], %[offsetof_args_tile_i]]\n" + "mov x26, #0x2\n" "mov x25, #0x2\n" - "str x16, [%x[params_struct], %[offsetof_args_tile_j]]\n" - "mov x15, #0x2\n" - "ldr x14, [%x[params_struct], %[offsetof_args_params]]\n" - "add x24, %x[params_struct], %[offsetof_args_min]\n" - "ldr x23, [%x[params_struct], %[offsetof_args_ld_input_row]]\n" - "add x21, %x[params_struct], %[offsetof_args_max]\n" - "ldr x13, [%x[params_struct], %[offsetof_args_ld_input_col]]\n" - "mov x22, #0x0\n" + "str x21, [%x[params_struct], %[offsetof_args_tile_j]]\n" + "ldr x24, [%x[params_struct], %[offsetof_args_ld_input_row]]\n" + "ldr x23, [%x[params_struct], %[offsetof_args_ld_output_row]]\n" + "mul x20, x22, x24\n" // offset = tile_i * ld_input_row + "ldr x14, [%x[params_struct], %[offsetof_args_ld_input_col]]\n" + "ldr x13, [%x[params_struct], %[offsetof_args_ld_output_col]]\n" + "mul x19, x22, x23\n" // offset = tile_i * ld_output_row + "mov x22, #0x10\n" // cntb _, ALL, #1 + "madd x20, x21, x14, x20\n" // offset += tile_j * ld_input_col "ldr x12, [%x[params_struct], %[offsetof_args_inptr]]\n" - "mul x19, x17, x23\n" // offset = tile_i * ld_input_row - "ldr x20, [%x[params_struct], %[offsetof_args_ld_output_row]]\n" - "madd x19, x16, x13, x19\n" // offset += tile_j * ld_input_col - "ldr x11, [%x[params_struct], %[offsetof_args_ld_output_col]]\n" - "mul x19, x19, x25\n" // offset *= kernel_stride * output_size - "ldr x10, [%x[params_struct], %[offsetof_args_outptr]]\n" - "add x12, x12, x19, LSL #1\n" // inptr[0] += offset * sizeof(__fp16) - "ld1r { v18.8h }, [x24]\n" - "add x9, x12, x23, LSL #1\n" - "ld1r { v17.8h }, [x21]\n" - "add x28, x9, x23, LSL #1\n" + "lsl x14, x14, #0x1\n" + "ldr x11, [%x[params_struct], %[offsetof_args_outptr]]\n" + "madd x19, x21, x13, x19\n" // offset += tile_j * ld_output_col + "lsr x21, %x[n_channels], #0x3\n" + "add x10, x14, x14\n" + "ldr x9, [%x[params_struct], %[offsetof_args_params]]\n" + "mul x20, x20, x26\n" // offset *= kernel_stride * output_size + "add x12, x12, x20, LSL #1\n" // inptr[0] += offset * sizeof(__fp16) + "add x28, x12, x24, LSL #1\n" + "mul x19, x19, x25\n" // offset *= output_tile_size + "add x27, x28, x24, LSL #1\n" + "add x11, x11, x19, LSL #1\n" // outptrs[0] += offset * sizeof(__fp16) + "add x20, %x[params_struct], %[offsetof_args_min]\n" + "add x19, %x[params_struct], %[offsetof_args_max]\n" + "ld1r { v18.8h }, [x20]\n" + "ld1r { v17.8h }, [x19]\n" + "add x26, x27, x24, LSL #1\n" + "add x25, x10, x14\n" + "add x24, x11, x23, LSL #1\n" "lsl x13, x13, #0x1\n" - "add x27, x28, x23, LSL #1\n" - "add x26, x13, x13\n" - "add x25, x26, x13\n" - "mul x19, x17, x20\n" // offset = tile_i * ld_output_row - "madd x19, x16, x11, x19\n" // offset += tile_j * ld_output_col - "mul x19, x19, x15\n" // offset *= output_tile_size - "add x10, x10, x19, LSL #1\n" // outptrs[0] += offset * sizeof(__fp16) - "add x24, x10, x20, LSL #1\n" - "lsl x11, x11, #0x1\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 q16, [x14, #0x0]\n" - "ldr q0, [x14, #0x10]\n" - "cmp x21, x19, LSL #4\n" - "ldr q1, [x14, #0x20]\n" - "ldr q2, [x14, #0x30]\n" - "ldr q3, [x14, #0x40]\n" - "ldr q4, [x14, #0x50]\n" - "ldr q5, [x14, #0x60]\n" - "ldr q6, [x14, #0x70]\n" - "ldr q7, [x14, #0x80]\n" - "ldr q8, [x14, #0x90]\n" - "add x14, x14, #0xa0\n" - "ldr q9, [x9, x13]\n" + "mov x20, #0x0\n" + "sub x19, XZR, x22\n" + "cbz x21, 4f\n" + "ldr q16, [x9, #0x0]\n" + "cmp x22, x21, LSL #4\n" + "ldr q0, [x9, #0x10]\n" + "ldr q1, [x9, #0x20]\n" + "ldr q2, [x9, #0x30]\n" + "ldr q3, [x9, #0x40]\n" + "ldr q4, [x9, #0x50]\n" + "ldr q5, [x9, #0x60]\n" + "ldr q6, [x9, #0x70]\n" + "ldr q7, [x9, #0x80]\n" + "ldr q8, [x9, #0x90]\n" + "ldr q9, [x28, x14]\n" + "add x9, x9, #0xa0\n" "ld1 { v10.8h }, [x12]\n" "ldr q11, [x12, x25]\n" - "ldr q12, [x9, x26]\n" - "ldr q13, [x28, x13]\n" + "ldr q12, [x28, x10]\n" + "ldr q13, [x27, x14]\n" "bge 3f\n" "2:" // Tile loop: Channel loop - "mov v31.16b, v16.16b\n fmla v31.8h, v4.8h, v9.8h\n" - "add x20, x20, #0x10\n" - "mov v30.16b, v16.16b\n fmla v30.8h, v3.8h, v9.8h\n" + "mov v28.16b, v16.16b\n fmla v28.8h, v4.8h, v9.8h\n" + "mov v29.16b, v16.16b\n fmla v29.8h, v3.8h, v9.8h\n" "add x22, x22, #0x10\n" - "mov v29.16b, v16.16b\n fmla v29.8h, v1.8h, v9.8h\n" - "add x21, x21, #0x10\n" - "mov v28.16b, v16.16b\n fmla v28.8h, v0.8h, v9.8h\n" - "ld1 { v9.8h }, [x27]\n" - "cmp x21, x19, LSL #4\n" - "fmla v31.8h, v0.8h, v10.8h\n" - "ldr q10, [x28, x26]\n" - "fmla v30.8h, v2.8h, v11.8h\n" - "ldr q11, [x27, x25]\n" - "fmla v29.8h, v2.8h, v12.8h\n" - "ldr q16, [x14, #0x0]\n" - "fmla v28.8h, v1.8h, v12.8h\n" - "fmla v31.8h, v5.8h, v12.8h\n" - "fmla v30.8h, v4.8h, v12.8h\n" - "ldr q12, [x12, x13]\n" - "fmla v29.8h, v6.8h, v9.8h\n" - "ldr q9, [x12, x26]\n" - "add x12, x12, #0x10\n" - "fmla v28.8h, v3.8h, v13.8h\n" - "fmla v31.8h, v7.8h, v13.8h\n" - "fmla v30.8h, v6.8h, v13.8h\n" - "fmla v29.8h, v4.8h, v13.8h\n" - "fmla v28.8h, v8.8h, v11.8h\n" - "ld1 { v11.8h }, [x9]\n" + "cmp x22, x21, LSL #4\n" + "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n" + "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n" + "ld1 { v9.8h }, [x26]\n" + "add x19, x19, #0x10\n" + "fmla v28.8h, v0.8h, v10.8h\n" + "fmla v29.8h, v2.8h, v11.8h\n" + "ldr q11, [x26, x25]\n" + "ldr q10, [x27, x10]\n" + "fmla v30.8h, v2.8h, v12.8h\n" "fmla v31.8h, v1.8h, v12.8h\n" - "fmla v30.8h, v0.8h, v12.8h\n" - "ldr q12, [x9, x25]\n" - "add x9, x9, #0x10\n" - "fmla v29.8h, v5.8h, v10.8h\n" - "fmla v28.8h, v4.8h, v10.8h\n" - "ldr q4, [x14, #0x50]\n" - "fmla v31.8h, v2.8h, v9.8h\n" - "fmla v30.8h, v1.8h, v9.8h\n" - "ld1 { v9.8h }, [x28]\n" - "ldr q1, [x14, #0x20]\n" - "fmla v29.8h, v0.8h, v11.8h\n" - "ldr q0, [x14, #0x10]\n" - "fmla v28.8h, v2.8h, v12.8h\n" - "ldr q2, [x14, #0x30]\n" - "fmla v31.8h, v8.8h, v10.8h\n" - "fmla v30.8h, v7.8h, v10.8h\n" - "ldr q10, [x28, x25]\n" + "add x20, x20, #0x10\n" + "ldr q16, [x9, #0x0]\n" + "fmla v28.8h, v5.8h, v12.8h\n" + "fmla v29.8h, v4.8h, v12.8h\n" + "ldr q12, [x12, x14]\n" + "fmla v30.8h, v6.8h, v9.8h\n" + "fmla v31.8h, v3.8h, v13.8h\n" + "ldr q9, [x12, x10]\n" + "add x12, x12, #0x10\n" + "fmla v28.8h, v7.8h, v13.8h\n" + "fmla v29.8h, v6.8h, v13.8h\n" + "fmla v30.8h, v4.8h, v13.8h\n" + "fmla v31.8h, v8.8h, v11.8h\n" + "ld1 { v11.8h }, [x28]\n" + "fmla v28.8h, v1.8h, v12.8h\n" + "fmla v29.8h, v0.8h, v12.8h\n" + "ldr q12, [x28, x25]\n" "add x28, x28, #0x10\n" - "fmla v29.8h, v3.8h, v9.8h\n" - "ldr q13, [x28, x13]\n" - "fmla v31.8h, v3.8h, v11.8h\n" - "ldr q11, [x27, x13]\n" - "fmla v30.8h, v5.8h, v12.8h\n" - "ldr q12, [x27, x26]\n" + "fmla v30.8h, v5.8h, v10.8h\n" + "fmla v31.8h, v4.8h, v10.8h\n" + "ldr q4, [x9, #0x50]\n" + "fmla v28.8h, v2.8h, v9.8h\n" + "fmla v29.8h, v1.8h, v9.8h\n" + "ld1 { v9.8h }, [x27]\n" + "ldr q1, [x9, #0x20]\n" + "fmla v30.8h, v0.8h, v11.8h\n" + "fmla v31.8h, v2.8h, v12.8h\n" + "ldr q0, [x9, #0x10]\n" + "ldr q2, [x9, #0x30]\n" + "fmla v28.8h, v8.8h, v10.8h\n" + "fmla v29.8h, v7.8h, v10.8h\n" + "ldr q10, [x27, x25]\n" "add x27, x27, #0x10\n" - "fmla v28.8h, v5.8h, v10.8h\n" - "ldr q3, [x14, #0x40]\n" - "ldr q5, [x14, #0x60]\n" - "fmla v31.8h, v6.8h, v9.8h\n" - "ldr q9, [x9, x13]\n" - "fmla v30.8h, v8.8h, v10.8h\n" - "ld1 { v10.8h }, [x12]\n" - "fmla v29.8h, v7.8h, v11.8h\n" - "fmla v28.8h, v6.8h, v11.8h\n" + "fmla v30.8h, v3.8h, v9.8h\n" + "fmla v31.8h, v5.8h, v10.8h\n" + "ldr q13, [x27, x14]\n" + "fmla v28.8h, v3.8h, v11.8h\n" + "ldr q11, [x26, x14]\n" + "fmla v29.8h, v5.8h, v12.8h\n" + "ldr q12, [x26, x10]\n" + "fmla v30.8h, v7.8h, v11.8h\n" + "fmla v31.8h, v6.8h, v11.8h\n" + "add x26, x26, #0x10\n" "ldr q11, [x12, x25]\n" - "ldr q6, [x14, #0x70]\n" - "fmax v31.8h, v31.8h, v18.8h\n" - "fmax v30.8h, v30.8h, v18.8h\n" - "fmla v29.8h, v8.8h, v12.8h\n" - "ldr q8, [x14, #0x90]\n" - "fmla v28.8h, v7.8h, v12.8h\n" - "ldr q12, [x9, x26]\n" - "fmin v31.8h, v31.8h, v17.8h\n" - "ldr q7, [x14, #0x80]\n" - "add x14, x14, #0xa0\n" - "fmin v30.8h, v30.8h, v17.8h\n" - "st1 { v31.8h }, [x10]\n" - "fmax v29.8h, v29.8h, v18.8h\n" + "fmla v28.8h, v6.8h, v9.8h\n" + "fmla v29.8h, v8.8h, v10.8h\n" "fmax v28.8h, v28.8h, v18.8h\n" - "str q30, [x10, x11]\n" - "fmin v29.8h, v29.8h, v17.8h\n" - "st1 { v29.8h }, [x24]\n" + "ldr q9, [x28, x14]\n" + "fmla v30.8h, v8.8h, v12.8h\n" + "fmla v31.8h, v7.8h, v12.8h\n" + "fmax v29.8h, v29.8h, v18.8h\n" + "ld1 { v10.8h }, [x12]\n" + "fmax v30.8h, v30.8h, v18.8h\n" + "fmax v31.8h, v31.8h, v18.8h\n" + "ldr q12, [x28, x10]\n" + "ldr q3, [x9, #0x40]\n" "fmin v28.8h, v28.8h, v17.8h\n" - "add x10, x10, #0x10\n" - "str q28, [x24, x11]\n" + "fmin v29.8h, v29.8h, v17.8h\n" + "st1 { v28.8h }, [x11]\n" + "ldr q5, [x9, #0x60]\n" + "fmin v30.8h, v30.8h, v17.8h\n" + "fmin v31.8h, v31.8h, v17.8h\n" + "str q29, [x11, x13]\n" + "add x11, x11, #0x10\n" + "st1 { v30.8h }, [x24]\n" + "ldr q6, [x9, #0x70]\n" + "ldr q7, [x9, #0x80]\n" + "str q31, [x24, x13]\n" "add x24, x24, #0x10\n" + "ldr q8, [x9, #0x90]\n" + "add x9, x9, #0xa0\n" "blt 2b\n" "3:" // Tile loop: Channel tail - "mov v31.16b, v16.16b\n fmla v31.8h, v4.8h, v9.8h\n" - "mov v30.16b, v16.16b\n fmla v30.8h, v3.8h, v9.8h\n" - "mov v29.16b, v16.16b\n fmla v29.8h, v1.8h, v9.8h\n" - "mov v28.16b, v16.16b\n fmla v28.8h, v0.8h, v9.8h\n" - "ld1 { v9.8h }, [x27]\n" - "fmla v31.8h, v0.8h, v10.8h\n" - "ldr q10, [x28, x26]\n" - "fmla v30.8h, v2.8h, v11.8h\n" - "ldr q11, [x27, x25]\n" - "fmla v29.8h, v2.8h, v12.8h\n" - "fmla v28.8h, v1.8h, v12.8h\n" - "fmla v31.8h, v5.8h, v12.8h\n" - "fmla v30.8h, v4.8h, v12.8h\n" - "ldr q12, [x12, x13]\n" - "fmla v29.8h, v6.8h, v9.8h\n" - "ldr q9, [x12, x26]\n" - "add x12, x12, #0x10\n" - "fmla v28.8h, v3.8h, v13.8h\n" - "fmla v31.8h, v7.8h, v13.8h\n" - "fmla v30.8h, v6.8h, v13.8h\n" - "fmla v29.8h, v4.8h, v13.8h\n" - "fmla v28.8h, v8.8h, v11.8h\n" - "ld1 { v11.8h }, [x9]\n" + "mov v28.16b, v16.16b\n fmla v28.8h, v4.8h, v9.8h\n" + "mov v29.16b, v16.16b\n fmla v29.8h, v3.8h, v9.8h\n" + "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n" + "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n" + "ld1 { v9.8h }, [x26]\n" + "fmla v28.8h, v0.8h, v10.8h\n" + "fmla v29.8h, v2.8h, v11.8h\n" + "ldr q11, [x26, x25]\n" + "ldr q10, [x27, x10]\n" + "fmla v30.8h, v2.8h, v12.8h\n" "fmla v31.8h, v1.8h, v12.8h\n" - "fmla v30.8h, v0.8h, v12.8h\n" - "ldr q12, [x9, x25]\n" - "add x9, x9, #0x10\n" - "fmla v29.8h, v5.8h, v10.8h\n" - "fmla v28.8h, v4.8h, v10.8h\n" - "fmla v31.8h, v2.8h, v9.8h\n" - "fmla v30.8h, v1.8h, v9.8h\n" - "ld1 { v9.8h }, [x28]\n" - "fmla v29.8h, v0.8h, v11.8h\n" - "fmla v28.8h, v2.8h, v12.8h\n" - "fmla v31.8h, v8.8h, v10.8h\n" - "fmla v30.8h, v7.8h, v10.8h\n" - "ldr q10, [x28, x25]\n" + "fmla v28.8h, v5.8h, v12.8h\n" + "fmla v29.8h, v4.8h, v12.8h\n" + "ldr q12, [x12, x14]\n" + "fmla v30.8h, v6.8h, v9.8h\n" + "fmla v31.8h, v3.8h, v13.8h\n" + "ldr q9, [x12, x10]\n" + "add x12, x12, #0x10\n" + "fmla v28.8h, v7.8h, v13.8h\n" + "fmla v29.8h, v6.8h, v13.8h\n" + "fmla v30.8h, v4.8h, v13.8h\n" + "fmla v31.8h, v8.8h, v11.8h\n" + "ld1 { v11.8h }, [x28]\n" + "fmla v28.8h, v1.8h, v12.8h\n" + "fmla v29.8h, v0.8h, v12.8h\n" + "ldr q12, [x28, x25]\n" "add x28, x28, #0x10\n" - "fmla v29.8h, v3.8h, v9.8h\n" - "fmla v31.8h, v3.8h, v11.8h\n" - "ldr q11, [x27, x13]\n" - "fmla v30.8h, v5.8h, v12.8h\n" - "ldr q12, [x27, x26]\n" + "fmla v30.8h, v5.8h, v10.8h\n" + "fmla v31.8h, v4.8h, v10.8h\n" + "fmla v28.8h, v2.8h, v9.8h\n" + "fmla v29.8h, v1.8h, v9.8h\n" + "ld1 { v9.8h }, [x27]\n" + "fmla v30.8h, v0.8h, v11.8h\n" + "fmla v31.8h, v2.8h, v12.8h\n" + "fmla v28.8h, v8.8h, v10.8h\n" + "fmla v29.8h, v7.8h, v10.8h\n" + "ldr q10, [x27, x25]\n" "add x27, x27, #0x10\n" - "fmla v28.8h, v5.8h, v10.8h\n" - "fmla v31.8h, v6.8h, v9.8h\n" - "fmla v30.8h, v8.8h, v10.8h\n" - "fmla v29.8h, v7.8h, v11.8h\n" - "fmla v28.8h, v6.8h, v11.8h\n" - "fmax v31.8h, v31.8h, v18.8h\n" - "fmax v30.8h, v30.8h, v18.8h\n" - "fmla v29.8h, v8.8h, v12.8h\n" - "fmla v28.8h, v7.8h, v12.8h\n" - "fmin v31.8h, v31.8h, v17.8h\n" - "st1 { v31.8h }, [x10]\n" - "fmin v30.8h, v30.8h, v17.8h\n" - "fmax v29.8h, v29.8h, v18.8h\n" - "str q30, [x10, x11]\n" - "fmin v29.8h, v29.8h, v17.8h\n" - "add x10, x10, #0x10\n" + "fmla v30.8h, v3.8h, v9.8h\n" + "fmla v31.8h, v5.8h, v10.8h\n" + "fmla v28.8h, v3.8h, v11.8h\n" + "ldr q11, [x26, x14]\n" + "fmla v29.8h, v5.8h, v12.8h\n" + "ldr q12, [x26, x10]\n" + "fmla v30.8h, v7.8h, v11.8h\n" + "fmla v31.8h, v6.8h, v11.8h\n" + "add x26, x26, #0x10\n" + "fmla v28.8h, v6.8h, v9.8h\n" + "fmla v29.8h, v8.8h, v10.8h\n" "fmax v28.8h, v28.8h, v18.8h\n" - "st1 { v29.8h }, [x24]\n" + "fmla v30.8h, v8.8h, v12.8h\n" + "fmla v31.8h, v7.8h, v12.8h\n" + "fmax v29.8h, v29.8h, v18.8h\n" + "fmax v30.8h, v30.8h, v18.8h\n" + "fmax v31.8h, v31.8h, v18.8h\n" "fmin v28.8h, v28.8h, v17.8h\n" - "str q28, [x24, x11]\n" + "fmin v29.8h, v29.8h, v17.8h\n" + "st1 { v28.8h }, [x11]\n" + "fmin v30.8h, v30.8h, v17.8h\n" + "fmin v31.8h, v31.8h, v17.8h\n" + "str q29, [x11, x13]\n" + "add x11, x11, #0x10\n" + "st1 { v30.8h }, [x24]\n" + "str q31, [x24, x13]\n" "add x24, x24, #0x10\n" "4:" // Tile loop: Oddments - "tst %x[n_channels], #0x1\n" - "beq 31f\n" - "ldr q16, [x14, #0x0]\n" - "ldr q0, [x14, #0x10]\n" - "add x23, x9, x13\n" - "ldr q1, [x14, #0x20]\n" + "tst %x[n_channels], #0x7\n" + "beq 57f\n" + "ldr q16, [x9, #0x0]\n" + "ldr q0, [x9, #0x10]\n" + "ldr q1, [x9, #0x20]\n" + "ldr q2, [x9, #0x30]\n" + "add x23, x28, x14\n" "add x22, x12, XZR\n" - "ldr q2, [x14, #0x30]\n" + "ldr q3, [x9, #0x40]\n" + "ldr q4, [x9, #0x50]\n" "add x21, x12, x25\n" - "ldr q3, [x14, #0x40]\n" - "add x20, x9, x26\n" - "ldr q4, [x14, #0x50]\n" - "add x19, x28, x13\n" - "ldr q5, [x14, #0x60]\n" - "ldr q6, [x14, #0x70]\n" - "ldr q7, [x14, #0x80]\n" - "ldr q8, [x14, #0x90]\n" + "add x20, x28, x10\n" + "ldr q5, [x9, #0x60]\n" + "ldr q6, [x9, #0x70]\n" + "add x19, x27, x14\n" + "ldr q7, [x9, #0x80]\n" + "ldr q8, [x9, #0x90]\n" + "tbz %x[n_channels], #2, 6f\n" + "ldr d9, [x23], #0x8\n" + "ldr d10, [x22], #0x8\n" + "ldr d11, [x21], #0x8\n" + "ldr d12, [x20], #0x8\n" + "ldr d13, [x19], #0x8\n" "tbz %x[n_channels], #1, 5f\n" + "ld1 { v9.s }[2], [x23], #0x4\n" + "ld1 { v10.s }[2], [x22], #0x4\n" + "ld1 { v11.s }[2], [x21], #0x4\n" + "ld1 { v12.s }[2], [x20], #0x4\n" + "ld1 { v13.s }[2], [x19], #0x4\n" + "tbz %x[n_channels], #0, 8f\n" + "ld1 { v9.h }[6], [x23]\n" + "ld1 { v10.h }[6], [x22]\n" + "ld1 { v11.h }[6], [x21]\n" + "ld1 { v12.h }[6], [x20]\n" + "ld1 { v13.h }[6], [x19]\n" + "b 8f\n" + "5:" // Tile loop: Oddments: Load inputs: (1, 1), (0, 0), (0, 3), (1, 2), (2, 1): Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 8f\n" + "ld1 { v9.h }[4], [x23]\n" + "ld1 { v10.h }[4], [x22]\n" + "ld1 { v11.h }[4], [x21]\n" + "ld1 { v12.h }[4], [x20]\n" + "ld1 { v13.h }[4], [x19]\n" + "b 8f\n" + "6:" // Tile loop: Oddments: Load inputs: (1, 1), (0, 0), (0, 3), (1, 2), (2, 1): Bit 2: Unset + "tbz %x[n_channels], #1, 7f\n" "ldr s9, [x23], #0x4\n" "ldr s10, [x22], #0x4\n" "ldr s11, [x21], #0x4\n" "ldr s12, [x20], #0x4\n" "ldr s13, [x19], #0x4\n" - "tbz %x[n_channels], #0, 6f\n" + "tbz %x[n_channels], #0, 8f\n" "ld1 { v9.h }[2], [x23]\n" "ld1 { v10.h }[2], [x22]\n" "ld1 { v11.h }[2], [x21]\n" "ld1 { v12.h }[2], [x20]\n" "ld1 { v13.h }[2], [x19]\n" - "b 6f\n" - "5:" // Tile loop: Oddments: Load inputs: (1, 1), (0, 0), (0, 3), (1, 2), (2, 1): Bit 1: Unset + "b 8f\n" + "7:" // Tile loop: Oddments: Load inputs: (1, 1), (0, 0), (0, 3), (1, 2), (2, 1): Bit 2: Unset: Bit 1: Unset "ldr h9, [x23, #0x0]\n" "ldr h10, [x22, #0x0]\n" "ldr h11, [x21, #0x0]\n" "ldr h12, [x20, #0x0]\n" "ldr h13, [x19, #0x0]\n" - "6:" // Tile loop: Oddments: Load inputs: (1, 1), (0, 0), (0, 3), (1, 2), (2, 1): Bit 1: End - "mov v31.16b, v16.16b\n fmla v31.8h, v4.8h, v9.8h\n" - "add x19, x27, XZR\n" - "mov v30.16b, v16.16b\n fmla v30.8h, v3.8h, v9.8h\n" - "mov v29.16b, v16.16b\n fmla v29.8h, v1.8h, v9.8h\n" - "mov v28.16b, v16.16b\n fmla v28.8h, v0.8h, v9.8h\n" - "fmla v31.8h, v0.8h, v10.8h\n" - "fmla v30.8h, v2.8h, v11.8h\n" - "fmla v29.8h, v2.8h, v12.8h\n" - "fmla v28.8h, v1.8h, v12.8h\n" - "fmla v31.8h, v5.8h, v12.8h\n" - "fmla v30.8h, v4.8h, v12.8h\n" - "tbz %x[n_channels], #1, 7f\n" + "8:" // Tile loop: Oddments: Load inputs: (1, 1), (0, 0), (0, 3), (1, 2), (2, 1): Bit 2: End + "mov v28.16b, v16.16b\n fmla v28.8h, v4.8h, v9.8h\n" + "mov v29.16b, v16.16b\n fmla v29.8h, v3.8h, v9.8h\n" + "add x19, x26, XZR\n" + "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n" + "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n" + "fmla v28.8h, v0.8h, v10.8h\n" + "fmla v29.8h, v2.8h, v11.8h\n" + "fmla v28.8h, v5.8h, v12.8h\n" + "fmla v29.8h, v4.8h, v12.8h\n" + "fmla v30.8h, v2.8h, v12.8h\n" + "fmla v31.8h, v1.8h, v12.8h\n" + "tbz %x[n_channels], #2, 10f\n" + "ldr d9, [x19], #0x8\n" + "tbz %x[n_channels], #1, 9f\n" + "ld1 { v9.s }[2], [x19], #0x4\n" + "tbz %x[n_channels], #0, 12f\n" + "ld1 { v9.h }[6], [x19]\n" + "b 12f\n" + "9:" // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 12f\n" + "ld1 { v9.h }[4], [x19]\n" + "b 12f\n" + "10:" // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: Unset + "tbz %x[n_channels], #1, 11f\n" "ldr s9, [x19], #0x4\n" - "tbz %x[n_channels], #0, 8f\n" + "tbz %x[n_channels], #0, 12f\n" "ld1 { v9.h }[2], [x19]\n" - "b 8f\n" - "7:" // Tile loop: Oddments: Load inputs: (3, 0): Bit 1: Unset + "b 12f\n" + "11:" // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: Unset: Bit 1: Unset "ldr h9, [x19, #0x0]\n" - "8:" // Tile loop: Oddments: Load inputs: (3, 0): Bit 1: End - "fmla v29.8h, v6.8h, v9.8h\n" - "add x19, x27, x25\n" - "fmla v31.8h, v7.8h, v13.8h\n" - "fmla v30.8h, v6.8h, v13.8h\n" - "fmla v28.8h, v3.8h, v13.8h\n" - "fmla v29.8h, v4.8h, v13.8h\n" - "tbz %x[n_channels], #1, 9f\n" + "12:" // Tile loop: Oddments: Load inputs: (3, 0): Bit 2: End + "fmla v30.8h, v6.8h, v9.8h\n" + "fmla v28.8h, v7.8h, v13.8h\n" + "add x19, x26, x25\n" + "fmla v29.8h, v6.8h, v13.8h\n" + "fmla v30.8h, v4.8h, v13.8h\n" + "fmla v31.8h, v3.8h, v13.8h\n" + "tbz %x[n_channels], #2, 14f\n" + "ldr d11, [x19], #0x8\n" + "tbz %x[n_channels], #1, 13f\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: (3, 3): 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: (3, 3): Bit 2: Unset + "tbz %x[n_channels], #1, 15f\n" "ldr s11, [x19], #0x4\n" - "tbz %x[n_channels], #0, 10f\n" + "tbz %x[n_channels], #0, 16f\n" "ld1 { v11.h }[2], [x19]\n" - "b 10f\n" - "9:" // Tile loop: Oddments: Load inputs: (3, 3): Bit 1: Unset + "b 16f\n" + "15:" // Tile loop: Oddments: Load inputs: (3, 3): Bit 2: Unset: Bit 1: Unset "ldr h11, [x19, #0x0]\n" - "10:" // Tile loop: Oddments: Load inputs: (3, 3): Bit 1: End - "fmla v28.8h, v8.8h, v11.8h\n" - "add x19, x12, x13\n" - "tbz %x[n_channels], #1, 11f\n" + "16:" // Tile loop: Oddments: Load inputs: (3, 3): Bit 2: End + "fmla v31.8h, v8.8h, v11.8h\n" + "add x19, x12, x14\n" + "tbz %x[n_channels], #2, 18f\n" + "ldr d12, [x19], #0x8\n" + "tbz %x[n_channels], #1, 17f\n" + "ld1 { v12.s }[2], [x19], #0x4\n" + "tbz %x[n_channels], #0, 20f\n" + "ld1 { v12.h }[6], [x19]\n" + "b 20f\n" + "17:" // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 20f\n" + "ld1 { v12.h }[4], [x19]\n" + "b 20f\n" + "18:" // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: Unset + "tbz %x[n_channels], #1, 19f\n" "ldr s12, [x19], #0x4\n" - "tbz %x[n_channels], #0, 12f\n" + "tbz %x[n_channels], #0, 20f\n" "ld1 { v12.h }[2], [x19]\n" - "b 12f\n" - "11:" // Tile loop: Oddments: Load inputs: (0, 1): Bit 1: Unset + "b 20f\n" + "19:" // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: Unset: Bit 1: Unset "ldr h12, [x19, #0x0]\n" - "12:" // Tile loop: Oddments: Load inputs: (0, 1): Bit 1: End - "fmla v31.8h, v1.8h, v12.8h\n" - "add x19, x12, x26\n" - "fmla v30.8h, v0.8h, v12.8h\n" - "tbz %x[n_channels], #1, 13f\n" + "20:" // Tile loop: Oddments: Load inputs: (0, 1): Bit 2: End + "fmla v28.8h, v1.8h, v12.8h\n" + "fmla v29.8h, v0.8h, v12.8h\n" + "add x19, x12, x10\n" + "tbz %x[n_channels], #2, 22f\n" + "ldr d9, [x19], #0x8\n" + "tbz %x[n_channels], #1, 21f\n" + "ld1 { v9.s }[2], [x19], #0x4\n" + "tbz %x[n_channels], #0, 24f\n" + "ld1 { v9.h }[6], [x19]\n" + "b 24f\n" + "21:" // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 24f\n" + "ld1 { v9.h }[4], [x19]\n" + "b 24f\n" + "22:" // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: Unset + "tbz %x[n_channels], #1, 23f\n" "ldr s9, [x19], #0x4\n" - "tbz %x[n_channels], #0, 14f\n" + "tbz %x[n_channels], #0, 24f\n" "ld1 { v9.h }[2], [x19]\n" - "b 14f\n" - "13:" // Tile loop: Oddments: Load inputs: (0, 2): Bit 1: Unset + "b 24f\n" + "23:" // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: Unset: Bit 1: Unset "ldr h9, [x19, #0x0]\n" - "14:" // Tile loop: Oddments: Load inputs: (0, 2): Bit 1: End - "fmla v31.8h, v2.8h, v9.8h\n" - "add x19, x28, x26\n" - "fmla v30.8h, v1.8h, v9.8h\n" - "tbz %x[n_channels], #1, 15f\n" + "24:" // Tile loop: Oddments: Load inputs: (0, 2): Bit 2: End + "fmla v28.8h, v2.8h, v9.8h\n" + "fmla v29.8h, v1.8h, v9.8h\n" + "add x19, x27, x10\n" + "tbz %x[n_channels], #2, 26f\n" + "ldr d10, [x19], #0x8\n" + "tbz %x[n_channels], #1, 25f\n" + "ld1 { v10.s }[2], [x19], #0x4\n" + "tbz %x[n_channels], #0, 28f\n" + "ld1 { v10.h }[6], [x19]\n" + "b 28f\n" + "25:" // Tile loop: Oddments: Load inputs: (2, 2): Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 28f\n" + "ld1 { v10.h }[4], [x19]\n" + "b 28f\n" + "26:" // Tile loop: Oddments: Load inputs: (2, 2): Bit 2: Unset + "tbz %x[n_channels], #1, 27f\n" "ldr s10, [x19], #0x4\n" - "tbz %x[n_channels], #0, 16f\n" + "tbz %x[n_channels], #0, 28f\n" "ld1 { v10.h }[2], [x19]\n" - "b 16f\n" - "15:" // Tile loop: Oddments: Load inputs: (2, 2): Bit 1: Unset + "b 28f\n" + "27:" // Tile loop: Oddments: Load inputs: (2, 2): Bit 2: Unset: Bit 1: Unset "ldr h10, [x19, #0x0]\n" - "16:" // Tile loop: Oddments: Load inputs: (2, 2): Bit 1: End - "fmla v31.8h, v8.8h, v10.8h\n" - "add x19, x9, XZR\n" - "fmla v30.8h, v7.8h, v10.8h\n" - "fmla v29.8h, v5.8h, v10.8h\n" - "fmla v28.8h, v4.8h, v10.8h\n" - "tbz %x[n_channels], #1, 17f\n" + "28:" // Tile loop: Oddments: Load inputs: (2, 2): Bit 2: End + "fmla v28.8h, v8.8h, v10.8h\n" + "fmla v29.8h, v7.8h, v10.8h\n" + "add x19, x28, XZR\n" + "fmla v30.8h, v5.8h, v10.8h\n" + "fmla v31.8h, v4.8h, v10.8h\n" + "tbz %x[n_channels], #2, 30f\n" + "ldr d11, [x19], #0x8\n" + "tbz %x[n_channels], #1, 29f\n" + "ld1 { v11.s }[2], [x19], #0x4\n" + "tbz %x[n_channels], #0, 32f\n" + "ld1 { v11.h }[6], [x19]\n" + "b 32f\n" + "29:" // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 32f\n" + "ld1 { v11.h }[4], [x19]\n" + "b 32f\n" + "30:" // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: Unset + "tbz %x[n_channels], #1, 31f\n" "ldr s11, [x19], #0x4\n" - "tbz %x[n_channels], #0, 18f\n" + "tbz %x[n_channels], #0, 32f\n" "ld1 { v11.h }[2], [x19]\n" - "b 18f\n" - "17:" // Tile loop: Oddments: Load inputs: (1, 0): Bit 1: Unset + "b 32f\n" + "31:" // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: Unset: Bit 1: Unset "ldr h11, [x19, #0x0]\n" - "18:" // Tile loop: Oddments: Load inputs: (1, 0): Bit 1: End - "fmla v31.8h, v3.8h, v11.8h\n" - "add x19, x9, x25\n" - "fmla v29.8h, v0.8h, v11.8h\n" - "tbz %x[n_channels], #1, 19f\n" + "32:" // Tile loop: Oddments: Load inputs: (1, 0): Bit 2: End + "fmla v28.8h, v3.8h, v11.8h\n" + "fmla v30.8h, v0.8h, v11.8h\n" + "add x19, x28, x25\n" + "tbz %x[n_channels], #2, 34f\n" + "ldr d12, [x19], #0x8\n" + "tbz %x[n_channels], #1, 33f\n" + "ld1 { v12.s }[2], [x19], #0x4\n" + "tbz %x[n_channels], #0, 36f\n" + "ld1 { v12.h }[6], [x19]\n" + "b 36f\n" + "33:" // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 36f\n" + "ld1 { v12.h }[4], [x19]\n" + "b 36f\n" + "34:" // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: Unset + "tbz %x[n_channels], #1, 35f\n" "ldr s12, [x19], #0x4\n" - "tbz %x[n_channels], #0, 20f\n" + "tbz %x[n_channels], #0, 36f\n" "ld1 { v12.h }[2], [x19]\n" - "b 20f\n" - "19:" // Tile loop: Oddments: Load inputs: (1, 3): Bit 1: Unset + "b 36f\n" + "35:" // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: Unset: Bit 1: Unset "ldr h12, [x19, #0x0]\n" - "20:" // Tile loop: Oddments: Load inputs: (1, 3): Bit 1: End - "fmla v30.8h, v5.8h, v12.8h\n" - "add x19, x28, XZR\n" - "fmla v28.8h, v2.8h, v12.8h\n" - "tbz %x[n_channels], #1, 21f\n" + "36:" // Tile loop: Oddments: Load inputs: (1, 3): Bit 2: End + "fmla v29.8h, v5.8h, v12.8h\n" + "fmla v31.8h, v2.8h, v12.8h\n" + "add x19, x27, XZR\n" + "tbz %x[n_channels], #2, 38f\n" + "ldr d9, [x19], #0x8\n" + "tbz %x[n_channels], #1, 37f\n" + "ld1 { v9.s }[2], [x19], #0x4\n" + "tbz %x[n_channels], #0, 40f\n" + "ld1 { v9.h }[6], [x19]\n" + "b 40f\n" + "37:" // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 40f\n" + "ld1 { v9.h }[4], [x19]\n" + "b 40f\n" + "38:" // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: Unset + "tbz %x[n_channels], #1, 39f\n" "ldr s9, [x19], #0x4\n" - "tbz %x[n_channels], #0, 22f\n" + "tbz %x[n_channels], #0, 40f\n" "ld1 { v9.h }[2], [x19]\n" - "b 22f\n" - "21:" // Tile loop: Oddments: Load inputs: (2, 0): Bit 1: Unset + "b 40f\n" + "39:" // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: Unset: Bit 1: Unset "ldr h9, [x19, #0x0]\n" - "22:" // Tile loop: Oddments: Load inputs: (2, 0): Bit 1: End - "fmla v31.8h, v6.8h, v9.8h\n" - "add x19, x28, x25\n" - "fmla v29.8h, v3.8h, v9.8h\n" - "tbz %x[n_channels], #1, 23f\n" + "40:" // Tile loop: Oddments: Load inputs: (2, 0): Bit 2: End + "fmla v28.8h, v6.8h, v9.8h\n" + "fmla v30.8h, v3.8h, v9.8h\n" + "add x19, x27, x25\n" + "tbz %x[n_channels], #2, 42f\n" + "ldr d10, [x19], #0x8\n" + "tbz %x[n_channels], #1, 41f\n" + "ld1 { v10.s }[2], [x19], #0x4\n" + "tbz %x[n_channels], #0, 44f\n" + "ld1 { v10.h }[6], [x19]\n" + "b 44f\n" + "41:" // Tile loop: Oddments: Load inputs: (2, 3): Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 44f\n" + "ld1 { v10.h }[4], [x19]\n" + "b 44f\n" + "42:" // Tile loop: Oddments: Load inputs: (2, 3): Bit 2: Unset + "tbz %x[n_channels], #1, 43f\n" "ldr s10, [x19], #0x4\n" - "tbz %x[n_channels], #0, 24f\n" + "tbz %x[n_channels], #0, 44f\n" "ld1 { v10.h }[2], [x19]\n" - "b 24f\n" - "23:" // Tile loop: Oddments: Load inputs: (2, 3): Bit 1: Unset + "b 44f\n" + "43:" // Tile loop: Oddments: Load inputs: (2, 3): Bit 2: Unset: Bit 1: Unset "ldr h10, [x19, #0x0]\n" - "24:" // Tile loop: Oddments: Load inputs: (2, 3): Bit 1: End - "fmla v30.8h, v8.8h, v10.8h\n" - "add x19, x27, x13\n" - "fmla v28.8h, v5.8h, v10.8h\n" - "tbz %x[n_channels], #1, 25f\n" + "44:" // Tile loop: Oddments: Load inputs: (2, 3): Bit 2: End + "fmla v29.8h, v8.8h, v10.8h\n" + "fmla v31.8h, v5.8h, v10.8h\n" + "add x19, x26, x14\n" + "tbz %x[n_channels], #2, 46f\n" + "ldr d11, [x19], #0x8\n" + "tbz %x[n_channels], #1, 45f\n" + "ld1 { v11.s }[2], [x19], #0x4\n" + "tbz %x[n_channels], #0, 48f\n" + "ld1 { v11.h }[6], [x19]\n" + "b 48f\n" + "45:" // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 48f\n" + "ld1 { v11.h }[4], [x19]\n" + "b 48f\n" + "46:" // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: Unset + "tbz %x[n_channels], #1, 47f\n" "ldr s11, [x19], #0x4\n" - "tbz %x[n_channels], #0, 26f\n" + "tbz %x[n_channels], #0, 48f\n" "ld1 { v11.h }[2], [x19]\n" - "b 26f\n" - "25:" // Tile loop: Oddments: Load inputs: (3, 1): Bit 1: Unset + "b 48f\n" + "47:" // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: Unset: Bit 1: Unset "ldr h11, [x19, #0x0]\n" - "26:" // Tile loop: Oddments: Load inputs: (3, 1): Bit 1: End - "fmla v29.8h, v7.8h, v11.8h\n" - "add x19, x27, x26\n" - "fmla v28.8h, v6.8h, v11.8h\n" - "tbz %x[n_channels], #1, 27f\n" + "48:" // Tile loop: Oddments: Load inputs: (3, 1): Bit 2: End + "fmla v30.8h, v7.8h, v11.8h\n" + "fmla v31.8h, v6.8h, v11.8h\n" + "add x19, x26, x10\n" + "tbz %x[n_channels], #2, 50f\n" + "ldr d12, [x19], #0x8\n" + "tbz %x[n_channels], #1, 49f\n" + "ld1 { v12.s }[2], [x19], #0x4\n" + "tbz %x[n_channels], #0, 52f\n" + "ld1 { v12.h }[6], [x19]\n" + "b 52f\n" + "49:" // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 52f\n" + "ld1 { v12.h }[4], [x19]\n" + "b 52f\n" + "50:" // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: Unset + "tbz %x[n_channels], #1, 51f\n" "ldr s12, [x19], #0x4\n" - "tbz %x[n_channels], #0, 28f\n" + "tbz %x[n_channels], #0, 52f\n" "ld1 { v12.h }[2], [x19]\n" - "b 28f\n" - "27:" // Tile loop: Oddments: Load inputs: (3, 2): Bit 1: Unset + "b 52f\n" + "51:" // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: Unset: Bit 1: Unset "ldr h12, [x19, #0x0]\n" - "28:" // Tile loop: Oddments: Load inputs: (3, 2): Bit 1: End - "fmla v29.8h, v8.8h, v12.8h\n" - "fmla v28.8h, v7.8h, v12.8h\n" - "fmax v31.8h, v31.8h, v18.8h\n" - "fmax v30.8h, v30.8h, v18.8h\n" - "fmin v31.8h, v31.8h, v17.8h\n" - "fmax v29.8h, v29.8h, v18.8h\n" - "fmin v30.8h, v30.8h, v17.8h\n" + "52:" // Tile loop: Oddments: Load inputs: (3, 2): Bit 2: End + "fmla v30.8h, v8.8h, v12.8h\n" + "fmla v31.8h, v7.8h, v12.8h\n" "fmax v28.8h, v28.8h, v18.8h\n" - "fmin v29.8h, v29.8h, v17.8h\n" + "fmax v29.8h, v29.8h, v18.8h\n" + "fmax v30.8h, v30.8h, v18.8h\n" + "fmax v31.8h, v31.8h, v18.8h\n" "fmin v28.8h, v28.8h, v17.8h\n" - "tbz %x[n_channels], #1, 29f\n" - "mov x19, x10\n" - "st1 { v31.s }[0], [x19], x11\n" - "add x10, x10, #0x4\n" - "st1 { v30.s }[0], [x19]\n" + "fmin v29.8h, v29.8h, v17.8h\n" + "fmin v30.8h, v30.8h, v17.8h\n" + "fmin v31.8h, v31.8h, v17.8h\n" + "tbz %x[n_channels], #2, 54f\n" + "mov x20, x11\n" + "mov x19, x24\n" + "st1 { v28.d }[0], [x20], x13\n" + "add x11, x11, #0x8\n" + "add x24, x24, #0x8\n" + "st1 { v30.d }[0], [x19], x13\n" + "st1 { v29.d }[0], [x20]\n" + "st1 { v31.d }[0], [x19]\n" + "tbz %x[n_channels], #1, 53f\n" + "mov x20, x11\n" + "mov x19, x24\n" + "st1 { v28.s }[2], [x20], x13\n" + "add x11, x11, #0x4\n" + "add x24, x24, #0x4\n" + "st1 { v30.s }[2], [x19], x13\n" + "st1 { v29.s }[2], [x20]\n" + "st1 { v31.s }[2], [x19]\n" + "tbz %x[n_channels], #0, 56f\n" + "mov x20, x11\n" + "mov x19, x24\n" + "st1 { v28.h }[6], [x20], x13\n" + "st1 { v30.h }[6], [x19], x13\n" + "st1 { v29.h }[6], [x20]\n" + "st1 { v31.h }[6], [x19]\n" + "b 56f\n" + "53:" // Tile loop: Oddments: Store: Bit 2: Bit 1: Unset + "tbz %x[n_channels], #0, 56f\n" + "mov x20, x11\n" + "mov x19, x24\n" + "st1 { v28.h }[4], [x20], x13\n" + "st1 { v30.h }[4], [x19], x13\n" + "st1 { v29.h }[4], [x20]\n" + "st1 { v31.h }[4], [x19]\n" + "b 56f\n" + "54:" // Tile loop: Oddments: Store: Bit 2: Unset + "tbz %x[n_channels], #1, 55f\n" + "mov x20, x11\n" "mov x19, x24\n" - "st1 { v29.s }[0], [x19], x11\n" + "st1 { v28.s }[0], [x20], x13\n" + "st1 { v30.s }[0], [x19], x13\n" + "add x11, x11, #0x4\n" "add x24, x24, #0x4\n" - "st1 { v28.s }[0], [x19]\n" - "tbz %x[n_channels], #0, 30f\n" - "mov x20, x10\n" - "st1 { v31.h }[2], [x20], x11\n" + "st1 { v29.s }[0], [x20]\n" + "st1 { v31.s }[0], [x19]\n" + "tbz %x[n_channels], #0, 56f\n" + "mov x20, x11\n" "mov x19, x24\n" - "st1 { v30.h }[2], [x20]\n" - "st1 { v29.h }[2], [x19], x11\n" - "st1 { v28.h }[2], [x19]\n" - "b 30f\n" - "29:" // Tile loop: Oddments: Store: Bit 1: Unset - "mov x20, x10\n" - "st1 { v31.h }[0], [x20], x11\n" + "st1 { v28.h }[2], [x20], x13\n" + "st1 { v30.h }[2], [x19], x13\n" + "st1 { v29.h }[2], [x20]\n" + "st1 { v31.h }[2], [x19]\n" + "b 56f\n" + "55:" // Tile loop: Oddments: Store: Bit 2: Unset: Bit 1: Unset + "mov x20, x11\n" "mov x19, x24\n" - "st1 { v30.h }[0], [x20]\n" - "st1 { v29.h }[0], [x19], x11\n" - "st1 { v28.h }[0], [x19]\n" - "30:" // Tile loop: Oddments: Store: Bit 1: End + "st1 { v28.h }[0], [x20], x13\n" + "st1 { v30.h }[0], [x19], x13\n" + "st1 { v29.h }[0], [x20]\n" + "st1 { v31.h }[0], [x19]\n" + "56:" // Tile loop: Oddments: Store: Bit 2: End - "31:" // Tile loop: End - "ldr x17, [%x[params_struct], %[offsetof_args_tile_i]]\n" - "add x21, x17, #0x1\n" - "ldr x16, [%x[params_struct], %[offsetof_args_tile_j]]\n" - "ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" - "add x16, x16, #0x1\n" + "57:" // Tile loop: End + "ldr x21, [%x[params_struct], %[offsetof_args_tile_j]]\n" + "ldr x22, [%x[params_struct], %[offsetof_args_tile_i]]\n" + "add x21, x21, #0x1\n" + "add x20, x22, #0x1\n" "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n" - "cmp x16, x19\n" - "csel x16, x16, XZR, LT\n" - "csel x17, x17, x21, LT\n" - "cmp x17, x20\n" + "cmp x21, x19\n" + "ldr x19, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n" + "csel x22, x22, x20, LT\n" + "csel x21, x21, XZR, LT\n" + "cmp x22, x19\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" (¶ms_struct) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v28", "v29", "v30", "v31", "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", "v16", "v17", "v18", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } -- cgit v1.2.1