diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst')
2 files changed, 822 insertions, 823 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp index 620319bc7c..9bfcd9cd3c 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_direct.cpp @@ -25,7 +25,7 @@ #include <cstddef> #include <cstdint> -#if __aarch64__ +#if defined(__aarch64__) namespace arm_conv { namespace depthwise { @@ -110,7 +110,7 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "lsr x23, %x[n_channels], #0x2\n" "ldr x14, [%x[params_struct], %[offsetof_args_params]]\n" "mul x21, x21, x27\n" // offset *= kernel_stride * output_size - "add x16, x16, x21, LSL #2\n" // inptr[0] += offset * sizeof(float) + "add x16, x16, x21, LSL #2\n" // inptr[0] += offset * sizeof(float) "add x13, x16, x25, LSL #2\n" "mul x20, x20, x26\n" // offset *= output_tile_size "add x12, x13, x25, LSL #2\n" @@ -120,9 +120,9 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "add x9, x11, x8\n" "add x28, x15, x22, LSL #2\n" "add x20, %x[params_struct], %[offsetof_args_min]\n" - "ld1r { v18.4s }, [x20]\n" + "ld1r { v15.4s }, [x20]\n" "add x20, %x[params_struct], %[offsetof_args_max]\n" - "ld1r { v17.4s }, [x20]\n" + "ld1r { v14.4s }, [x20]\n" "add x27, x10, x25, LSL #2\n" "add x26, x9, x8\n" "add x25, x28, x22, LSL #2\n" @@ -130,7 +130,7 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "mov x21, #0x0\n" "sub x20, XZR, x24\n" "cbz x23, 4f\n" - "ldr q16, [x14, #0x0]\n" + "ldr q31, [x14, #0x0]\n" "ldr q0, [x14, #0x10]\n" "cmp x24, x23, LSL #4\n" "ldr q1, [x14, #0x20]\n" @@ -149,304 +149,304 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "ldr q13, [x13, x11]\n" "bge 3f\n" "2:" // Tile loop: Channel loop - "mov v24.16b, v16.16b\n fmla v24.4s, v7.4s, v9.4s\n" - "mov v23.16b, v16.16b\n fmla v23.4s, v8.4s, v9.4s\n" + "mov v29.16b, v31.16b\n fmla v29.4s, v7.4s, v9.4s\n" + "mov v28.16b, v31.16b\n fmla v28.4s, v8.4s, v9.4s\n" "add x24, x24, #0x10\n" "cmp x24, x23, LSL #4\n" - "mov v25.16b, v16.16b\n fmla v25.4s, v6.4s, v9.4s\n" - "fmla v24.4s, v4.4s, v13.4s\n" + "mov v27.16b, v31.16b\n fmla v27.4s, v6.4s, v9.4s\n" + "fmla v29.4s, v4.4s, v13.4s\n" "add x20, x20, #0x10\n" "add x21, x21, #0x10\n" - "mov v26.16b, v16.16b\n fmla v26.4s, v5.4s, v9.4s\n" - "mov v27.16b, v16.16b\n fmla v27.4s, v4.4s, v9.4s\n" - "mov v28.16b, v16.16b\n fmla v28.4s, v3.4s, v9.4s\n" - "fmla v23.4s, v0.4s, v10.4s\n" - "ldr q10, [x12, x9]\n" - "fmla v25.4s, v2.4s, v11.4s\n" - "ldr q11, [x12, x8]\n" - "mov v29.16b, v16.16b\n fmla v29.4s, v2.4s, v9.4s\n" - "fmla v24.4s, v6.4s, v11.4s\n" - "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v9.4s\n" - "fmla v23.4s, v5.4s, v13.4s\n" - "fmla v25.4s, v3.4s, v13.4s\n" + "mov v26.16b, v31.16b\n fmla v26.4s, v5.4s, v9.4s\n" + "mov v25.16b, v31.16b\n fmla v25.4s, v4.4s, v9.4s\n" + "mov v24.16b, v31.16b\n fmla v24.4s, v3.4s, v9.4s\n" + "fmla v28.4s, v0.4s, v10.4s\n" + "ldr q23, [x12, x9]\n" + "fmla v27.4s, v2.4s, v11.4s\n" + "ldr q18, [x12, x8]\n" + "mov v22.16b, v31.16b\n fmla v22.4s, v2.4s, v9.4s\n" + "fmla v29.4s, v6.4s, v18.4s\n" + "mov v21.16b, v31.16b\n fmla v21.4s, v0.4s, v9.4s\n" + "fmla v28.4s, v5.4s, v13.4s\n" + "fmla v27.4s, v3.4s, v13.4s\n" "fmla v26.4s, v2.4s, v13.4s\n" - "fmla v27.4s, v1.4s, v13.4s\n" - "fmla v28.4s, v0.4s, v13.4s\n" - "ldr q13, [x16, x8]\n" - "fmla v29.4s, v6.4s, v12.4s\n" - "ldr q12, [x27, x26]\n" - "mov v30.16b, v16.16b\n fmla v30.4s, v1.4s, v9.4s\n" - "ldr q16, [x14, #0x0]\n" + "fmla v25.4s, v1.4s, v13.4s\n" "fmla v24.4s, v0.4s, v13.4s\n" - "fmla v31.4s, v8.4s, v12.4s\n" - "ldr q12, [x16, x9]\n" - "fmla v23.4s, v7.4s, v11.4s\n" - "fmla v30.4s, v0.4s, v11.4s\n" - "fmla v26.4s, v4.4s, v11.4s\n" - "fmla v27.4s, v3.4s, v11.4s\n" - "fmla v29.4s, v1.4s, v11.4s\n" - "ld1 { v11.4s }, [x13]\n" - "fmla v24.4s, v2.4s, v12.4s\n" - "fmla v25.4s, v1.4s, v12.4s\n" - "ld1 { v12.4s }, [x10]\n" - "fmla v28.4s, v4.4s, v10.4s\n" - "fmla v23.4s, v1.4s, v13.4s\n" - "ldr q13, [x13, x26]\n" - "fmla v30.4s, v2.4s, v10.4s\n" - "fmla v31.4s, v1.4s, v10.4s\n" - "fmla v24.4s, v8.4s, v10.4s\n" - "fmla v25.4s, v7.4s, v10.4s\n" - "fmla v27.4s, v5.4s, v10.4s\n" - "ldr q10, [x10, x11]\n" - "fmla v26.4s, v0.4s, v11.4s\n" - "fmla v29.4s, v3.4s, v12.4s\n" - "fmla v28.4s, v2.4s, v13.4s\n" - "fmla v30.4s, v4.4s, v10.4s\n" - "fmla v31.4s, v3.4s, v10.4s\n" - "fmla v23.4s, v3.4s, v11.4s\n" - "ldr q11, [x10, x26]\n" - "fmla v25.4s, v5.4s, v13.4s\n" - "ldr q13, [x27, x8]\n" - "fmla v26.4s, v6.4s, v12.4s\n" - "ldr q12, [x13, x8]\n" - "fmla v27.4s, v7.4s, v10.4s\n" - "fmla v29.4s, v5.4s, v10.4s\n" - "fmla v28.4s, v6.4s, v10.4s\n" - "fmla v31.4s, v5.4s, v11.4s\n" - "fmla v30.4s, v6.4s, v13.4s\n" - "fmla v26.4s, v8.4s, v10.4s\n" - "fmla v29.4s, v7.4s, v13.4s\n" - "ldr q13, [x27, x9]\n" - "fmla v24.4s, v3.4s, v12.4s\n" - "fmla v27.4s, v0.4s, v12.4s\n" - "fmla v28.4s, v8.4s, v11.4s\n" - "ldr q11, [x13, x9]\n" - "fmla v30.4s, v8.4s, v13.4s\n" + "ldr q17, [x16, x8]\n" + "fmla v22.4s, v6.4s, v12.4s\n" + "ldr q16, [x27, x26]\n" + "mov v20.16b, v31.16b\n fmla v20.4s, v1.4s, v9.4s\n" + "ldr q31, [x14, #0x0]\n" + "fmla v29.4s, v0.4s, v17.4s\n" + "fmla v21.4s, v8.4s, v16.4s\n" + "ldr q16, [x16, x9]\n" + "fmla v28.4s, v7.4s, v18.4s\n" + "fmla v20.4s, v0.4s, v18.4s\n" + "fmla v26.4s, v4.4s, v18.4s\n" + "fmla v25.4s, v3.4s, v18.4s\n" + "fmla v22.4s, v1.4s, v18.4s\n" + "ld1 { v19.4s }, [x13]\n" + "fmla v29.4s, v2.4s, v16.4s\n" + "fmla v27.4s, v1.4s, v16.4s\n" + "ld1 { v18.4s }, [x10]\n" + "fmla v24.4s, v4.4s, v23.4s\n" + "fmla v28.4s, v1.4s, v17.4s\n" + "ldr q16, [x13, x26]\n" + "fmla v20.4s, v2.4s, v23.4s\n" + "fmla v21.4s, v1.4s, v23.4s\n" + "fmla v29.4s, v8.4s, v23.4s\n" + "fmla v27.4s, v7.4s, v23.4s\n" + "fmla v25.4s, v5.4s, v23.4s\n" + "ldr q17, [x10, x11]\n" + "fmla v26.4s, v0.4s, v19.4s\n" + "fmla v22.4s, v3.4s, v18.4s\n" + "fmla v24.4s, v2.4s, v16.4s\n" + "fmla v20.4s, v4.4s, v17.4s\n" + "fmla v21.4s, v3.4s, v17.4s\n" + "fmla v28.4s, v3.4s, v19.4s\n" + "ldr q19, [x10, x26]\n" + "fmla v27.4s, v5.4s, v16.4s\n" + "ldr q16, [x27, x8]\n" + "fmla v26.4s, v6.4s, v18.4s\n" + "ldr q18, [x13, x8]\n" + "fmla v25.4s, v7.4s, v17.4s\n" + "fmla v22.4s, v5.4s, v17.4s\n" + "fmla v24.4s, v6.4s, v17.4s\n" + "fmla v21.4s, v5.4s, v19.4s\n" + "fmla v20.4s, v6.4s, v16.4s\n" + "fmla v26.4s, v8.4s, v17.4s\n" + "fmla v22.4s, v7.4s, v16.4s\n" + "ldr q17, [x27, x9]\n" + "fmla v29.4s, v3.4s, v18.4s\n" + "fmla v25.4s, v0.4s, v18.4s\n" + "fmla v24.4s, v8.4s, v19.4s\n" + "ldr q16, [x13, x9]\n" + "fmla v20.4s, v8.4s, v17.4s\n" "add x13, x13, #0x10\n" - "fmla v31.4s, v7.4s, v13.4s\n" - "ldr q13, [x10, x9]\n" - "fmla v23.4s, v4.4s, v12.4s\n" - "fmla v26.4s, v1.4s, v12.4s\n" - "ldr q12, [x10, x8]\n" - "fmla v24.4s, v5.4s, v11.4s\n" + "fmla v21.4s, v7.4s, v17.4s\n" + "ldr q19, [x10, x9]\n" + "fmla v28.4s, v4.4s, v18.4s\n" + "fmla v26.4s, v1.4s, v18.4s\n" + "ldr q17, [x10, x8]\n" + "fmla v29.4s, v5.4s, v16.4s\n" "add x10, x10, #0x10\n" - "fmla v25.4s, v4.4s, v11.4s\n" - "fmla v27.4s, v2.4s, v11.4s\n" - "fmla v28.4s, v1.4s, v11.4s\n" - "ldr q11, [x16, x11]\n" - "fmla v29.4s, v4.4s, v12.4s\n" + "fmla v27.4s, v4.4s, v16.4s\n" + "fmla v25.4s, v2.4s, v16.4s\n" + "fmla v24.4s, v1.4s, v16.4s\n" + "ldr q16, [x16, x11]\n" + "fmla v22.4s, v4.4s, v17.4s\n" "add x16, x16, #0x10\n" "ld1 { v10.4s }, [x16]\n" - "fmla v30.4s, v3.4s, v12.4s\n" - "fmla v31.4s, v4.4s, v13.4s\n" + "fmla v20.4s, v3.4s, v17.4s\n" + "fmla v21.4s, v4.4s, v19.4s\n" "ldr q4, [x14, #0x50]\n" - "fmla v26.4s, v7.4s, v12.4s\n" - "fmla v27.4s, v6.4s, v12.4s\n" - "ld1 { v12.4s }, [x12]\n" - "fmla v23.4s, v2.4s, v11.4s\n" - "fmla v24.4s, v1.4s, v11.4s\n" + "fmla v26.4s, v7.4s, v17.4s\n" + "fmla v25.4s, v6.4s, v17.4s\n" + "ld1 { v18.4s }, [x12]\n" + "fmla v28.4s, v2.4s, v16.4s\n" + "fmla v29.4s, v1.4s, v16.4s\n" "ldr q1, [x14, #0x20]\n" - "fmax v24.4s, v24.4s, v18.4s\n" - "fmla v25.4s, v0.4s, v11.4s\n" - "ldr q11, [x12, x26]\n" - "fmla v28.4s, v7.4s, v13.4s\n" + "fmax v29.4s, v29.4s, v15.4s\n" + "fmla v27.4s, v0.4s, v16.4s\n" + "ldr q17, [x12, x26]\n" + "fmla v24.4s, v7.4s, v19.4s\n" "add x12, x12, #0x10\n" "ldr q9, [x12, x11]\n" - "fmla v30.4s, v5.4s, v13.4s\n" - "fmla v29.4s, v0.4s, v12.4s\n" + "fmla v20.4s, v5.4s, v19.4s\n" + "fmla v22.4s, v0.4s, v18.4s\n" "ldr q0, [x14, #0x10]\n" - "fmla v31.4s, v2.4s, v11.4s\n" + "fmla v21.4s, v2.4s, v17.4s\n" "ldr q2, [x14, #0x30]\n" - "fmla v27.4s, v8.4s, v13.4s\n" - "ldr q13, [x27, x11]\n" - "fmla v23.4s, v6.4s, v12.4s\n" - "fmla v26.4s, v3.4s, v12.4s\n" + "fmla v25.4s, v8.4s, v19.4s\n" + "ldr q16, [x27, x11]\n" + "fmla v28.4s, v6.4s, v18.4s\n" + "fmla v26.4s, v3.4s, v18.4s\n" "ldr q3, [x14, #0x40]\n" - "fmax v23.4s, v23.4s, v18.4s\n" - "fmla v25.4s, v8.4s, v11.4s\n" - "fmla v28.4s, v5.4s, v11.4s\n" + "fmax v28.4s, v28.4s, v15.4s\n" + "fmla v27.4s, v8.4s, v17.4s\n" + "fmla v24.4s, v5.4s, v17.4s\n" "ldr q11, [x16, x26]\n" "ldr q5, [x14, #0x60]\n" - "fmla v29.4s, v8.4s, v13.4s\n" + "fmla v22.4s, v8.4s, v16.4s\n" "ldr q8, [x14, #0x90]\n" - "fmla v30.4s, v7.4s, v13.4s\n" + "fmla v20.4s, v7.4s, v16.4s\n" "ldr q7, [x14, #0x80]\n" - "fmla v31.4s, v6.4s, v13.4s\n" + "fmla v21.4s, v6.4s, v16.4s\n" "ldr q13, [x13, x11]\n" "ldr q6, [x14, #0x70]\n" - "fmax v25.4s, v25.4s, v18.4s\n" - "fmax v26.4s, v26.4s, v18.4s\n" - "fmax v27.4s, v27.4s, v18.4s\n" + "fmax v27.4s, v27.4s, v15.4s\n" + "fmax v26.4s, v26.4s, v15.4s\n" + "fmax v25.4s, v25.4s, v15.4s\n" "add x27, x27, #0x10\n" "ld1 { v12.4s }, [x27]\n" - "fmax v28.4s, v28.4s, v18.4s\n" - "fmax v29.4s, v29.4s, v18.4s\n" + "fmax v24.4s, v24.4s, v15.4s\n" + "fmax v22.4s, v22.4s, v15.4s\n" "add x14, x14, #0xa0\n" - "fmax v30.4s, v30.4s, v18.4s\n" - "fmax v31.4s, v31.4s, v18.4s\n" - "fmin v23.4s, v23.4s, v17.4s\n" - "fmin v24.4s, v24.4s, v17.4s\n" - "st1 { v23.4s }, [x15]\n" - "fmin v25.4s, v25.4s, v17.4s\n" - "fmin v26.4s, v26.4s, v17.4s\n" - "str q24, [x15, x17]\n" - "fmin v27.4s, v27.4s, v17.4s\n" - "fmin v28.4s, v28.4s, v17.4s\n" - "str q25, [x15, x22]\n" + "fmax v20.4s, v20.4s, v15.4s\n" + "fmax v21.4s, v21.4s, v15.4s\n" + "fmin v28.4s, v28.4s, v14.4s\n" + "fmin v29.4s, v29.4s, v14.4s\n" + "st1 { v28.4s }, [x15]\n" + "fmin v27.4s, v27.4s, v14.4s\n" + "fmin v26.4s, v26.4s, v14.4s\n" + "str q29, [x15, x17]\n" + "fmin v25.4s, v25.4s, v14.4s\n" + "fmin v24.4s, v24.4s, v14.4s\n" + "str q27, [x15, x22]\n" "add x15, x15, #0x10\n" - "fmin v29.4s, v29.4s, v17.4s\n" - "fmin v30.4s, v30.4s, v17.4s\n" + "fmin v22.4s, v22.4s, v14.4s\n" + "fmin v20.4s, v20.4s, v14.4s\n" "st1 { v26.4s }, [x28]\n" - "fmin v31.4s, v31.4s, v17.4s\n" - "str q27, [x28, x17]\n" - "str q28, [x28, x22]\n" + "fmin v21.4s, v21.4s, v14.4s\n" + "str q25, [x28, x17]\n" + "str q24, [x28, x22]\n" "add x28, x28, #0x10\n" - "st1 { v29.4s }, [x25]\n" - "str q30, [x25, x17]\n" - "str q31, [x25, x22]\n" + "st1 { v22.4s }, [x25]\n" + "str q20, [x25, x17]\n" + "str q21, [x25, x22]\n" "add x25, x25, #0x10\n" "blt 2b\n" "3:" // Tile loop: Channel tail - "mov v24.16b, v16.16b\n fmla v24.4s, v7.4s, v9.4s\n" - "mov v23.16b, v16.16b\n fmla v23.4s, v8.4s, v9.4s\n" - "mov v25.16b, v16.16b\n fmla v25.4s, v6.4s, v9.4s\n" - "fmla v24.4s, v4.4s, v13.4s\n" - "mov v26.16b, v16.16b\n fmla v26.4s, v5.4s, v9.4s\n" - "mov v27.16b, v16.16b\n fmla v27.4s, v4.4s, v9.4s\n" - "mov v28.16b, v16.16b\n fmla v28.4s, v3.4s, v9.4s\n" - "fmla v23.4s, v0.4s, v10.4s\n" - "ldr q10, [x12, x9]\n" - "fmla v25.4s, v2.4s, v11.4s\n" - "ldr q11, [x12, x8]\n" - "mov v29.16b, v16.16b\n fmla v29.4s, v2.4s, v9.4s\n" - "fmla v24.4s, v6.4s, v11.4s\n" - "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v9.4s\n" - "fmla v23.4s, v5.4s, v13.4s\n" - "fmla v25.4s, v3.4s, v13.4s\n" + "mov v29.16b, v31.16b\n fmla v29.4s, v7.4s, v9.4s\n" + "mov v28.16b, v31.16b\n fmla v28.4s, v8.4s, v9.4s\n" + "mov v27.16b, v31.16b\n fmla v27.4s, v6.4s, v9.4s\n" + "fmla v29.4s, v4.4s, v13.4s\n" + "mov v26.16b, v31.16b\n fmla v26.4s, v5.4s, v9.4s\n" + "mov v25.16b, v31.16b\n fmla v25.4s, v4.4s, v9.4s\n" + "mov v24.16b, v31.16b\n fmla v24.4s, v3.4s, v9.4s\n" + "fmla v28.4s, v0.4s, v10.4s\n" + "ldr q23, [x12, x9]\n" + "fmla v27.4s, v2.4s, v11.4s\n" + "ldr q18, [x12, x8]\n" + "mov v22.16b, v31.16b\n fmla v22.4s, v2.4s, v9.4s\n" + "fmla v29.4s, v6.4s, v18.4s\n" + "mov v21.16b, v31.16b\n fmla v21.4s, v0.4s, v9.4s\n" + "fmla v28.4s, v5.4s, v13.4s\n" + "fmla v27.4s, v3.4s, v13.4s\n" "fmla v26.4s, v2.4s, v13.4s\n" - "fmla v27.4s, v1.4s, v13.4s\n" - "fmla v28.4s, v0.4s, v13.4s\n" - "ldr q13, [x16, x8]\n" - "fmla v29.4s, v6.4s, v12.4s\n" - "ldr q12, [x27, x26]\n" - "mov v30.16b, v16.16b\n fmla v30.4s, v1.4s, v9.4s\n" + "fmla v25.4s, v1.4s, v13.4s\n" "fmla v24.4s, v0.4s, v13.4s\n" - "fmla v31.4s, v8.4s, v12.4s\n" - "ldr q12, [x16, x9]\n" - "fmla v23.4s, v7.4s, v11.4s\n" - "fmla v30.4s, v0.4s, v11.4s\n" - "fmla v26.4s, v4.4s, v11.4s\n" - "fmla v27.4s, v3.4s, v11.4s\n" - "fmla v29.4s, v1.4s, v11.4s\n" - "ld1 { v11.4s }, [x13]\n" - "fmla v24.4s, v2.4s, v12.4s\n" - "fmla v25.4s, v1.4s, v12.4s\n" - "ld1 { v12.4s }, [x10]\n" - "fmla v28.4s, v4.4s, v10.4s\n" - "fmla v23.4s, v1.4s, v13.4s\n" - "ldr q13, [x13, x26]\n" - "fmla v30.4s, v2.4s, v10.4s\n" - "fmla v31.4s, v1.4s, v10.4s\n" - "fmla v24.4s, v8.4s, v10.4s\n" - "fmla v25.4s, v7.4s, v10.4s\n" - "fmla v27.4s, v5.4s, v10.4s\n" - "ldr q10, [x10, x11]\n" - "fmla v26.4s, v0.4s, v11.4s\n" - "fmla v29.4s, v3.4s, v12.4s\n" - "fmla v28.4s, v2.4s, v13.4s\n" - "fmla v30.4s, v4.4s, v10.4s\n" - "fmla v31.4s, v3.4s, v10.4s\n" - "fmla v23.4s, v3.4s, v11.4s\n" - "ldr q11, [x10, x26]\n" - "fmla v25.4s, v5.4s, v13.4s\n" - "ldr q13, [x27, x8]\n" - "fmla v26.4s, v6.4s, v12.4s\n" - "ldr q12, [x13, x8]\n" - "fmla v27.4s, v7.4s, v10.4s\n" - "fmla v29.4s, v5.4s, v10.4s\n" - "fmla v28.4s, v6.4s, v10.4s\n" - "fmla v31.4s, v5.4s, v11.4s\n" - "fmla v30.4s, v6.4s, v13.4s\n" - "fmla v26.4s, v8.4s, v10.4s\n" - "fmla v29.4s, v7.4s, v13.4s\n" - "ldr q13, [x27, x9]\n" - "fmla v24.4s, v3.4s, v12.4s\n" - "fmla v27.4s, v0.4s, v12.4s\n" - "fmla v28.4s, v8.4s, v11.4s\n" - "ldr q11, [x13, x9]\n" - "fmla v30.4s, v8.4s, v13.4s\n" + "ldr q17, [x16, x8]\n" + "fmla v22.4s, v6.4s, v12.4s\n" + "ldr q16, [x27, x26]\n" + "mov v20.16b, v31.16b\n fmla v20.4s, v1.4s, v9.4s\n" + "fmla v29.4s, v0.4s, v17.4s\n" + "fmla v21.4s, v8.4s, v16.4s\n" + "ldr q16, [x16, x9]\n" + "fmla v28.4s, v7.4s, v18.4s\n" + "fmla v20.4s, v0.4s, v18.4s\n" + "fmla v26.4s, v4.4s, v18.4s\n" + "fmla v25.4s, v3.4s, v18.4s\n" + "fmla v22.4s, v1.4s, v18.4s\n" + "ld1 { v19.4s }, [x13]\n" + "fmla v29.4s, v2.4s, v16.4s\n" + "fmla v27.4s, v1.4s, v16.4s\n" + "ld1 { v18.4s }, [x10]\n" + "fmla v24.4s, v4.4s, v23.4s\n" + "fmla v28.4s, v1.4s, v17.4s\n" + "ldr q16, [x13, x26]\n" + "fmla v20.4s, v2.4s, v23.4s\n" + "fmla v21.4s, v1.4s, v23.4s\n" + "fmla v29.4s, v8.4s, v23.4s\n" + "fmla v27.4s, v7.4s, v23.4s\n" + "fmla v25.4s, v5.4s, v23.4s\n" + "ldr q17, [x10, x11]\n" + "fmla v26.4s, v0.4s, v19.4s\n" + "fmla v22.4s, v3.4s, v18.4s\n" + "fmla v24.4s, v2.4s, v16.4s\n" + "fmla v20.4s, v4.4s, v17.4s\n" + "fmla v21.4s, v3.4s, v17.4s\n" + "fmla v28.4s, v3.4s, v19.4s\n" + "ldr q19, [x10, x26]\n" + "fmla v27.4s, v5.4s, v16.4s\n" + "ldr q16, [x27, x8]\n" + "fmla v26.4s, v6.4s, v18.4s\n" + "ldr q18, [x13, x8]\n" + "fmla v25.4s, v7.4s, v17.4s\n" + "fmla v22.4s, v5.4s, v17.4s\n" + "fmla v24.4s, v6.4s, v17.4s\n" + "fmla v21.4s, v5.4s, v19.4s\n" + "fmla v20.4s, v6.4s, v16.4s\n" + "fmla v26.4s, v8.4s, v17.4s\n" + "fmla v22.4s, v7.4s, v16.4s\n" + "ldr q17, [x27, x9]\n" + "fmla v29.4s, v3.4s, v18.4s\n" + "fmla v25.4s, v0.4s, v18.4s\n" + "fmla v24.4s, v8.4s, v19.4s\n" + "ldr q16, [x13, x9]\n" + "fmla v20.4s, v8.4s, v17.4s\n" "add x13, x13, #0x10\n" - "fmla v31.4s, v7.4s, v13.4s\n" - "ldr q13, [x10, x9]\n" - "fmla v23.4s, v4.4s, v12.4s\n" - "fmla v26.4s, v1.4s, v12.4s\n" - "ldr q12, [x10, x8]\n" - "fmla v24.4s, v5.4s, v11.4s\n" + "fmla v21.4s, v7.4s, v17.4s\n" + "ldr q19, [x10, x9]\n" + "fmla v28.4s, v4.4s, v18.4s\n" + "fmla v26.4s, v1.4s, v18.4s\n" + "ldr q17, [x10, x8]\n" + "fmla v29.4s, v5.4s, v16.4s\n" "add x10, x10, #0x10\n" - "fmla v25.4s, v4.4s, v11.4s\n" - "fmla v27.4s, v2.4s, v11.4s\n" - "fmla v28.4s, v1.4s, v11.4s\n" - "ldr q11, [x16, x11]\n" - "fmla v29.4s, v4.4s, v12.4s\n" + "fmla v27.4s, v4.4s, v16.4s\n" + "fmla v25.4s, v2.4s, v16.4s\n" + "fmla v24.4s, v1.4s, v16.4s\n" + "ldr q16, [x16, x11]\n" + "fmla v22.4s, v4.4s, v17.4s\n" "add x16, x16, #0x10\n" - "fmla v30.4s, v3.4s, v12.4s\n" - "fmla v31.4s, v4.4s, v13.4s\n" - "fmla v26.4s, v7.4s, v12.4s\n" - "fmla v27.4s, v6.4s, v12.4s\n" - "ld1 { v12.4s }, [x12]\n" - "fmla v23.4s, v2.4s, v11.4s\n" - "fmla v24.4s, v1.4s, v11.4s\n" - "fmax v24.4s, v24.4s, v18.4s\n" - "fmla v25.4s, v0.4s, v11.4s\n" - "ldr q11, [x12, x26]\n" - "fmla v28.4s, v7.4s, v13.4s\n" - "fmin v24.4s, v24.4s, v17.4s\n" - "fmla v30.4s, v5.4s, v13.4s\n" - "fmla v29.4s, v0.4s, v12.4s\n" + "fmla v20.4s, v3.4s, v17.4s\n" + "fmla v21.4s, v4.4s, v19.4s\n" + "fmla v26.4s, v7.4s, v17.4s\n" + "fmla v25.4s, v6.4s, v17.4s\n" + "ld1 { v18.4s }, [x12]\n" + "fmla v28.4s, v2.4s, v16.4s\n" + "fmla v29.4s, v1.4s, v16.4s\n" + "fmax v29.4s, v29.4s, v15.4s\n" + "fmla v27.4s, v0.4s, v16.4s\n" + "ldr q17, [x12, x26]\n" + "fmla v24.4s, v7.4s, v19.4s\n" + "fmin v29.4s, v29.4s, v14.4s\n" + "fmla v20.4s, v5.4s, v19.4s\n" + "fmla v22.4s, v0.4s, v18.4s\n" "add x12, x12, #0x10\n" - "fmla v31.4s, v2.4s, v11.4s\n" - "fmla v27.4s, v8.4s, v13.4s\n" - "ldr q13, [x27, x11]\n" - "fmax v27.4s, v27.4s, v18.4s\n" - "fmla v23.4s, v6.4s, v12.4s\n" - "fmla v26.4s, v3.4s, v12.4s\n" - "fmax v23.4s, v23.4s, v18.4s\n" + "fmla v21.4s, v2.4s, v17.4s\n" + "fmla v25.4s, v8.4s, v19.4s\n" + "ldr q16, [x27, x11]\n" + "fmax v25.4s, v25.4s, v15.4s\n" + "fmla v28.4s, v6.4s, v18.4s\n" + "fmla v26.4s, v3.4s, v18.4s\n" + "fmax v28.4s, v28.4s, v15.4s\n" "add x27, x27, #0x10\n" - "fmla v25.4s, v8.4s, v11.4s\n" - "fmla v28.4s, v5.4s, v11.4s\n" - "fmax v25.4s, v25.4s, v18.4s\n" - "fmla v29.4s, v8.4s, v13.4s\n" - "fmla v30.4s, v7.4s, v13.4s\n" - "fmax v26.4s, v26.4s, v18.4s\n" - "fmla v31.4s, v6.4s, v13.4s\n" - "fmax v28.4s, v28.4s, v18.4s\n" - "fmax v29.4s, v29.4s, v18.4s\n" - "fmax v30.4s, v30.4s, v18.4s\n" - "fmax v31.4s, v31.4s, v18.4s\n" - "fmin v23.4s, v23.4s, v17.4s\n" - "st1 { v23.4s }, [x15]\n" - "fmin v25.4s, v25.4s, v17.4s\n" - "fmin v26.4s, v26.4s, v17.4s\n" - "str q24, [x15, x17]\n" - "fmin v27.4s, v27.4s, v17.4s\n" - "fmin v28.4s, v28.4s, v17.4s\n" - "str q25, [x15, x22]\n" + "fmla v27.4s, v8.4s, v17.4s\n" + "fmla v24.4s, v5.4s, v17.4s\n" + "fmax v27.4s, v27.4s, v15.4s\n" + "fmla v22.4s, v8.4s, v16.4s\n" + "fmla v20.4s, v7.4s, v16.4s\n" + "fmax v26.4s, v26.4s, v15.4s\n" + "fmla v21.4s, v6.4s, v16.4s\n" + "fmax v24.4s, v24.4s, v15.4s\n" + "fmax v22.4s, v22.4s, v15.4s\n" + "fmax v20.4s, v20.4s, v15.4s\n" + "fmax v21.4s, v21.4s, v15.4s\n" + "fmin v28.4s, v28.4s, v14.4s\n" + "st1 { v28.4s }, [x15]\n" + "fmin v27.4s, v27.4s, v14.4s\n" + "fmin v26.4s, v26.4s, v14.4s\n" + "str q29, [x15, x17]\n" + "fmin v25.4s, v25.4s, v14.4s\n" + "fmin v24.4s, v24.4s, v14.4s\n" + "str q27, [x15, x22]\n" "add x15, x15, #0x10\n" - "fmin v29.4s, v29.4s, v17.4s\n" - "fmin v30.4s, v30.4s, v17.4s\n" + "fmin v22.4s, v22.4s, v14.4s\n" + "fmin v20.4s, v20.4s, v14.4s\n" "st1 { v26.4s }, [x28]\n" - "fmin v31.4s, v31.4s, v17.4s\n" - "str q27, [x28, x17]\n" - "str q28, [x28, x22]\n" + "fmin v21.4s, v21.4s, v14.4s\n" + "str q25, [x28, x17]\n" + "str q24, [x28, x22]\n" "add x28, x28, #0x10\n" - "st1 { v29.4s }, [x25]\n" - "str q30, [x25, x17]\n" - "str q31, [x25, x22]\n" + "st1 { v22.4s }, [x25]\n" + "str q20, [x25, x17]\n" + "str q21, [x25, x22]\n" "add x25, x25, #0x10\n" "4:" // Tile loop: Oddments "tst %x[n_channels], #0x3\n" "beq 49f\n" - "ldr q16, [x14, #0x0]\n" + "ldr q31, [x14, #0x0]\n" "ldr q0, [x14, #0x10]\n" "add x24, x12, x11\n" "add x23, x16, XZR\n" @@ -481,18 +481,18 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "ldr s12, [x21, #0x0]\n" "ldr s13, [x20, #0x0]\n" "6:" // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 4), (4, 0), (1, 2): Bit 1: End - "mov v23.16b, v16.16b\n fmla v23.4s, v8.4s, v9.4s\n" - "mov v25.16b, v16.16b\n fmla v25.4s, v6.4s, v9.4s\n" + "mov v23.16b, v31.16b\n fmla v23.4s, v8.4s, v9.4s\n" + "mov v25.16b, v31.16b\n fmla v25.4s, v6.4s, v9.4s\n" "add x20, x27, x26\n" - "mov v24.16b, v16.16b\n fmla v24.4s, v7.4s, v9.4s\n" - "mov v26.16b, v16.16b\n fmla v26.4s, v5.4s, v9.4s\n" - "mov v27.16b, v16.16b\n fmla v27.4s, v4.4s, v9.4s\n" - "mov v28.16b, v16.16b\n fmla v28.4s, v3.4s, v9.4s\n" - "mov v29.16b, v16.16b\n fmla v29.4s, v2.4s, v9.4s\n" + "mov v24.16b, v31.16b\n fmla v24.4s, v7.4s, v9.4s\n" + "mov v26.16b, v31.16b\n fmla v26.4s, v5.4s, v9.4s\n" + "mov v27.16b, v31.16b\n fmla v27.4s, v4.4s, v9.4s\n" + "mov v28.16b, v31.16b\n fmla v28.4s, v3.4s, v9.4s\n" + "mov v29.16b, v31.16b\n fmla v29.4s, v2.4s, v9.4s\n" "fmla v23.4s, v0.4s, v10.4s\n" "fmla v25.4s, v2.4s, v11.4s\n" - "mov v30.16b, v16.16b\n fmla v30.4s, v1.4s, v9.4s\n" - "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v9.4s\n" + "mov v30.16b, v31.16b\n fmla v30.4s, v1.4s, v9.4s\n" + "fmla v31.4s, v0.4s, v9.4s\n" "fmla v29.4s, v6.4s, v12.4s\n" "fmla v23.4s, v5.4s, v13.4s\n" "fmla v24.4s, v4.4s, v13.4s\n" @@ -741,25 +741,25 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "46:" // Tile loop: Oddments: Load inputs: (4, 2): Bit 1: End "fmla v29.4s, v8.4s, v13.4s\n" "fmla v30.4s, v7.4s, v13.4s\n" - "fmax v23.4s, v23.4s, v18.4s\n" + "fmax v23.4s, v23.4s, v15.4s\n" "fmla v31.4s, v6.4s, v13.4s\n" - "fmax v24.4s, v24.4s, v18.4s\n" - "fmax v25.4s, v25.4s, v18.4s\n" - "fmax v26.4s, v26.4s, v18.4s\n" - "fmax v27.4s, v27.4s, v18.4s\n" - "fmax v28.4s, v28.4s, v18.4s\n" - "fmax v29.4s, v29.4s, v18.4s\n" - "fmax v30.4s, v30.4s, v18.4s\n" - "fmax v31.4s, v31.4s, v18.4s\n" - "fmin v23.4s, v23.4s, v17.4s\n" - "fmin v24.4s, v24.4s, v17.4s\n" - "fmin v25.4s, v25.4s, v17.4s\n" - "fmin v26.4s, v26.4s, v17.4s\n" - "fmin v27.4s, v27.4s, v17.4s\n" - "fmin v28.4s, v28.4s, v17.4s\n" - "fmin v29.4s, v29.4s, v17.4s\n" - "fmin v30.4s, v30.4s, v17.4s\n" - "fmin v31.4s, v31.4s, v17.4s\n" + "fmax v24.4s, v24.4s, v15.4s\n" + "fmax v25.4s, v25.4s, v15.4s\n" + "fmax v26.4s, v26.4s, v15.4s\n" + "fmax v27.4s, v27.4s, v15.4s\n" + "fmax v28.4s, v28.4s, v15.4s\n" + "fmax v29.4s, v29.4s, v15.4s\n" + "fmax v30.4s, v30.4s, v15.4s\n" + "fmax v31.4s, v31.4s, v15.4s\n" + "fmin v23.4s, v23.4s, v14.4s\n" + "fmin v24.4s, v24.4s, v14.4s\n" + "fmin v25.4s, v25.4s, v14.4s\n" + "fmin v26.4s, v26.4s, v14.4s\n" + "fmin v27.4s, v27.4s, v14.4s\n" + "fmin v28.4s, v28.4s, v14.4s\n" + "fmin v29.4s, v29.4s, v14.4s\n" + "fmin v30.4s, v30.4s, v14.4s\n" + "fmin v31.4s, v31.4s, v14.4s\n" "tbz %x[n_channels], #1, 47f\n" "mov x22, x15\n" "mov x21, x28\n" @@ -804,7 +804,6 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "st1 { v28.s }[0], [x21]\n" "st1 { v31.s }[0], [x20]\n" "48:" // Tile loop: Oddments: Store: Bit 1: End - "49:" // Tile loop: End "ldr x23, [%x[params_struct], %[offsetof_args_tile_j]]\n" "ldr x24, [%x[params_struct], %[offsetof_args_tile_i]]\n" @@ -819,11 +818,11 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_direct_impl( "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", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "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", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } } // namespace depthwise } // namespace arm_conv -#endif // __aarch64__ +#endif // defined(__aarch64__) diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp index 15053a337a..972f7eb535 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp @@ -25,7 +25,7 @@ #include <cstddef> #include <cstdint> -#if __aarch64__ +#if defined(__aarch64__) namespace arm_conv { namespace depthwise { @@ -87,405 +87,405 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( activation_min, activation_max); __asm__ __volatile__( - "mov x8, #0x10\n" // cntb _, ALL, #1 - "lsr x17, %x[n_channels], #0x2\n" - "ldr x16, [%x[params_struct], %[offsetof_args_outptrs]]\n" - "ldr x15, [%x[params_struct], %[offsetof_args_params]]\n" + "mov x7, #0x10\n" // cntb _, ALL, #1 + "lsr x8, %x[n_channels], #0x2\n" + "ldr x17, [%x[params_struct], %[offsetof_args_outptrs]]\n" + "ldr x16, [%x[params_struct], %[offsetof_args_params]]\n" "add x20, %x[params_struct], %[offsetof_args_min]\n" - "ld1r { v18.4s }, [x20]\n" + "ld1r { v15.4s }, [x20]\n" "add x20, %x[params_struct], %[offsetof_args_max]\n" - "ld1r { v17.4s }, [x20]\n" - "add x14, %x[params_struct], %[offsetof_Args_inptrs]\n" - "mov x13, #0x0\n" - "sub x12, XZR, x8\n" - "cbz x17, 3f\n" - "ldr q16, [x15, #0x0]\n" - "ldr q0, [x15, #0x10]\n" - "cmp x8, x17, LSL #4\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" - "add x15, x15, #0xa0\n" - "ldp x11, x10, [x14, #0x0]\n" - "ldr q9, [x11, x13]\n" - "ldr q10, [x10, x13]\n" - "ldp x9, x28, [x14, #0x10]\n" - "ldr q11, [x9, x13]\n" - "ldr q12, [x28, x13]\n" - "ldr x27, [x14, #0x20]\n" - "ldr q13, [x27, x13]\n" + "ld1r { v14.4s }, [x20]\n" + "add x15, %x[params_struct], %[offsetof_Args_inptrs]\n" + "mov x14, #0x0\n" + "sub x13, XZR, x7\n" + "cbz x8, 3f\n" + "ldr q31, [x16, #0x0]\n" + "ldr q0, [x16, #0x10]\n" + "cmp x7, x8, LSL #4\n" + "ldr q1, [x16, #0x20]\n" + "ldr q2, [x16, #0x30]\n" + "ldr q3, [x16, #0x40]\n" + "ldr q4, [x16, #0x50]\n" + "ldr q5, [x16, #0x60]\n" + "ldr q6, [x16, #0x70]\n" + "ldr q7, [x16, #0x80]\n" + "ldr q8, [x16, #0x90]\n" + "add x16, x16, #0xa0\n" + "ldp x21, x20, [x15, #0x0]\n" + "ldr q9, [x21, x14]\n" + "ldr q10, [x20, x14]\n" + "ldp x21, x20, [x15, #0x10]\n" + "ldr q11, [x21, x14]\n" + "ldr q12, [x20, x14]\n" + "ldr x20, [x15, #0x20]\n" + "ldr q13, [x20, x14]\n" "bge 2f\n" "1:" // Channel loop - "mov v23.16b, v16.16b\n fmla v23.4s, v8.4s, v9.4s\n" - "mov v24.16b, v16.16b\n fmla v24.4s, v7.4s, v9.4s\n" - "ldr x26, [x14, #0x30]\n" - "ldr x25, [x14, #0x38]\n" - "mov v25.16b, v16.16b\n fmla v25.4s, v6.4s, v9.4s\n" - "fmla v23.4s, v0.4s, v10.4s\n" - "ldr x24, [x14, #0x28]\n" - "ldr x10, [x14, #0x48]\n" - "ldr q10, [x10, x13]\n" - "fmla v24.4s, v4.4s, v13.4s\n" - "mov v26.16b, v16.16b\n fmla v26.4s, v5.4s, v9.4s\n" - "ldr x11, [x14, #0x40]\n" - "mov v27.16b, v16.16b\n fmla v27.4s, v4.4s, v9.4s\n" - "mov v28.16b, v16.16b\n fmla v28.4s, v3.4s, v9.4s\n" - "ldr x9, [x14, #0x50]\n" - "ldr x28, [x14, #0x58]\n" - "fmla v25.4s, v2.4s, v11.4s\n" - "ldr q11, [x26, x13]\n" - "mov v29.16b, v16.16b\n fmla v29.4s, v2.4s, v9.4s\n" - "ldr x27, [x14, #0x60]\n" - "fmla v23.4s, v5.4s, v13.4s\n" - "fmla v24.4s, v6.4s, v11.4s\n" - "ldr x26, [x14, #0x70]\n" - "ldr x10, [x14, #0x88]\n" - "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v9.4s\n" - "fmla v25.4s, v3.4s, v13.4s\n" - "ldr x23, [x16, #0x0]\n" - "add x12, x12, #0x10\n" + "mov v29.16b, v31.16b\n fmla v29.4s, v8.4s, v9.4s\n" + "mov v28.16b, v31.16b\n fmla v28.4s, v7.4s, v9.4s\n" + "ldr x26, [x15, #0x30]\n" + "ldr x23, [x15, #0x38]\n" + "mov v27.16b, v31.16b\n fmla v27.4s, v6.4s, v9.4s\n" + "fmla v29.4s, v0.4s, v10.4s\n" + "ldr x22, [x15, #0x28]\n" + "ldr x20, [x15, #0x48]\n" + "ldr q19, [x20, x14]\n" + "fmla v28.4s, v4.4s, v13.4s\n" + "mov v26.16b, v31.16b\n fmla v26.4s, v5.4s, v9.4s\n" + "ldr x21, [x15, #0x40]\n" + "mov v25.16b, v31.16b\n fmla v25.4s, v4.4s, v9.4s\n" + "mov v24.16b, v31.16b\n fmla v24.4s, v3.4s, v9.4s\n" + "ldr x25, [x15, #0x50]\n" + "ldr x24, [x15, #0x58]\n" + "fmla v27.4s, v2.4s, v11.4s\n" + "ldr q17, [x26, x14]\n" + "mov v23.16b, v31.16b\n fmla v23.4s, v2.4s, v9.4s\n" + "ldr x20, [x15, #0x60]\n" + "fmla v29.4s, v5.4s, v13.4s\n" + "fmla v28.4s, v6.4s, v17.4s\n" + "ldr x12, [x15, #0x70]\n" + "ldr x11, [x15, #0x88]\n" + "mov v22.16b, v31.16b\n fmla v22.4s, v0.4s, v9.4s\n" + "fmla v27.4s, v3.4s, v13.4s\n" + "ldr x10, [x17, #0x0]\n" + "add x13, x13, #0x10\n" "fmla v26.4s, v2.4s, v13.4s\n" - "fmla v27.4s, v1.4s, v13.4s\n" - "ldr x22, [x16, #0x8]\n" - "ldr x21, [x16, #0x10]\n" - "fmla v28.4s, v0.4s, v13.4s\n" - "ldr q13, [x25, x13]\n" - "fmla v29.4s, v6.4s, v12.4s\n" - "ldr q12, [x24, x13]\n" - "mov v30.16b, v16.16b\n fmla v30.4s, v1.4s, v9.4s\n" - "ldr q16, [x15, #0x0]\n" - "fmla v23.4s, v7.4s, v11.4s\n" - "ldr x24, [x14, #0x68]\n" + "fmla v25.4s, v1.4s, v13.4s\n" + "ldr x9, [x17, #0x8]\n" + "ldr x28, [x17, #0x10]\n" "fmla v24.4s, v0.4s, v13.4s\n" - "fmla v31.4s, v8.4s, v12.4s\n" - "ldr q12, [x11, x13]\n" - "ldr x25, [x14, #0x78]\n" - "fmla v26.4s, v4.4s, v11.4s\n" - "fmla v27.4s, v3.4s, v11.4s\n" - "ldr x11, [x14, #0x80]\n" - "ldr x20, [x16, #0x18]\n" - "fmla v30.4s, v0.4s, v11.4s\n" - "fmla v28.4s, v4.4s, v10.4s\n" - "fmla v29.4s, v1.4s, v11.4s\n" - "ldr q11, [x9, x13]\n" - "fmla v23.4s, v1.4s, v13.4s\n" - "ldr q13, [x28, x13]\n" - "fmla v24.4s, v2.4s, v12.4s\n" - "fmla v25.4s, v1.4s, v12.4s\n" - "ldr q12, [x27, x13]\n" - "ldr x9, [x14, #0x90]\n" - "fmla v27.4s, v5.4s, v10.4s\n" - "fmla v30.4s, v2.4s, v10.4s\n" - "ldr x27, [x14, #0xa0]\n" - "ldr x28, [x14, #0x98]\n" - "fmla v26.4s, v0.4s, v11.4s\n" - "fmla v28.4s, v2.4s, v13.4s\n" - "fmla v24.4s, v8.4s, v10.4s\n" - "fmla v25.4s, v7.4s, v10.4s\n" - "fmla v31.4s, v1.4s, v10.4s\n" - "ldr q10, [x24, x13]\n" - "fmla v29.4s, v3.4s, v12.4s\n" - "ldr x24, [x14, #0xa8]\n" - "fmla v26.4s, v6.4s, v12.4s\n" - "ldr q12, [x11, x13]\n" - "fmla v27.4s, v7.4s, v10.4s\n" - "ldr x11, [x14, #0xc0]\n" - "fmla v28.4s, v6.4s, v10.4s\n" - "fmla v30.4s, v4.4s, v10.4s\n" - "fmla v23.4s, v3.4s, v11.4s\n" - "ldr q11, [x26, x13]\n" - "fmla v25.4s, v5.4s, v13.4s\n" - "ldr q13, [x25, x13]\n" - "fmla v29.4s, v5.4s, v10.4s\n" - "fmla v31.4s, v3.4s, v10.4s\n" - "ldr x26, [x14, #0xb0]\n" - "ldr x25, [x14, #0xb8]\n" - "fmla v26.4s, v8.4s, v10.4s\n" - "fmla v28.4s, v8.4s, v11.4s\n" - "fmla v30.4s, v6.4s, v13.4s\n" - "fmla v24.4s, v3.4s, v12.4s\n" - "fmla v27.4s, v0.4s, v12.4s\n" - "fmla v31.4s, v5.4s, v11.4s\n" - "ldr q11, [x10, x13]\n" - "fmla v29.4s, v7.4s, v13.4s\n" - "ldr q13, [x9, x13]\n" - "fmla v23.4s, v4.4s, v12.4s\n" - "fmla v26.4s, v1.4s, v12.4s\n" - "ldr q12, [x28, x13]\n" - "fmla v24.4s, v5.4s, v11.4s\n" - "fmla v25.4s, v4.4s, v11.4s\n" - "fmla v27.4s, v2.4s, v11.4s\n" - "fmla v28.4s, v1.4s, v11.4s\n" - "ldr q11, [x27, x13]\n" - "fmla v30.4s, v8.4s, v13.4s\n" - "ldr x27, [x14, #0x20]\n" - "fmla v31.4s, v7.4s, v13.4s\n" - "ldr q13, [x24, x13]\n" - "fmla v23.4s, v2.4s, v11.4s\n" - "fmla v26.4s, v7.4s, v12.4s\n" - "fmla v27.4s, v6.4s, v12.4s\n" - "fmla v29.4s, v4.4s, v12.4s\n" - "fmla v30.4s, v3.4s, v12.4s\n" - "ldr q12, [x26, x13]\n" - "fmla v31.4s, v4.4s, v13.4s\n" - "ldr q4, [x15, #0x50]\n" - "fmla v24.4s, v1.4s, v11.4s\n" - "ldr q1, [x15, #0x20]\n" - "fmla v25.4s, v0.4s, v11.4s\n" - "ldr q11, [x25, x13]\n" + "ldr q18, [x23, x14]\n" "fmla v23.4s, v6.4s, v12.4s\n" - "fmax v23.4s, v23.4s, v18.4s\n" - "fmla v28.4s, v7.4s, v13.4s\n" - "fmla v30.4s, v5.4s, v13.4s\n" - "fmin v23.4s, v23.4s, v17.4s\n" - "str q23, [x23, x12]\n" - "fmla v29.4s, v0.4s, v12.4s\n" - "ldr q0, [x15, #0x10]\n" - "fmla v31.4s, v2.4s, v11.4s\n" - "ldr q2, [x15, #0x30]\n" - "fmla v27.4s, v8.4s, v13.4s\n" - "ldr q13, [x11, x13]\n" - "fmla v26.4s, v3.4s, v12.4s\n" - "ldr q3, [x15, #0x40]\n" - "fmla v25.4s, v8.4s, v11.4s\n" - "fmla v28.4s, v5.4s, v11.4s\n" - "ldr q5, [x15, #0x60]\n" - "fmax v24.4s, v24.4s, v18.4s\n" - "fmla v29.4s, v8.4s, v13.4s\n" - "ldr q8, [x15, #0x90]\n" - "fmla v30.4s, v7.4s, v13.4s\n" - "ldr q7, [x15, #0x80]\n" - "fmla v31.4s, v6.4s, v13.4s\n" - "ldr q13, [x27, x8]\n" - "ldr q6, [x15, #0x70]\n" - "fmax v25.4s, v25.4s, v18.4s\n" - "fmax v26.4s, v26.4s, v18.4s\n" - "fmax v27.4s, v27.4s, v18.4s\n" - "ldr x23, [x16, #0x20]\n" - "ldp x11, x10, [x14, #0x0]\n" - "ldr q9, [x11, x8]\n" - "ldr q10, [x10, x8]\n" - "fmin v24.4s, v24.4s, v17.4s\n" - "fmin v25.4s, v25.4s, v17.4s\n" - "ldp x9, x28, [x14, #0x10]\n" - "ldr q11, [x9, x8]\n" - "fmin v26.4s, v26.4s, v17.4s\n" - "fmin v27.4s, v27.4s, v17.4s\n" - "ldr q12, [x28, x8]\n" - "fmax v28.4s, v28.4s, v18.4s\n" - "fmax v29.4s, v29.4s, v18.4s\n" - "str q24, [x22, x12]\n" - "fmax v30.4s, v30.4s, v18.4s\n" - "fmax v31.4s, v31.4s, v18.4s\n" - "str q25, [x21, x12]\n" - "ldr x22, [x16, #0x28]\n" - "str q26, [x20, x12]\n" - "ldr x21, [x16, #0x30]\n" - "ldr x20, [x16, #0x38]\n" - "add x8, x8, #0x10\n" - "str q27, [x23, x12]\n" - "ldr x23, [x16, #0x40]\n" - "cmp x8, x17, LSL #4\n" - "fmin v28.4s, v28.4s, v17.4s\n" - "fmin v29.4s, v29.4s, v17.4s\n" - "fmin v30.4s, v30.4s, v17.4s\n" - "add x13, x13, #0x10\n" - "str q28, [x22, x12]\n" - "fmin v31.4s, v31.4s, v17.4s\n" - "str q29, [x21, x12]\n" - "add x15, x15, #0xa0\n" - "str q30, [x20, x12]\n" - "str q31, [x23, x12]\n" + "ldr q16, [x22, x14]\n" + "mov v21.16b, v31.16b\n fmla v21.4s, v1.4s, v9.4s\n" + "ldr q31, [x16, #0x0]\n" + "fmla v29.4s, v7.4s, v17.4s\n" + "ldr x23, [x15, #0x68]\n" + "fmla v28.4s, v0.4s, v18.4s\n" + "fmla v22.4s, v8.4s, v16.4s\n" + "ldr q16, [x21, x14]\n" + "ldr x22, [x15, #0x78]\n" + "fmla v26.4s, v4.4s, v17.4s\n" + "fmla v25.4s, v3.4s, v17.4s\n" + "ldr x21, [x15, #0x80]\n" + "ldr x27, [x17, #0x18]\n" + "fmla v21.4s, v0.4s, v17.4s\n" + "fmla v24.4s, v4.4s, v19.4s\n" + "fmla v23.4s, v1.4s, v17.4s\n" + "ldr q17, [x25, x14]\n" + "fmla v29.4s, v1.4s, v18.4s\n" + "ldr q20, [x24, x14]\n" + "fmla v28.4s, v2.4s, v16.4s\n" + "fmla v27.4s, v1.4s, v16.4s\n" + "ldr q16, [x20, x14]\n" + "ldr x26, [x15, #0x90]\n" + "fmla v25.4s, v5.4s, v19.4s\n" + "fmla v21.4s, v2.4s, v19.4s\n" + "ldr x25, [x15, #0xa0]\n" + "ldr x20, [x15, #0x98]\n" + "fmla v26.4s, v0.4s, v17.4s\n" + "fmla v24.4s, v2.4s, v20.4s\n" + "fmla v28.4s, v8.4s, v19.4s\n" + "fmla v27.4s, v7.4s, v19.4s\n" + "fmla v22.4s, v1.4s, v19.4s\n" + "ldr q19, [x23, x14]\n" + "fmla v23.4s, v3.4s, v16.4s\n" + "ldr x24, [x15, #0xa8]\n" + "fmla v26.4s, v6.4s, v16.4s\n" + "ldr q18, [x21, x14]\n" + "fmla v25.4s, v7.4s, v19.4s\n" + "ldr x23, [x15, #0xc0]\n" + "fmla v24.4s, v6.4s, v19.4s\n" + "fmla v21.4s, v4.4s, v19.4s\n" + "fmla v29.4s, v3.4s, v17.4s\n" + "ldr q17, [x12, x14]\n" + "fmla v27.4s, v5.4s, v20.4s\n" + "ldr q16, [x22, x14]\n" + "fmla v23.4s, v5.4s, v19.4s\n" + "fmla v22.4s, v3.4s, v19.4s\n" + "ldr x22, [x15, #0xb0]\n" + "ldr x21, [x15, #0xb8]\n" + "fmla v26.4s, v8.4s, v19.4s\n" + "fmla v24.4s, v8.4s, v17.4s\n" + "fmla v21.4s, v6.4s, v16.4s\n" + "fmla v28.4s, v3.4s, v18.4s\n" + "fmla v25.4s, v0.4s, v18.4s\n" + "fmla v22.4s, v5.4s, v17.4s\n" + "ldr q17, [x11, x14]\n" + "fmla v23.4s, v7.4s, v16.4s\n" + "ldr q16, [x26, x14]\n" + "fmla v29.4s, v4.4s, v18.4s\n" + "fmla v26.4s, v1.4s, v18.4s\n" + "ldr q18, [x20, x14]\n" + "fmla v28.4s, v5.4s, v17.4s\n" + "fmla v27.4s, v4.4s, v17.4s\n" + "fmla v25.4s, v2.4s, v17.4s\n" + "fmla v24.4s, v1.4s, v17.4s\n" + "ldr q17, [x25, x14]\n" + "fmla v21.4s, v8.4s, v16.4s\n" + "ldr x20, [x15, #0x20]\n" + "fmla v22.4s, v7.4s, v16.4s\n" + "ldr q16, [x24, x14]\n" + "fmla v29.4s, v2.4s, v17.4s\n" + "fmla v26.4s, v7.4s, v18.4s\n" + "fmla v25.4s, v6.4s, v18.4s\n" + "fmla v23.4s, v4.4s, v18.4s\n" + "fmla v21.4s, v3.4s, v18.4s\n" + "ldr q18, [x22, x14]\n" + "fmla v22.4s, v4.4s, v16.4s\n" + "ldr q4, [x16, #0x50]\n" + "fmla v28.4s, v1.4s, v17.4s\n" + "ldr q1, [x16, #0x20]\n" + "fmla v27.4s, v0.4s, v17.4s\n" + "ldr q17, [x21, x14]\n" + "fmla v29.4s, v6.4s, v18.4s\n" + "fmax v29.4s, v29.4s, v15.4s\n" + "fmla v24.4s, v7.4s, v16.4s\n" + "fmla v21.4s, v5.4s, v16.4s\n" + "fmin v29.4s, v29.4s, v14.4s\n" + "str q29, [x10, x13]\n" + "fmla v23.4s, v0.4s, v18.4s\n" + "ldr q0, [x16, #0x10]\n" + "fmla v22.4s, v2.4s, v17.4s\n" + "ldr q2, [x16, #0x30]\n" + "fmla v25.4s, v8.4s, v16.4s\n" + "ldr q16, [x23, x14]\n" + "fmla v26.4s, v3.4s, v18.4s\n" + "ldr q3, [x16, #0x40]\n" + "fmla v27.4s, v8.4s, v17.4s\n" + "fmla v24.4s, v5.4s, v17.4s\n" + "ldr q5, [x16, #0x60]\n" + "fmax v28.4s, v28.4s, v15.4s\n" + "fmla v23.4s, v8.4s, v16.4s\n" + "ldr q8, [x16, #0x90]\n" + "fmla v21.4s, v7.4s, v16.4s\n" + "ldr q7, [x16, #0x80]\n" + "fmla v22.4s, v6.4s, v16.4s\n" + "ldr q13, [x20, x7]\n" + "ldr q6, [x16, #0x70]\n" + "fmax v27.4s, v27.4s, v15.4s\n" + "fmax v26.4s, v26.4s, v15.4s\n" + "fmax v25.4s, v25.4s, v15.4s\n" + "ldr x24, [x17, #0x20]\n" + "ldp x21, x20, [x15, #0x0]\n" + "ldr q9, [x21, x7]\n" + "ldr q10, [x20, x7]\n" + "fmin v28.4s, v28.4s, v14.4s\n" + "fmin v27.4s, v27.4s, v14.4s\n" + "ldp x21, x20, [x15, #0x10]\n" + "ldr q11, [x21, x7]\n" + "fmin v26.4s, v26.4s, v14.4s\n" + "fmin v25.4s, v25.4s, v14.4s\n" + "ldr q12, [x20, x7]\n" + "fmax v24.4s, v24.4s, v15.4s\n" + "fmax v23.4s, v23.4s, v15.4s\n" + "str q28, [x9, x13]\n" + "fmax v21.4s, v21.4s, v15.4s\n" + "fmax v22.4s, v22.4s, v15.4s\n" + "str q27, [x28, x13]\n" + "ldr x23, [x17, #0x28]\n" + "str q26, [x27, x13]\n" + "ldr x22, [x17, #0x30]\n" + "ldr x21, [x17, #0x38]\n" + "add x7, x7, #0x10\n" + "str q25, [x24, x13]\n" + "ldr x20, [x17, #0x40]\n" + "cmp x7, x8, LSL #4\n" + "fmin v24.4s, v24.4s, v14.4s\n" + "fmin v23.4s, v23.4s, v14.4s\n" + "fmin v21.4s, v21.4s, v14.4s\n" + "add x14, x14, #0x10\n" + "str q24, [x23, x13]\n" + "fmin v22.4s, v22.4s, v14.4s\n" + "str q23, [x22, x13]\n" + "add x16, x16, #0xa0\n" + "str q21, [x21, x13]\n" + "str q22, [x20, x13]\n" "blt 1b\n" "2:" // Channel tail - "mov v23.16b, v16.16b\n fmla v23.4s, v8.4s, v9.4s\n" - "mov v24.16b, v16.16b\n fmla v24.4s, v7.4s, v9.4s\n" - "ldr x26, [x14, #0x30]\n" - "ldr x25, [x14, #0x38]\n" - "mov v25.16b, v16.16b\n fmla v25.4s, v6.4s, v9.4s\n" - "fmla v23.4s, v0.4s, v10.4s\n" - "ldr x24, [x14, #0x28]\n" - "ldr x10, [x14, #0x48]\n" - "ldr q10, [x10, x13]\n" - "fmla v24.4s, v4.4s, v13.4s\n" - "mov v26.16b, v16.16b\n fmla v26.4s, v5.4s, v9.4s\n" - "ldr x11, [x14, #0x40]\n" - "mov v27.16b, v16.16b\n fmla v27.4s, v4.4s, v9.4s\n" - "mov v28.16b, v16.16b\n fmla v28.4s, v3.4s, v9.4s\n" - "ldr x9, [x14, #0x50]\n" - "ldr x28, [x14, #0x58]\n" - "fmla v25.4s, v2.4s, v11.4s\n" - "ldr q11, [x26, x13]\n" - "mov v29.16b, v16.16b\n fmla v29.4s, v2.4s, v9.4s\n" - "ldr x27, [x14, #0x60]\n" - "fmla v23.4s, v5.4s, v13.4s\n" - "fmla v24.4s, v6.4s, v11.4s\n" - "ldr x26, [x14, #0x70]\n" - "ldr x10, [x14, #0x88]\n" - "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v9.4s\n" - "fmla v25.4s, v3.4s, v13.4s\n" - "ldr x23, [x16, #0x0]\n" - "add x12, x12, #0x10\n" + "mov v29.16b, v31.16b\n fmla v29.4s, v8.4s, v9.4s\n" + "mov v28.16b, v31.16b\n fmla v28.4s, v7.4s, v9.4s\n" + "ldr x23, [x15, #0x30]\n" + "ldr x22, [x15, #0x38]\n" + "mov v27.16b, v31.16b\n fmla v27.4s, v6.4s, v9.4s\n" + "fmla v29.4s, v0.4s, v10.4s\n" + "ldr x21, [x15, #0x28]\n" + "ldr x20, [x15, #0x48]\n" + "ldr q19, [x20, x14]\n" + "fmla v28.4s, v4.4s, v13.4s\n" + "mov v26.16b, v31.16b\n fmla v26.4s, v5.4s, v9.4s\n" + "ldr x20, [x15, #0x40]\n" + "mov v25.16b, v31.16b\n fmla v25.4s, v4.4s, v9.4s\n" + "mov v24.16b, v31.16b\n fmla v24.4s, v3.4s, v9.4s\n" + "ldr x25, [x15, #0x50]\n" + "ldr x24, [x15, #0x58]\n" + "fmla v27.4s, v2.4s, v11.4s\n" + "ldr q17, [x23, x14]\n" + "mov v23.16b, v31.16b\n fmla v23.4s, v2.4s, v9.4s\n" + "ldr x23, [x15, #0x60]\n" + "fmla v29.4s, v5.4s, v13.4s\n" + "fmla v28.4s, v6.4s, v17.4s\n" + "ldr x12, [x15, #0x70]\n" + "ldr x11, [x15, #0x88]\n" + "mov v22.16b, v31.16b\n fmla v22.4s, v0.4s, v9.4s\n" + "fmla v27.4s, v3.4s, v13.4s\n" + "ldr x10, [x17, #0x0]\n" + "add x13, x13, #0x10\n" "fmla v26.4s, v2.4s, v13.4s\n" - "fmla v27.4s, v1.4s, v13.4s\n" - "ldr x22, [x16, #0x8]\n" - "ldr x21, [x16, #0x10]\n" - "fmla v28.4s, v0.4s, v13.4s\n" - "ldr q13, [x25, x13]\n" - "fmla v29.4s, v6.4s, v12.4s\n" - "ldr q12, [x24, x13]\n" - "mov v30.16b, v16.16b\n fmla v30.4s, v1.4s, v9.4s\n" - "fmla v23.4s, v7.4s, v11.4s\n" - "ldr x24, [x14, #0x68]\n" - "ldr x25, [x14, #0x78]\n" + "fmla v25.4s, v1.4s, v13.4s\n" + "ldr x9, [x17, #0x8]\n" + "ldr x28, [x17, #0x10]\n" "fmla v24.4s, v0.4s, v13.4s\n" - "fmla v31.4s, v8.4s, v12.4s\n" - "ldr q12, [x11, x13]\n" - "ldr x11, [x14, #0x80]\n" - "fmla v26.4s, v4.4s, v11.4s\n" - "fmla v27.4s, v3.4s, v11.4s\n" - "ldr x20, [x16, #0x18]\n" - "fmla v30.4s, v0.4s, v11.4s\n" - "fmla v28.4s, v4.4s, v10.4s\n" - "fmla v29.4s, v1.4s, v11.4s\n" - "ldr q11, [x9, x13]\n" - "fmla v23.4s, v1.4s, v13.4s\n" - "ldr q13, [x28, x13]\n" - "fmla v24.4s, v2.4s, v12.4s\n" - "fmla v25.4s, v1.4s, v12.4s\n" - "ldr q12, [x27, x13]\n" - "ldr x9, [x14, #0x90]\n" - "fmla v27.4s, v5.4s, v10.4s\n" - "fmla v30.4s, v2.4s, v10.4s\n" - "ldr x27, [x14, #0xa0]\n" - "ldr x28, [x14, #0x98]\n" - "fmla v26.4s, v0.4s, v11.4s\n" - "fmla v28.4s, v2.4s, v13.4s\n" - "fmla v24.4s, v8.4s, v10.4s\n" - "fmla v25.4s, v7.4s, v10.4s\n" - "fmla v31.4s, v1.4s, v10.4s\n" - "ldr q10, [x24, x13]\n" - "fmla v29.4s, v3.4s, v12.4s\n" - "ldr x24, [x14, #0xa8]\n" - "fmla v26.4s, v6.4s, v12.4s\n" - "ldr q12, [x11, x13]\n" - "fmla v27.4s, v7.4s, v10.4s\n" - "ldr x11, [x14, #0xc0]\n" - "fmla v28.4s, v6.4s, v10.4s\n" - "fmla v30.4s, v4.4s, v10.4s\n" - "fmla v23.4s, v3.4s, v11.4s\n" - "ldr q11, [x26, x13]\n" - "fmla v25.4s, v5.4s, v13.4s\n" - "ldr q13, [x25, x13]\n" - "fmla v29.4s, v5.4s, v10.4s\n" - "fmla v31.4s, v3.4s, v10.4s\n" - "ldr x26, [x14, #0xb0]\n" - "ldr x25, [x14, #0xb8]\n" - "fmla v26.4s, v8.4s, v10.4s\n" - "fmla v28.4s, v8.4s, v11.4s\n" - "fmla v30.4s, v6.4s, v13.4s\n" - "fmla v24.4s, v3.4s, v12.4s\n" - "fmla v27.4s, v0.4s, v12.4s\n" - "fmla v31.4s, v5.4s, v11.4s\n" - "ldr q11, [x10, x13]\n" - "fmla v29.4s, v7.4s, v13.4s\n" - "ldr q13, [x9, x13]\n" - "fmla v23.4s, v4.4s, v12.4s\n" - "fmla v26.4s, v1.4s, v12.4s\n" - "ldr q12, [x28, x13]\n" - "fmla v24.4s, v5.4s, v11.4s\n" - "fmla v25.4s, v4.4s, v11.4s\n" - "fmla v27.4s, v2.4s, v11.4s\n" - "fmla v28.4s, v1.4s, v11.4s\n" - "ldr q11, [x27, x13]\n" - "fmla v30.4s, v8.4s, v13.4s\n" - "fmla v31.4s, v7.4s, v13.4s\n" - "ldr q13, [x24, x13]\n" - "fmla v23.4s, v2.4s, v11.4s\n" - "fmla v26.4s, v7.4s, v12.4s\n" - "fmla v27.4s, v6.4s, v12.4s\n" - "fmla v29.4s, v4.4s, v12.4s\n" - "fmla v30.4s, v3.4s, v12.4s\n" - "ldr q12, [x26, x13]\n" - "fmla v31.4s, v4.4s, v13.4s\n" - "fmla v24.4s, v1.4s, v11.4s\n" - "fmax v24.4s, v24.4s, v18.4s\n" - "fmla v25.4s, v0.4s, v11.4s\n" - "ldr q11, [x25, x13]\n" + "ldr q18, [x22, x14]\n" "fmla v23.4s, v6.4s, v12.4s\n" - "fmax v23.4s, v23.4s, v18.4s\n" - "fmla v28.4s, v7.4s, v13.4s\n" - "fmla v30.4s, v5.4s, v13.4s\n" - "fmin v23.4s, v23.4s, v17.4s\n" - "str q23, [x23, x12]\n" - "fmla v29.4s, v0.4s, v12.4s\n" - "fmla v31.4s, v2.4s, v11.4s\n" - "ldr x23, [x16, #0x20]\n" - "fmin v24.4s, v24.4s, v17.4s\n" - "fmla v27.4s, v8.4s, v13.4s\n" - "ldr q13, [x11, x13]\n" - "fmla v26.4s, v3.4s, v12.4s\n" - "fmax v26.4s, v26.4s, v18.4s\n" - "fmla v25.4s, v8.4s, v11.4s\n" - "fmla v28.4s, v5.4s, v11.4s\n" - "fmax v25.4s, v25.4s, v18.4s\n" - "str q24, [x22, x12]\n" - "fmla v29.4s, v8.4s, v13.4s\n" - "fmla v30.4s, v7.4s, v13.4s\n" - "fmax v27.4s, v27.4s, v18.4s\n" - "ldr x22, [x16, #0x28]\n" - "fmla v31.4s, v6.4s, v13.4s\n" - "fmin v25.4s, v25.4s, v17.4s\n" - "str q25, [x21, x12]\n" - "ldr x21, [x16, #0x30]\n" - "fmin v26.4s, v26.4s, v17.4s\n" - "fmin v27.4s, v27.4s, v17.4s\n" - "str q26, [x20, x12]\n" - "ldr x20, [x16, #0x38]\n" - "fmax v28.4s, v28.4s, v18.4s\n" - "fmax v29.4s, v29.4s, v18.4s\n" - "str q27, [x23, x12]\n" - "ldr x23, [x16, #0x40]\n" - "fmax v30.4s, v30.4s, v18.4s\n" - "fmax v31.4s, v31.4s, v18.4s\n" - "add x13, x13, #0x10\n" - "fmin v28.4s, v28.4s, v17.4s\n" - "fmin v29.4s, v29.4s, v17.4s\n" - "str q28, [x22, x12]\n" - "fmin v30.4s, v30.4s, v17.4s\n" - "fmin v31.4s, v31.4s, v17.4s\n" - "str q29, [x21, x12]\n" - "str q30, [x20, x12]\n" - "str q31, [x23, x12]\n" + "ldr q16, [x21, x14]\n" + "mov v21.16b, v31.16b\n fmla v21.4s, v1.4s, v9.4s\n" + "fmla v29.4s, v7.4s, v17.4s\n" + "ldr x22, [x15, #0x68]\n" + "ldr x21, [x15, #0x78]\n" + "fmla v28.4s, v0.4s, v18.4s\n" + "fmla v22.4s, v8.4s, v16.4s\n" + "ldr q16, [x20, x14]\n" + "ldr x20, [x15, #0x80]\n" + "fmla v26.4s, v4.4s, v17.4s\n" + "fmla v25.4s, v3.4s, v17.4s\n" + "ldr x27, [x17, #0x18]\n" + "fmla v21.4s, v0.4s, v17.4s\n" + "fmla v24.4s, v4.4s, v19.4s\n" + "fmla v23.4s, v1.4s, v17.4s\n" + "ldr q17, [x25, x14]\n" + "fmla v29.4s, v1.4s, v18.4s\n" + "ldr q20, [x24, x14]\n" + "fmla v28.4s, v2.4s, v16.4s\n" + "fmla v27.4s, v1.4s, v16.4s\n" + "ldr q16, [x23, x14]\n" + "ldr x26, [x15, #0x90]\n" + "fmla v25.4s, v5.4s, v19.4s\n" + "fmla v21.4s, v2.4s, v19.4s\n" + "ldr x25, [x15, #0xa0]\n" + "ldr x24, [x15, #0x98]\n" + "fmla v26.4s, v0.4s, v17.4s\n" + "fmla v24.4s, v2.4s, v20.4s\n" + "fmla v28.4s, v8.4s, v19.4s\n" + "fmla v27.4s, v7.4s, v19.4s\n" + "fmla v22.4s, v1.4s, v19.4s\n" + "ldr q19, [x22, x14]\n" + "fmla v23.4s, v3.4s, v16.4s\n" + "ldr x23, [x15, #0xa8]\n" + "fmla v26.4s, v6.4s, v16.4s\n" + "ldr q18, [x20, x14]\n" + "fmla v25.4s, v7.4s, v19.4s\n" + "ldr x22, [x15, #0xc0]\n" + "fmla v24.4s, v6.4s, v19.4s\n" + "fmla v21.4s, v4.4s, v19.4s\n" + "fmla v29.4s, v3.4s, v17.4s\n" + "ldr q17, [x12, x14]\n" + "fmla v27.4s, v5.4s, v20.4s\n" + "ldr q16, [x21, x14]\n" + "fmla v23.4s, v5.4s, v19.4s\n" + "fmla v22.4s, v3.4s, v19.4s\n" + "ldr x21, [x15, #0xb0]\n" + "ldr x20, [x15, #0xb8]\n" + "fmla v26.4s, v8.4s, v19.4s\n" + "fmla v24.4s, v8.4s, v17.4s\n" + "fmla v21.4s, v6.4s, v16.4s\n" + "fmla v28.4s, v3.4s, v18.4s\n" + "fmla v25.4s, v0.4s, v18.4s\n" + "fmla v22.4s, v5.4s, v17.4s\n" + "ldr q17, [x11, x14]\n" + "fmla v23.4s, v7.4s, v16.4s\n" + "ldr q16, [x26, x14]\n" + "fmla v29.4s, v4.4s, v18.4s\n" + "fmla v26.4s, v1.4s, v18.4s\n" + "ldr q18, [x24, x14]\n" + "fmla v28.4s, v5.4s, v17.4s\n" + "fmla v27.4s, v4.4s, v17.4s\n" + "fmla v25.4s, v2.4s, v17.4s\n" + "fmla v24.4s, v1.4s, v17.4s\n" + "ldr q17, [x25, x14]\n" + "fmla v21.4s, v8.4s, v16.4s\n" + "fmla v22.4s, v7.4s, v16.4s\n" + "ldr q16, [x23, x14]\n" + "fmla v29.4s, v2.4s, v17.4s\n" + "fmla v26.4s, v7.4s, v18.4s\n" + "fmla v25.4s, v6.4s, v18.4s\n" + "fmla v23.4s, v4.4s, v18.4s\n" + "fmla v21.4s, v3.4s, v18.4s\n" + "ldr q18, [x21, x14]\n" + "fmla v22.4s, v4.4s, v16.4s\n" + "fmla v28.4s, v1.4s, v17.4s\n" + "fmax v28.4s, v28.4s, v15.4s\n" + "fmla v27.4s, v0.4s, v17.4s\n" + "ldr q17, [x20, x14]\n" + "fmla v29.4s, v6.4s, v18.4s\n" + "fmax v29.4s, v29.4s, v15.4s\n" + "fmla v24.4s, v7.4s, v16.4s\n" + "fmla v21.4s, v5.4s, v16.4s\n" + "fmin v29.4s, v29.4s, v14.4s\n" + "str q29, [x10, x13]\n" + "fmla v23.4s, v0.4s, v18.4s\n" + "fmla v22.4s, v2.4s, v17.4s\n" + "ldr x20, [x17, #0x20]\n" + "fmin v28.4s, v28.4s, v14.4s\n" + "fmla v25.4s, v8.4s, v16.4s\n" + "ldr q16, [x22, x14]\n" + "fmla v26.4s, v3.4s, v18.4s\n" + "fmax v26.4s, v26.4s, v15.4s\n" + "fmla v27.4s, v8.4s, v17.4s\n" + "fmla v24.4s, v5.4s, v17.4s\n" + "fmax v27.4s, v27.4s, v15.4s\n" + "str q28, [x9, x13]\n" + "fmla v23.4s, v8.4s, v16.4s\n" + "fmla v21.4s, v7.4s, v16.4s\n" + "fmax v25.4s, v25.4s, v15.4s\n" + "ldr x23, [x17, #0x28]\n" + "fmla v22.4s, v6.4s, v16.4s\n" + "fmin v27.4s, v27.4s, v14.4s\n" + "str q27, [x28, x13]\n" + "ldr x22, [x17, #0x30]\n" + "fmin v26.4s, v26.4s, v14.4s\n" + "fmin v25.4s, v25.4s, v14.4s\n" + "str q26, [x27, x13]\n" + "ldr x21, [x17, #0x38]\n" + "fmax v24.4s, v24.4s, v15.4s\n" + "fmax v23.4s, v23.4s, v15.4s\n" + "str q25, [x20, x13]\n" + "ldr x20, [x17, #0x40]\n" + "fmax v21.4s, v21.4s, v15.4s\n" + "fmax v22.4s, v22.4s, v15.4s\n" + "add x14, x14, #0x10\n" + "fmin v24.4s, v24.4s, v14.4s\n" + "fmin v23.4s, v23.4s, v14.4s\n" + "str q24, [x23, x13]\n" + "fmin v21.4s, v21.4s, v14.4s\n" + "fmin v22.4s, v22.4s, v14.4s\n" + "str q23, [x22, x13]\n" + "str q21, [x21, x13]\n" + "str q22, [x20, x13]\n" "3:" // Oddments "tst %x[n_channels], #0x3\n" "beq 48f\n" - "ldr q16, [x15, #0x0]\n" - "ldr q0, [x15, #0x10]\n" - "mov x12, x13\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 x24, [x14, #0x0]\n" - "ldr x23, [x14, #0x8]\n" - "add x24, x24, x13\n" - "add x23, x23, x13\n" - "ldr x22, [x14, #0x10]\n" - "ldr x21, [x14, #0x18]\n" - "add x22, x22, x13\n" - "add x21, x21, x13\n" - "ldr x20, [x14, #0x20]\n" - "add x20, x20, x13\n" + "ldr q31, [x16, #0x0]\n" + "ldr q0, [x16, #0x10]\n" + "mov x13, x14\n" + "ldr q1, [x16, #0x20]\n" + "ldr q2, [x16, #0x30]\n" + "ldr q3, [x16, #0x40]\n" + "ldr q4, [x16, #0x50]\n" + "ldr q5, [x16, #0x60]\n" + "ldr q6, [x16, #0x70]\n" + "ldr q7, [x16, #0x80]\n" + "ldr q8, [x16, #0x90]\n" + "ldr x24, [x15, #0x0]\n" + "ldr x23, [x15, #0x8]\n" + "add x24, x24, x14\n" + "add x23, x23, x14\n" + "ldr x22, [x15, #0x10]\n" + "ldr x21, [x15, #0x18]\n" + "add x22, x22, x14\n" + "add x21, x21, x14\n" + "ldr x20, [x15, #0x20]\n" + "add x20, x20, x14\n" "tbz %x[n_channels], #1, 4f\n" "ld1 { v9.d }[0], [x24], #0x8\n" "ld1 { v10.d }[0], [x23], #0x8\n" @@ -506,19 +506,19 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "ld1 { v12.s }[0], [x21], #0x4\n" "ld1 { v13.s }[0], [x20], #0x4\n" "5:" // Oddments: Load inputs (2, 2), (0, 0), (0, 4), (4, 0), (1, 2): Bit 1: End - "mov v23.16b, v16.16b\n fmla v23.4s, v8.4s, v9.4s\n" - "mov v25.16b, v16.16b\n fmla v25.4s, v6.4s, v9.4s\n" - "ldr x20, [x14, #0x28]\n" - "add x20, x20, x13\n" - "mov v24.16b, v16.16b\n fmla v24.4s, v7.4s, v9.4s\n" - "mov v26.16b, v16.16b\n fmla v26.4s, v5.4s, v9.4s\n" - "mov v27.16b, v16.16b\n fmla v27.4s, v4.4s, v9.4s\n" - "mov v28.16b, v16.16b\n fmla v28.4s, v3.4s, v9.4s\n" - "mov v29.16b, v16.16b\n fmla v29.4s, v2.4s, v9.4s\n" + "mov v23.16b, v31.16b\n fmla v23.4s, v8.4s, v9.4s\n" + "mov v25.16b, v31.16b\n fmla v25.4s, v6.4s, v9.4s\n" + "ldr x20, [x15, #0x28]\n" + "add x20, x20, x14\n" + "mov v24.16b, v31.16b\n fmla v24.4s, v7.4s, v9.4s\n" + "mov v26.16b, v31.16b\n fmla v26.4s, v5.4s, v9.4s\n" + "mov v27.16b, v31.16b\n fmla v27.4s, v4.4s, v9.4s\n" + "mov v28.16b, v31.16b\n fmla v28.4s, v3.4s, v9.4s\n" + "mov v29.16b, v31.16b\n fmla v29.4s, v2.4s, v9.4s\n" "fmla v23.4s, v0.4s, v10.4s\n" "fmla v25.4s, v2.4s, v11.4s\n" - "mov v30.16b, v16.16b\n fmla v30.4s, v1.4s, v9.4s\n" - "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v9.4s\n" + "mov v30.16b, v31.16b\n fmla v30.4s, v1.4s, v9.4s\n" + "fmla v31.4s, v0.4s, v9.4s\n" "fmla v29.4s, v6.4s, v12.4s\n" "fmla v23.4s, v5.4s, v13.4s\n" "fmla v24.4s, v4.4s, v13.4s\n" @@ -534,9 +534,9 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "6:" // Oddments: Load input (4, 4): Bit 1: Unset "ld1 { v12.s }[0], [x20], #0x4\n" "7:" // Oddments: Load input (4, 4): Bit 1: End - "ldr x20, [x14, #0x30]\n" + "ldr x20, [x15, #0x30]\n" "fmla v31.4s, v8.4s, v12.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "tbz %x[n_channels], #1, 8f\n" "ld1 { v11.d }[0], [x20], #0x8\n" "tbz %x[n_channels], #0, 9f\n" @@ -545,10 +545,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "8:" // Oddments: Load input (2, 1): Bit 1: Unset "ld1 { v11.s }[0], [x20], #0x4\n" "9:" // Oddments: Load input (2, 1): Bit 1: End - "ldr x20, [x14, #0x38]\n" + "ldr x20, [x15, #0x38]\n" "fmla v23.4s, v7.4s, v11.4s\n" "fmla v24.4s, v6.4s, v11.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "fmla v26.4s, v4.4s, v11.4s\n" "fmla v27.4s, v3.4s, v11.4s\n" "fmla v29.4s, v1.4s, v11.4s\n" @@ -561,10 +561,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "10:" // Oddments: Load input (0, 1): Bit 1: Unset "ld1 { v13.s }[0], [x20], #0x4\n" "11:" // Oddments: Load input (0, 1): Bit 1: End - "ldr x20, [x14, #0x40]\n" + "ldr x20, [x15, #0x40]\n" "fmla v23.4s, v1.4s, v13.4s\n" "fmla v24.4s, v0.4s, v13.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "tbz %x[n_channels], #1, 12f\n" "ld1 { v12.d }[0], [x20], #0x8\n" "tbz %x[n_channels], #0, 13f\n" @@ -573,10 +573,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "12:" // Oddments: Load input (0, 3): Bit 1: Unset "ld1 { v12.s }[0], [x20], #0x4\n" "13:" // Oddments: Load input (0, 3): Bit 1: End - "ldr x20, [x14, #0x48]\n" + "ldr x20, [x15, #0x48]\n" "fmla v24.4s, v2.4s, v12.4s\n" "fmla v25.4s, v1.4s, v12.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "tbz %x[n_channels], #1, 14f\n" "ld1 { v10.d }[0], [x20], #0x8\n" "tbz %x[n_channels], #0, 15f\n" @@ -585,10 +585,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "14:" // Oddments: Load input (2, 3): Bit 1: Unset "ld1 { v10.s }[0], [x20], #0x4\n" "15:" // Oddments: Load input (2, 3): Bit 1: End - "ldr x20, [x14, #0x50]\n" + "ldr x20, [x15, #0x50]\n" "fmla v24.4s, v8.4s, v10.4s\n" "fmla v25.4s, v7.4s, v10.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "fmla v27.4s, v5.4s, v10.4s\n" "fmla v28.4s, v4.4s, v10.4s\n" "fmla v30.4s, v2.4s, v10.4s\n" @@ -601,10 +601,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "16:" // Oddments: Load input (1, 0): Bit 1: Unset "ld1 { v11.s }[0], [x20], #0x4\n" "17:" // Oddments: Load input (1, 0): Bit 1: End - "ldr x20, [x14, #0x58]\n" + "ldr x20, [x15, #0x58]\n" "fmla v23.4s, v3.4s, v11.4s\n" "fmla v26.4s, v0.4s, v11.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "tbz %x[n_channels], #1, 18f\n" "ld1 { v13.d }[0], [x20], #0x8\n" "tbz %x[n_channels], #0, 19f\n" @@ -613,10 +613,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "18:" // Oddments: Load input (1, 4): Bit 1: Unset "ld1 { v13.s }[0], [x20], #0x4\n" "19:" // Oddments: Load input (1, 4): Bit 1: End - "ldr x20, [x14, #0x60]\n" + "ldr x20, [x15, #0x60]\n" "fmla v25.4s, v5.4s, v13.4s\n" "fmla v28.4s, v2.4s, v13.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "tbz %x[n_channels], #1, 20f\n" "ld1 { v12.d }[0], [x20], #0x8\n" "tbz %x[n_channels], #0, 21f\n" @@ -625,10 +625,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "20:" // Oddments: Load input (3, 0): Bit 1: Unset "ld1 { v12.s }[0], [x20], #0x4\n" "21:" // Oddments: Load input (3, 0): Bit 1: End - "ldr x20, [x14, #0x68]\n" + "ldr x20, [x15, #0x68]\n" "fmla v26.4s, v6.4s, v12.4s\n" "fmla v29.4s, v3.4s, v12.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "tbz %x[n_channels], #1, 22f\n" "ld1 { v10.d }[0], [x20], #0x8\n" "tbz %x[n_channels], #0, 23f\n" @@ -637,10 +637,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "22:" // Oddments: Load input (3, 2): Bit 1: Unset "ld1 { v10.s }[0], [x20], #0x4\n" "23:" // Oddments: Load input (3, 2): Bit 1: End - "ldr x20, [x14, #0x70]\n" + "ldr x20, [x15, #0x70]\n" "fmla v26.4s, v8.4s, v10.4s\n" "fmla v27.4s, v7.4s, v10.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "fmla v28.4s, v6.4s, v10.4s\n" "fmla v29.4s, v5.4s, v10.4s\n" "fmla v30.4s, v4.4s, v10.4s\n" @@ -653,10 +653,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "24:" // Oddments: Load input (3, 4): Bit 1: Unset "ld1 { v11.s }[0], [x20], #0x4\n" "25:" // Oddments: Load input (3, 4): Bit 1: End - "ldr x20, [x14, #0x78]\n" + "ldr x20, [x15, #0x78]\n" "fmla v28.4s, v8.4s, v11.4s\n" "fmla v31.4s, v5.4s, v11.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "tbz %x[n_channels], #1, 26f\n" "ld1 { v13.d }[0], [x20], #0x8\n" "tbz %x[n_channels], #0, 27f\n" @@ -665,10 +665,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "26:" // Oddments: Load input (4, 1): Bit 1: Unset "ld1 { v13.s }[0], [x20], #0x4\n" "27:" // Oddments: Load input (4, 1): Bit 1: End - "ldr x20, [x14, #0x80]\n" + "ldr x20, [x15, #0x80]\n" "fmla v29.4s, v7.4s, v13.4s\n" "fmla v30.4s, v6.4s, v13.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "tbz %x[n_channels], #1, 28f\n" "ld1 { v12.d }[0], [x20], #0x8\n" "tbz %x[n_channels], #0, 29f\n" @@ -677,10 +677,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "28:" // Oddments: Load input (1, 1): Bit 1: Unset "ld1 { v12.s }[0], [x20], #0x4\n" "29:" // Oddments: Load input (1, 1): Bit 1: End - "ldr x20, [x14, #0x88]\n" + "ldr x20, [x15, #0x88]\n" "fmla v23.4s, v4.4s, v12.4s\n" "fmla v24.4s, v3.4s, v12.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "fmla v26.4s, v1.4s, v12.4s\n" "fmla v27.4s, v0.4s, v12.4s\n" "tbz %x[n_channels], #1, 30f\n" @@ -691,10 +691,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "30:" // Oddments: Load input (1, 3): Bit 1: Unset "ld1 { v11.s }[0], [x20], #0x4\n" "31:" // Oddments: Load input (1, 3): Bit 1: End - "ldr x20, [x14, #0x90]\n" + "ldr x20, [x15, #0x90]\n" "fmla v24.4s, v5.4s, v11.4s\n" "fmla v25.4s, v4.4s, v11.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "fmla v27.4s, v2.4s, v11.4s\n" "fmla v28.4s, v1.4s, v11.4s\n" "tbz %x[n_channels], #1, 32f\n" @@ -705,10 +705,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "32:" // Oddments: Load input (4, 3): Bit 1: Unset "ld1 { v13.s }[0], [x20], #0x4\n" "33:" // Oddments: Load input (4, 3): Bit 1: End - "ldr x20, [x14, #0x98]\n" + "ldr x20, [x15, #0x98]\n" "fmla v30.4s, v8.4s, v13.4s\n" "fmla v31.4s, v7.4s, v13.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "tbz %x[n_channels], #1, 34f\n" "ld1 { v12.d }[0], [x20], #0x8\n" "tbz %x[n_channels], #0, 35f\n" @@ -717,10 +717,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "34:" // Oddments: Load input (3, 1): Bit 1: Unset "ld1 { v12.s }[0], [x20], #0x4\n" "35:" // Oddments: Load input (3, 1): Bit 1: End - "ldr x20, [x14, #0xa0]\n" + "ldr x20, [x15, #0xa0]\n" "fmla v26.4s, v7.4s, v12.4s\n" "fmla v27.4s, v6.4s, v12.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "fmla v29.4s, v4.4s, v12.4s\n" "fmla v30.4s, v3.4s, v12.4s\n" "tbz %x[n_channels], #1, 36f\n" @@ -731,10 +731,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "36:" // Oddments: Load input (0, 2): Bit 1: Unset "ld1 { v11.s }[0], [x20], #0x4\n" "37:" // Oddments: Load input (0, 2): Bit 1: End - "ldr x20, [x14, #0xa8]\n" + "ldr x20, [x15, #0xa8]\n" "fmla v23.4s, v2.4s, v11.4s\n" "fmla v24.4s, v1.4s, v11.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "fmla v25.4s, v0.4s, v11.4s\n" "tbz %x[n_channels], #1, 38f\n" "ld1 { v13.d }[0], [x20], #0x8\n" @@ -744,10 +744,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "38:" // Oddments: Load input (3, 3): Bit 1: Unset "ld1 { v13.s }[0], [x20], #0x4\n" "39:" // Oddments: Load input (3, 3): Bit 1: End - "ldr x20, [x14, #0xb0]\n" + "ldr x20, [x15, #0xb0]\n" "fmla v27.4s, v8.4s, v13.4s\n" "fmla v28.4s, v7.4s, v13.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "fmla v30.4s, v5.4s, v13.4s\n" "fmla v31.4s, v4.4s, v13.4s\n" "tbz %x[n_channels], #1, 40f\n" @@ -758,10 +758,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "40:" // Oddments: Load input (2, 0): Bit 1: Unset "ld1 { v12.s }[0], [x20], #0x4\n" "41:" // Oddments: Load input (2, 0): Bit 1: End - "ldr x20, [x14, #0xb8]\n" + "ldr x20, [x15, #0xb8]\n" "fmla v23.4s, v6.4s, v12.4s\n" "fmla v26.4s, v3.4s, v12.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "fmla v29.4s, v0.4s, v12.4s\n" "tbz %x[n_channels], #1, 42f\n" "ld1 { v11.d }[0], [x20], #0x8\n" @@ -771,10 +771,10 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "42:" // Oddments: Load input (2, 4): Bit 1: Unset "ld1 { v11.s }[0], [x20], #0x4\n" "43:" // Oddments: Load input (2, 4): Bit 1: End - "ldr x20, [x14, #0xc0]\n" + "ldr x20, [x15, #0xc0]\n" "fmla v25.4s, v8.4s, v11.4s\n" "fmla v28.4s, v5.4s, v11.4s\n" - "add x20, x20, x13\n" + "add x20, x20, x14\n" "fmla v31.4s, v2.4s, v11.4s\n" "tbz %x[n_channels], #1, 44f\n" "ld1 { v13.d }[0], [x20], #0x8\n" @@ -786,120 +786,120 @@ void a64_fp32_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl( "45:" // Oddments: Load input (4, 2): Bit 1: End "fmla v29.4s, v8.4s, v13.4s\n" "fmla v30.4s, v7.4s, v13.4s\n" - "fmax v23.4s, v23.4s, v18.4s\n" + "fmax v23.4s, v23.4s, v15.4s\n" "fmla v31.4s, v6.4s, v13.4s\n" - "fmax v24.4s, v24.4s, v18.4s\n" - "fmax v25.4s, v25.4s, v18.4s\n" - "fmax v26.4s, v26.4s, v18.4s\n" - "fmax v27.4s, v27.4s, v18.4s\n" - "fmax v28.4s, v28.4s, v18.4s\n" - "fmax v29.4s, v29.4s, v18.4s\n" - "fmax v30.4s, v30.4s, v18.4s\n" - "fmax v31.4s, v31.4s, v18.4s\n" - "fmin v23.4s, v23.4s, v17.4s\n" - "fmin v24.4s, v24.4s, v17.4s\n" - "fmin v25.4s, v25.4s, v17.4s\n" - "fmin v26.4s, v26.4s, v17.4s\n" - "fmin v27.4s, v27.4s, v17.4s\n" - "fmin v28.4s, v28.4s, v17.4s\n" - "fmin v29.4s, v29.4s, v17.4s\n" - "fmin v30.4s, v30.4s, v17.4s\n" - "fmin v31.4s, v31.4s, v17.4s\n" + "fmax v24.4s, v24.4s, v15.4s\n" + "fmax v25.4s, v25.4s, v15.4s\n" + "fmax v26.4s, v26.4s, v15.4s\n" + "fmax v27.4s, v27.4s, v15.4s\n" + "fmax v28.4s, v28.4s, v15.4s\n" + "fmax v29.4s, v29.4s, v15.4s\n" + "fmax v30.4s, v30.4s, v15.4s\n" + "fmax v31.4s, v31.4s, v15.4s\n" + "fmin v23.4s, v23.4s, v14.4s\n" + "fmin v24.4s, v24.4s, v14.4s\n" + "fmin v25.4s, v25.4s, v14.4s\n" + "fmin v26.4s, v26.4s, v14.4s\n" + "fmin v27.4s, v27.4s, v14.4s\n" + "fmin v28.4s, v28.4s, v14.4s\n" + "fmin v29.4s, v29.4s, v14.4s\n" + "fmin v30.4s, v30.4s, v14.4s\n" + "fmin v31.4s, v31.4s, v14.4s\n" "tbz %x[n_channels], #1, 46f\n" - "ldr x23, [x16, #0x0]\n" - "add x23, x23, x12\n" - "st1 { v23.d }[0], [x23]\n" - "ldr x22, [x16, #0x8]\n" - "ldr x21, [x16, #0x10]\n" - "ldr x20, [x16, #0x18]\n" - "add x22, x22, x12\n" - "add x21, x21, x12\n" - "ldr x23, [x16, #0x20]\n" - "add x20, x20, x12\n" - "add x23, x23, x12\n" - "st1 { v24.d }[0], [x22]\n" - "st1 { v25.d }[0], [x21]\n" - "ldr x22, [x16, #0x28]\n" - "ldr x21, [x16, #0x30]\n" - "add x22, x22, x12\n" - "st1 { v26.d }[0], [x20]\n" - "ldr x20, [x16, #0x38]\n" - "add x21, x21, x12\n" - "add x20, x20, x12\n" - "st1 { v27.d }[0], [x23]\n" - "ldr x23, [x16, #0x40]\n" - "add x23, x23, x12\n" - "add x12, x12, #0x8\n" - "st1 { v28.d }[0], [x22]\n" - "st1 { v29.d }[0], [x21]\n" - "st1 { v30.d }[0], [x20]\n" - "st1 { v31.d }[0], [x23]\n" + "ldr x20, [x17, #0x0]\n" + "add x20, x20, x13\n" + "st1 { v23.d }[0], [x20]\n" + "ldr x23, [x17, #0x8]\n" + "ldr x22, [x17, #0x10]\n" + "ldr x21, [x17, #0x18]\n" + "add x23, x23, x13\n" + "add x22, x22, x13\n" + "ldr x20, [x17, #0x20]\n" + "add x21, x21, x13\n" + "add x20, x20, x13\n" + "st1 { v24.d }[0], [x23]\n" + "st1 { v25.d }[0], [x22]\n" + "ldr x23, [x17, #0x28]\n" + "ldr x22, [x17, #0x30]\n" + "add x23, x23, x13\n" + "st1 { v26.d }[0], [x21]\n" + "ldr x21, [x17, #0x38]\n" + "add x22, x22, x13\n" + "add x21, x21, x13\n" + "st1 { v27.d }[0], [x20]\n" + "ldr x20, [x17, #0x40]\n" + "add x20, x20, x13\n" + "add x13, x13, #0x8\n" + "st1 { v28.d }[0], [x23]\n" + "st1 { v29.d }[0], [x22]\n" + "st1 { v30.d }[0], [x21]\n" + "st1 { v31.d }[0], [x20]\n" "tbz %x[n_channels], #0, 47f\n" - "ldr x23, [x16, #0x0]\n" - "add x23, x23, x12\n" - "st1 { v23.s }[2], [x23]\n" - "ldr x22, [x16, #0x8]\n" - "ldr x21, [x16, #0x10]\n" - "ldr x20, [x16, #0x18]\n" - "add x22, x22, x12\n" - "add x21, x21, x12\n" - "ldr x23, [x16, #0x20]\n" - "add x20, x20, x12\n" - "add x23, x23, x12\n" - "st1 { v24.s }[2], [x22]\n" - "st1 { v25.s }[2], [x21]\n" - "ldr x22, [x16, #0x28]\n" - "ldr x21, [x16, #0x30]\n" - "add x22, x22, x12\n" - "st1 { v26.s }[2], [x20]\n" - "ldr x20, [x16, #0x38]\n" - "add x21, x21, x12\n" - "add x20, x20, x12\n" - "st1 { v27.s }[2], [x23]\n" - "ldr x23, [x16, #0x40]\n" - "add x23, x23, x12\n" - "st1 { v28.s }[2], [x22]\n" - "st1 { v29.s }[2], [x21]\n" - "st1 { v30.s }[2], [x20]\n" - "st1 { v31.s }[2], [x23]\n" + "ldr x20, [x17, #0x0]\n" + "add x20, x20, x13\n" + "st1 { v23.s }[2], [x20]\n" + "ldr x23, [x17, #0x8]\n" + "ldr x22, [x17, #0x10]\n" + "ldr x21, [x17, #0x18]\n" + "add x23, x23, x13\n" + "add x22, x22, x13\n" + "ldr x20, [x17, #0x20]\n" + "add x21, x21, x13\n" + "add x20, x20, x13\n" + "st1 { v24.s }[2], [x23]\n" + "st1 { v25.s }[2], [x22]\n" + "ldr x23, [x17, #0x28]\n" + "ldr x22, [x17, #0x30]\n" + "add x23, x23, x13\n" + "st1 { v26.s }[2], [x21]\n" + "ldr x21, [x17, #0x38]\n" + "add x22, x22, x13\n" + "add x21, x21, x13\n" + "st1 { v27.s }[2], [x20]\n" + "ldr x20, [x17, #0x40]\n" + "add x20, x20, x13\n" + "st1 { v28.s }[2], [x23]\n" + "st1 { v29.s }[2], [x22]\n" + "st1 { v30.s }[2], [x21]\n" + "st1 { v31.s }[2], [x20]\n" "b 47f\n" "46:" // Oddments: Store: Bit 1: Unset - "ldr x23, [x16, #0x0]\n" - "add x23, x23, x12\n" - "st1 { v23.s }[0], [x23]\n" - "ldr x22, [x16, #0x8]\n" - "ldr x21, [x16, #0x10]\n" - "ldr x20, [x16, #0x18]\n" - "add x22, x22, x12\n" - "add x21, x21, x12\n" - "ldr x23, [x16, #0x20]\n" - "add x20, x20, x12\n" - "add x23, x23, x12\n" - "st1 { v24.s }[0], [x22]\n" - "st1 { v25.s }[0], [x21]\n" - "ldr x22, [x16, #0x28]\n" - "ldr x21, [x16, #0x30]\n" - "add x22, x22, x12\n" - "st1 { v26.s }[0], [x20]\n" - "ldr x20, [x16, #0x38]\n" - "add x21, x21, x12\n" - "add x20, x20, x12\n" - "st1 { v27.s }[0], [x23]\n" - "ldr x23, [x16, #0x40]\n" - "add x23, x23, x12\n" - "st1 { v28.s }[0], [x22]\n" - "st1 { v29.s }[0], [x21]\n" - "st1 { v30.s }[0], [x20]\n" - "st1 { v31.s }[0], [x23]\n" + "ldr x20, [x17, #0x0]\n" + "add x20, x20, x13\n" + "st1 { v23.s }[0], [x20]\n" + "ldr x23, [x17, #0x8]\n" + "ldr x22, [x17, #0x10]\n" + "ldr x21, [x17, #0x18]\n" + "add x23, x23, x13\n" + "add x22, x22, x13\n" + "ldr x20, [x17, #0x20]\n" + "add x21, x21, x13\n" + "add x20, x20, x13\n" + "st1 { v24.s }[0], [x23]\n" + "st1 { v25.s }[0], [x22]\n" + "ldr x23, [x17, #0x28]\n" + "ldr x22, [x17, #0x30]\n" + "add x23, x23, x13\n" + "st1 { v26.s }[0], [x21]\n" + "ldr x21, [x17, #0x38]\n" + "add x22, x22, x13\n" + "add x21, x21, x13\n" + "st1 { v27.s }[0], [x20]\n" + "ldr x20, [x17, #0x40]\n" + "add x20, x20, x13\n" + "st1 { v28.s }[0], [x23]\n" + "st1 { v29.s }[0], [x22]\n" + "st1 { v30.s }[0], [x21]\n" + "st1 { v31.s }[0], [x20]\n" "47:" // Oddments: Store: Bit 1: End "48:" // End : : [n_channels] "r" ((unsigned long) n_channels), [offsetof_Args_inptrs] "I" (offsetof(Args, inptrs)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_outptrs] "I" (offsetof(Args, outptrs)), [offsetof_args_params] "I" (offsetof(Args, params)), [params_struct] "r" (¶ms_struct) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "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", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } } // namespace depthwise } // namespace arm_conv -#endif // __aarch64__ +#endif // defined(__aarch64__) |