From 7d9a626aaba9837cb82d189a9c4f0bcef58825bb Mon Sep 17 00:00:00 2001 From: Michael Tyler Date: Wed, 1 Feb 2023 16:37:07 +0000 Subject: Update CPU kernels to remove x19 and w19 Resolves: COMPMID-5805 Change-Id: Idf720bbb136474810086f5089c5ed23b3f79835a Signed-off-by: Michael Tyler Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9081 Benchmark: Arm Jenkins Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins Reviewed-by: Gunes Bayir Reviewed-by: Viet-Hoa Do --- .../generic.cpp | 866 ++++++++++----------- 1 file changed, 427 insertions(+), 439 deletions(-) (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp') diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp index cbe3d2cd1c..342a297dd4 100644 --- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp +++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -40,487 +40,475 @@ void a64_s8q_packed_to_nhwc_3x3_s2_with_multiplier_output2x4_dot_depthfirst_impl ) { __asm__ __volatile__( - "movi v5.16b, #0x1\n" - "ldr x22, [%x[inptrs], #0x0]\n" - "add SP, SP, #-0x80\n" - "ushr v5.4s, v5.4s, #0x8\n" - "ldr x20, [%x[inptrs], #0x8]\n" - "add x21, %x[qp], %[offsetof_Requantize32_b_offset]\n" + "ldr q14, [%x[params], #0x0]\n" + "ldr q5, [%x[params], #0x10]\n" + "movi v15.16b, #0x1\n" + "ushr v15.4s, v15.4s, #0x8\n" + "ldr q6, [%x[params], #0x20]\n" + "ldr q7, [%x[params], #0x30]\n" "movi v26.4s, #0x0\n" - "ldr x19, [%x[inptrs], #0x10]\n" - "mov x11, #0x0\n" - "movi v1.4s, #0x0\n" - "ld1 { v15.16b }, [x22]\n" - "mov x10, #0x0\n" - "movi v22.4s, #0x0\n" - "ld1 { v29.16b }, [x20]\n" - "add x9, %x[qp], %[offsetof_Requantize32_c_offset]\n" - "movi v25.4s, #0x0\n" - "ld1 { v0.16b }, [x19]\n" - "add x28, %x[qp], %[offsetof_Requantize32_minval]\n" - "movi v13.4s, #0x0\n" + "movi v27.4s, #0x0\n" + "ldr x20, [%x[inptrs], #0x8]\n" + "ld1 { v1.16b }, [x20]\n" + "mov v29.16b, v1.16b\n" + "mov v16.16b, v1.16b\n" + "ldr x20, [%x[inptrs], #0x10]\n" + "ld1 { v2.16b }, [x20]\n" + "mov v28.16b, v1.16b\n" + "mov v22.16b, v2.16b\n" + "ldr x20, [%x[inptrs], #0x20]\n" + "ld1 { v4.16b }, [x20]\n" + "mov v31.16b, v2.16b\n" + "mov v30.16b, v2.16b\n" + "ldr x20, [%x[inptrs], #0x0]\n" + "ld1 { v0.16b }, [x20]\n" + "mov v23.16b, v4.16b\n" + "mov v21.16b, v4.16b\n" "ldr x20, [%x[inptrs], #0x18]\n" - "add x27, %x[qp], %[offsetof_Requantize32_maxval]\n" - "mov v20.16b, v15.16b\n" - "ldr x19, [%x[inptrs], #0x20]\n" + "ld1 { v3.16b }, [x20]\n" + "mov v20.16b, v4.16b\n" + "ext v29.16b, v29.16b, v29.16b, #0x2\n" + "ext v16.16b, v16.16b, v16.16b, #0x4\n" + "ext v28.16b, v28.16b, v28.16b, #0x6\n" + "add x20, %x[qp], %[offsetof_Requantize32_b_offset]\n" + "ld1r { v13.4s }, [x20]\n" + "ext v22.16b, v22.16b, v22.16b, #0x2\n" + "ext v31.16b, v31.16b, v31.16b, #0x4\n" + "add x20, %x[qp], %[offsetof_Requantize32_c_offset]\n" + "ld1r { v12.4s }, [x20]\n" + "ext v30.16b, v30.16b, v30.16b, #0x6\n" + "ext v23.16b, v23.16b, v23.16b, #0x2\n" + "add x20, %x[qp], %[offsetof_Requantize32_minval]\n" + "ld1r { v11.4s }, [x20]\n" + "ext v21.16b, v21.16b, v21.16b, #0x4\n" + "ext v20.16b, v20.16b, v20.16b, #0x6\n" + "add x20, %x[qp], %[offsetof_Requantize32_maxval]\n" + "ld1r { v10.4s }, [x20]\n" + "mov v25.16b, v0.16b\n" + "mov v19.16b, v0.16b\n" "cmp %x[n_channels], #0x4\n" - "ext v20.16b, v20.16b, v20.16b, #0x2\n" - "ld1r { v4.4s }, [x21]\n" - "mov v17.16b, v15.16b\n" - "ld1 { v2.16b }, [x20]\n" - "ext v17.16b, v17.16b, v17.16b, #0x4\n" - "ld1 { v7.16b }, [x19]\n" - "mov v23.16b, v15.16b\n" - "ldp x26, x25, [%x[outptrs], #0x0]\n" - "ext v23.16b, v23.16b, v23.16b, #0x6\n" - "ldp x24, x23, [%x[outptrs], #0x10]\n" - "mov v18.16b, v29.16b\n" - "ldp x22, x21, [%x[outptrs], #0x20]\n" - "zip1 v15.4s, v15.4s, v17.4s\n" - "ldp x20, x19, [%x[outptrs], #0x30]\n" - "ext v18.16b, v18.16b, v18.16b, #0x2\n" - "ld1r { v14.4s }, [x9]\n" - "zip1 v20.4s, v20.4s, v23.4s\n" - "ld1r { v27.4s }, [x28]\n" - "zip1 v15.4s, v15.4s, v20.4s\n" - "ld1r { v23.4s }, [x27]\n" - "mov v17.16b, v29.16b\n" - "ldr q6, [%x[params], #0x0]\n" - "ext v17.16b, v17.16b, v17.16b, #0x4\n" - "ldr q8, [%x[params], #0x10]\n" - "mov v11.16b, v29.16b\n" - "ldr q9, [%x[params], #0x20]\n" - "ext v11.16b, v11.16b, v11.16b, #0x6\n" - "ldr q10, [%x[params], #0x30]\n" + "mov x9, #0x0\n" + "mov v18.16b, v0.16b\n" + "mov v24.16b, v3.16b\n" + "mov x28, #0x0\n" + "ldp x27, x26, [%x[outptrs], #0x0]\n" + "mov v17.16b, v3.16b\n" + "ext v25.16b, v25.16b, v25.16b, #0x2\n" + "ldp x25, x24, [%x[outptrs], #0x10]\n" + "ldp x23, x22, [%x[outptrs], #0x20]\n" + "ext v19.16b, v19.16b, v19.16b, #0x4\n" + "ext v18.16b, v18.16b, v18.16b, #0x6\n" + "ldp x21, x20, [%x[outptrs], #0x30]\n" "add %x[params], %x[params], #0x40\n" - "zip1 v29.4s, v29.4s, v17.4s\n" - "mov v12.16b, v0.16b\n" - "ext v12.16b, v12.16b, v12.16b, #0x2\n" - "zip1 v18.4s, v18.4s, v11.4s\n" - "zip1 v29.4s, v29.4s, v18.4s\n" - "mov v17.16b, v0.16b\n" - "ext v17.16b, v17.16b, v17.16b, #0x4\n" - "mov v11.16b, v0.16b\n" - "ext v11.16b, v11.16b, v11.16b, #0x6\n" - "mov v18.16b, v2.16b\n" - "zip1 v0.4s, v0.4s, v17.4s\n" - "ext v18.16b, v18.16b, v18.16b, #0x2\n" - "zip1 v12.4s, v12.4s, v11.4s\n" - "zip1 v0.4s, v0.4s, v12.4s\n" - "mov v17.16b, v2.16b\n" + "zip1 v1.4s, v1.4s, v16.4s\n" + "mov v16.16b, v3.16b\n" + "zip1 v29.4s, v29.4s, v28.4s\n" + "zip1 v2.4s, v2.4s, v31.4s\n" + "zip1 v22.4s, v22.4s, v30.4s\n" + "ext v24.16b, v24.16b, v24.16b, #0x2\n" "ext v17.16b, v17.16b, v17.16b, #0x4\n" - "mov v19.16b, v2.16b\n" - "ext v19.16b, v19.16b, v19.16b, #0x6\n" - "mov v28.16b, v7.16b\n" - "zip1 v2.4s, v2.4s, v17.4s\n" - "ext v28.16b, v28.16b, v28.16b, #0x2\n" - "zip1 v18.4s, v18.4s, v19.4s\n" - "zip1 v2.4s, v2.4s, v18.4s\n" - "mov v18.16b, v7.16b\n" - "ext v18.16b, v18.16b, v18.16b, #0x4\n" - "mov v21.16b, v7.16b\n" - "ext v21.16b, v21.16b, v21.16b, #0x6\n" - "movi v30.4s, #0x0\n" - "zip1 v7.4s, v7.4s, v18.4s\n" - "movi v3.4s, #0x0\n" - "zip1 v28.4s, v28.4s, v21.4s\n" - "zip1 v7.4s, v7.4s, v28.4s\n" - "movi v12.4s, #0x0\n" - "movi v11.4s, #0x0\n" - "movi v19.4s, #0x0\n" + "ext v16.16b, v16.16b, v16.16b, #0x6\n" + "zip1 v4.4s, v4.4s, v21.4s\n" + "zip1 v23.4s, v23.4s, v20.4s\n" + "zip1 v0.4s, v0.4s, v19.4s\n" + "zip1 v25.4s, v25.4s, v18.4s\n" + "zip1 v1.4s, v1.4s, v29.4s\n" + "zip1 v2.4s, v2.4s, v22.4s\n" + ".inst 0x4f81e1fa // sdot v26.4s, v15.16b, v1.4b[0]\n" + "zip1 v3.4s, v3.4s, v17.4s\n" + "zip1 v24.4s, v24.4s, v16.4s\n" + ".inst 0x4fa1e1fb // sdot v27.4s, v15.16b, v1.4b[1]\n" + "zip1 v4.4s, v4.4s, v23.4s\n" + "movi v23.4s, #0x0\n" + ".inst 0x4f81e9f7 // sdot v23.4s, v15.16b, v1.4b[2]\n" + "movi v22.4s, #0x0\n" "movi v21.4s, #0x0\n" + ".inst 0x4fa1e9f6 // sdot v22.4s, v15.16b, v1.4b[3]\n" + "movi v20.4s, #0x0\n" + "movi v9.4s, #0x0\n" + ".inst 0x4f82e1f5 // sdot v21.4s, v15.16b, v2.4b[0]\n" + "movi v8.4s, #0x0\n" + "movi v19.4s, #0x0\n" + ".inst 0x4fa2e1f4 // sdot v20.4s, v15.16b, v2.4b[1]\n" + "movi v18.4s, #0x0\n" "movi v17.4s, #0x0\n" + ".inst 0x4f82e9e9 // sdot v9.4s, v15.16b, v2.4b[2]\n" "movi v16.4s, #0x0\n" - "movi v28.4s, #0x0\n" - "movi v18.4s, #0x0\n" - "movi v20.4s, #0x0\n" - "movi v24.4s, #0x0\n" + "zip1 v0.4s, v0.4s, v25.4s\n" + ".inst 0x4fa2e9e8 // sdot v8.4s, v15.16b, v2.4b[3]\n" + "zip1 v3.4s, v3.4s, v24.4s\n" + ".inst 0x4f84e1f3 // sdot v19.4s, v15.16b, v4.4b[0]\n" + ".inst 0x4fa4e1f2 // sdot v18.4s, v15.16b, v4.4b[1]\n" + ".inst 0x4f84e9f1 // sdot v17.4s, v15.16b, v4.4b[2]\n" + ".inst 0x4fa4e9f0 // sdot v16.4s, v15.16b, v4.4b[3]\n" "movi v31.4s, #0x0\n" - ".inst 0x4f8fe0ba // sdot v26.4s, v5.16b, v15.4b[0]\n" - ".inst 0x4fafe0a1 // sdot v1.4s, v5.16b, v15.4b[1]\n" - ".inst 0x4f8fe8b6 // sdot v22.4s, v5.16b, v15.4b[2]\n" - ".inst 0x4fafe8b9 // sdot v25.4s, v5.16b, v15.4b[3]\n" - ".inst 0x4f9de0ad // sdot v13.4s, v5.16b, v29.4b[0]\n" - ".inst 0x4fbde0be // sdot v30.4s, v5.16b, v29.4b[1]\n" - ".inst 0x4f9de8a3 // sdot v3.4s, v5.16b, v29.4b[2]\n" - ".inst 0x4fbde8ac // sdot v12.4s, v5.16b, v29.4b[3]\n" - ".inst 0x4f80e0ab // sdot v11.4s, v5.16b, v0.4b[0]\n" - ".inst 0x4fa0e0b3 // sdot v19.4s, v5.16b, v0.4b[1]\n" - ".inst 0x4f80e8b5 // sdot v21.4s, v5.16b, v0.4b[2]\n" - ".inst 0x4fa0e8b1 // sdot v17.4s, v5.16b, v0.4b[3]\n" - ".inst 0x4f82e0b0 // sdot v16.4s, v5.16b, v2.4b[0]\n" - ".inst 0x4fa2e0bc // sdot v28.4s, v5.16b, v2.4b[1]\n" - ".inst 0x4f82e8b2 // sdot v18.4s, v5.16b, v2.4b[2]\n" - ".inst 0x4fa2e8b4 // sdot v20.4s, v5.16b, v2.4b[3]\n" - ".inst 0x4f87e0b8 // sdot v24.4s, v5.16b, v7.4b[0]\n" - ".inst 0x4fa7e0bf // sdot v31.4s, v5.16b, v7.4b[1]\n" - "mov v26.16b, v26.16b\n" - "mov v1.16b, v1.16b\n" - "mov v22.16b, v22.16b\n" - "mov v25.16b, v25.16b\n" - "add v26.4s, v26.4s, v13.4s\n" - "movi v13.4s, #0x0\n" - ".inst 0x4f87e8ad // sdot v13.4s, v5.16b, v7.4b[2]\n" - "add v1.4s, v1.4s, v30.4s\n" "movi v30.4s, #0x0\n" - ".inst 0x4fa7e8be // sdot v30.4s, v5.16b, v7.4b[3]\n" - "add v22.4s, v22.4s, v3.4s\n" - "add v25.4s, v25.4s, v12.4s\n" - "add v26.4s, v26.4s, v11.4s\n" - "add v1.4s, v1.4s, v19.4s\n" - "add v22.4s, v22.4s, v21.4s\n" - "add v25.4s, v25.4s, v17.4s\n" - "mov v11.16b, v11.16b\n" - "mov v3.16b, v19.16b\n" - "mov v19.16b, v21.16b\n" - "mov v21.16b, v17.16b\n" - "add v11.4s, v11.4s, v16.4s\n" - "add v3.4s, v3.4s, v28.4s\n" - "add v19.4s, v19.4s, v18.4s\n" - "add v21.4s, v21.4s, v20.4s\n" - "add v11.4s, v11.4s, v24.4s\n" - "add v3.4s, v3.4s, v31.4s\n" - "add v19.4s, v19.4s, v13.4s\n" - "add v21.4s, v21.4s, v30.4s\n" - "neg v4.4s, v4.4s\n" - "mul v26.4s, v26.4s, v4.4s\n" - "str q26, [SP, #0x0]\n" - "mul v1.4s, v1.4s, v4.4s\n" - "mul v22.4s, v22.4s, v4.4s\n" - "str q1, [SP, #0x10]\n" - "mul v25.4s, v25.4s, v4.4s\n" - "mul v11.4s, v11.4s, v4.4s\n" - "str q22, [SP, #0x20]\n" - "mul v3.4s, v3.4s, v4.4s\n" - "str q25, [SP, #0x30]\n" - "mul v19.4s, v19.4s, v4.4s\n" - "mul v21.4s, v21.4s, v4.4s\n" - "str q11, [SP, #0x40]\n" - "add v26.4s, v26.4s, v6.4s\n" - "str q3, [SP, #0x50]\n" - "add v1.4s, v1.4s, v6.4s\n" - "str q19, [SP, #0x60]\n" - "add v22.4s, v22.4s, v6.4s\n" - "add v25.4s, v25.4s, v6.4s\n" - "str q21, [SP, #0x70]\n" - "add v11.4s, v11.4s, v6.4s\n" - "add v3.4s, v3.4s, v6.4s\n" - "add v19.4s, v19.4s, v6.4s\n" - "add v21.4s, v21.4s, v6.4s\n" + "movi v29.4s, #0x0\n" + ".inst 0x4f80e1ff // sdot v31.4s, v15.16b, v0.4b[0]\n" + "movi v28.4s, #0x0\n" + ".inst 0x4fa0e1fe // sdot v30.4s, v15.16b, v0.4b[1]\n" + ".inst 0x4f80e9fd // sdot v29.4s, v15.16b, v0.4b[2]\n" + ".inst 0x4fa0e9fc // sdot v28.4s, v15.16b, v0.4b[3]\n" + "add v24.4s, v26.4s, v21.4s\n" + "add v25.4s, v27.4s, v20.4s\n" + "add v26.4s, v23.4s, v9.4s\n" + "add v27.4s, v22.4s, v8.4s\n" + "add v23.4s, v19.4s, v21.4s\n" + "movi v22.4s, #0x0\n" + ".inst 0x4f83e1f6 // sdot v22.4s, v15.16b, v3.4b[0]\n" + "add v21.4s, v18.4s, v20.4s\n" + "movi v20.4s, #0x0\n" + ".inst 0x4fa3e1f4 // sdot v20.4s, v15.16b, v3.4b[1]\n" + "add v19.4s, v17.4s, v9.4s\n" + "movi v18.4s, #0x0\n" + ".inst 0x4f83e9f2 // sdot v18.4s, v15.16b, v3.4b[2]\n" + "add v17.4s, v16.4s, v8.4s\n" + "movi v16.4s, #0x0\n" + ".inst 0x4fa3e9f0 // sdot v16.4s, v15.16b, v3.4b[3]\n" + "add v24.4s, v24.4s, v31.4s\n" + "add v25.4s, v25.4s, v30.4s\n" + "add v26.4s, v26.4s, v29.4s\n" + "add v27.4s, v27.4s, v28.4s\n" + "add v28.4s, v23.4s, v22.4s\n" + "add v29.4s, v21.4s, v20.4s\n" + "add v30.4s, v19.4s, v18.4s\n" + "add v31.4s, v17.4s, v16.4s\n" + "neg v13.4s, v13.4s\n" + "mul v24.4s, v24.4s, v13.4s\n" + "mul v25.4s, v25.4s, v13.4s\n" + "mul v26.4s, v26.4s, v13.4s\n" + "mul v27.4s, v27.4s, v13.4s\n" + "mul v28.4s, v28.4s, v13.4s\n" + "mul v29.4s, v29.4s, v13.4s\n" + "mul v30.4s, v30.4s, v13.4s\n" + "mul v31.4s, v31.4s, v13.4s\n" + "zip1 v19.4s, v24.4s, v26.4s\n" + "zip1 v18.4s, v25.4s, v27.4s\n" + "zip1 v17.4s, v28.4s, v30.4s\n" + "zip1 v16.4s, v29.4s, v31.4s\n" + "zip1 v22.4s, v19.4s, v18.4s\n" + "zip1 v23.4s, v17.4s, v16.4s\n" + "add v24.4s, v24.4s, v14.4s\n" + "add v25.4s, v25.4s, v14.4s\n" + "add v26.4s, v26.4s, v14.4s\n" + "add v27.4s, v27.4s, v14.4s\n" + "add v28.4s, v28.4s, v14.4s\n" + "add v29.4s, v29.4s, v14.4s\n" + "add v30.4s, v30.4s, v14.4s\n" + "add v31.4s, v31.4s, v14.4s\n" "ble 2f\n" "1:" // Loop - ".inst 0x4f8fe11a // sdot v26.4s, v8.16b, v15.4b[0]\n" - "ldr q20, [%x[params], #0x0]\n" - "add x11, x11, #0x10\n" - ".inst 0x4fafe101 // sdot v1.4s, v8.16b, v15.4b[1]\n" - "ldr q4, [%x[params], #0x10]\n" + "ldr q21, [%x[params], #0x0]\n" + "ldr q20, [%x[params], #0x10]\n" + ".inst 0x4f80e0b8 // sdot v24.4s, v5.16b, v0.4b[0]\n" + ".inst 0x4fa0e0b9 // sdot v25.4s, v5.16b, v0.4b[1]\n" + "ldr q14, [%x[params], #0x20]\n" + ".inst 0x4f80e8ba // sdot v26.4s, v5.16b, v0.4b[2]\n" + ".inst 0x4fa0e8bb // sdot v27.4s, v5.16b, v0.4b[3]\n" "sub %x[n_channels], %x[n_channels], #0x4\n" - ".inst 0x4f8fe916 // sdot v22.4s, v8.16b, v15.4b[2]\n" - "ldr q6, [%x[params], #0x20]\n" + ".inst 0x4f81e0d8 // sdot v24.4s, v6.16b, v1.4b[0]\n" + ".inst 0x4fa1e0d9 // sdot v25.4s, v6.16b, v1.4b[1]\n" "cmp %x[n_channels], #0x4\n" - ".inst 0x4fafe919 // sdot v25.4s, v8.16b, v15.4b[3]\n" - ".inst 0x4f80e10b // sdot v11.4s, v8.16b, v0.4b[0]\n" - ".inst 0x4fa0e103 // sdot v3.4s, v8.16b, v0.4b[1]\n" - ".inst 0x4f80e913 // sdot v19.4s, v8.16b, v0.4b[2]\n" - ".inst 0x4fa0e915 // sdot v21.4s, v8.16b, v0.4b[3]\n" - "ldr q8, [%x[params], #0x30]\n" - ".inst 0x4f9de13a // sdot v26.4s, v9.16b, v29.4b[0]\n" - ".inst 0x4fbde121 // sdot v1.4s, v9.16b, v29.4b[1]\n" - ".inst 0x4f9de936 // sdot v22.4s, v9.16b, v29.4b[2]\n" - ".inst 0x4fbde939 // sdot v25.4s, v9.16b, v29.4b[3]\n" - ".inst 0x4f82e12b // sdot v11.4s, v9.16b, v2.4b[0]\n" - ".inst 0x4fa2e123 // sdot v3.4s, v9.16b, v2.4b[1]\n" - ".inst 0x4f82e933 // sdot v19.4s, v9.16b, v2.4b[2]\n" - ".inst 0x4fa2e935 // sdot v21.4s, v9.16b, v2.4b[3]\n" - "ldr q9, [%x[params], #0x40]\n" - ".inst 0x4f80e15a // sdot v26.4s, v10.16b, v0.4b[0]\n" - ".inst 0x4fa0e141 // sdot v1.4s, v10.16b, v0.4b[1]\n" - ".inst 0x4f80e956 // sdot v22.4s, v10.16b, v0.4b[2]\n" - ".inst 0x4fa0e959 // sdot v25.4s, v10.16b, v0.4b[3]\n" - ".inst 0x4f87e14b // sdot v11.4s, v10.16b, v7.4b[0]\n" - ".inst 0x4fa7e143 // sdot v3.4s, v10.16b, v7.4b[1]\n" - ".inst 0x4f87e953 // sdot v19.4s, v10.16b, v7.4b[2]\n" - ".inst 0x4fa7e955 // sdot v21.4s, v10.16b, v7.4b[3]\n" - "ldr q10, [%x[params], #0x50]\n" + "add x9, x9, #0x10\n" + ".inst 0x4f81e8da // sdot v26.4s, v6.16b, v1.4b[2]\n" + ".inst 0x4fa1e8db // sdot v27.4s, v6.16b, v1.4b[3]\n" + ".inst 0x4f82e0bc // sdot v28.4s, v5.16b, v2.4b[0]\n" + ".inst 0x4fa2e0bd // sdot v29.4s, v5.16b, v2.4b[1]\n" + ".inst 0x4f82e8be // sdot v30.4s, v5.16b, v2.4b[2]\n" + ".inst 0x4fa2e8bf // sdot v31.4s, v5.16b, v2.4b[3]\n" + "ldr q5, [%x[params], #0x30]\n" + ".inst 0x4f82e0f8 // sdot v24.4s, v7.16b, v2.4b[0]\n" + ".inst 0x4fa2e0f9 // sdot v25.4s, v7.16b, v2.4b[1]\n" + "sqrdmulh v24.4s, v24.4s, v21.4s\n" + ".inst 0x4f82e8fa // sdot v26.4s, v7.16b, v2.4b[2]\n" + ".inst 0x4fa2e8fb // sdot v27.4s, v7.16b, v2.4b[3]\n" + "sqrdmulh v25.4s, v25.4s, v21.4s\n" + ".inst 0x4f83e0dc // sdot v28.4s, v6.16b, v3.4b[0]\n" + ".inst 0x4fa3e0dd // sdot v29.4s, v6.16b, v3.4b[1]\n" + "sqrdmulh v26.4s, v26.4s, v21.4s\n" + ".inst 0x4f83e8de // sdot v30.4s, v6.16b, v3.4b[2]\n" + ".inst 0x4fa3e8df // sdot v31.4s, v6.16b, v3.4b[3]\n" + "ldr q6, [%x[params], #0x40]\n" + "sqrdmulh v27.4s, v27.4s, v21.4s\n" + ".inst 0x4f84e0fc // sdot v28.4s, v7.16b, v4.4b[0]\n" + ".inst 0x4fa4e0fd // sdot v29.4s, v7.16b, v4.4b[1]\n" + "and v19.16b, v24.16b, v20.16b\n" + ".inst 0x4f84e8fe // sdot v30.4s, v7.16b, v4.4b[2]\n" + ".inst 0x4fa4e8ff // sdot v31.4s, v7.16b, v4.4b[3]\n" + "ldr q7, [%x[params], #0x50]\n" + "and v18.16b, v25.16b, v20.16b\n" + "and v17.16b, v26.16b, v20.16b\n" + "and v16.16b, v27.16b, v20.16b\n" "add %x[params], %x[params], #0x60\n" - "sqrdmulh v26.4s, v26.4s, v20.4s\n" - "sqrdmulh v1.4s, v1.4s, v20.4s\n" - "sqrdmulh v22.4s, v22.4s, v20.4s\n" - "sqrdmulh v25.4s, v25.4s, v20.4s\n" - "sqrdmulh v11.4s, v11.4s, v20.4s\n" - "and v30.16b, v26.16b, v4.16b\n" - "and v17.16b, v1.16b, v4.16b\n" - "and v16.16b, v22.16b, v4.16b\n" - "sshr v30.4s, v30.4s, #0x1f\n" + "sshr v19.4s, v19.4s, #0x1f\n" + "sshr v18.4s, v18.4s, #0x1f\n" "sshr v17.4s, v17.4s, #0x1f\n" "sshr v16.4s, v16.4s, #0x1f\n" - "sqadd v26.4s, v26.4s, v30.4s\n" - "sqadd v1.4s, v1.4s, v17.4s\n" - "sqadd v22.4s, v22.4s, v16.4s\n" - "and v16.16b, v25.16b, v4.16b\n" - "srshl v26.4s, v26.4s, v4.4s\n" - "srshl v1.4s, v1.4s, v4.4s\n" - "srshl v22.4s, v22.4s, v4.4s\n" - "sshr v16.4s, v16.4s, #0x1f\n" - "add v26.4s, v26.4s, v14.4s\n" - "add v1.4s, v1.4s, v14.4s\n" - "add v22.4s, v22.4s, v14.4s\n" - "smin v26.4s, v26.4s, v23.4s\n" - "smin v1.4s, v1.4s, v23.4s\n" - "smin v22.4s, v22.4s, v23.4s\n" - "smax v26.4s, v26.4s, v27.4s\n" - "smax v1.4s, v1.4s, v27.4s\n" - "smax v22.4s, v22.4s, v27.4s\n" - "uzp1 v26.16b, v26.16b, v26.16b\n" - "uzp1 v1.16b, v1.16b, v1.16b\n" - "uzp1 v26.16b, v26.16b, v26.16b\n" - "str s26, [x26, x10]\n" - "uzp1 v1.16b, v1.16b, v1.16b\n" - "uzp1 v22.16b, v22.16b, v22.16b\n" - "ldr q26, [SP, #0x0]\n" - "sqadd v25.4s, v25.4s, v16.4s\n" - "str s1, [x25, x10]\n" - "uzp1 v22.16b, v22.16b, v22.16b\n" - "ldr q1, [SP, #0x10]\n" - "and v16.16b, v11.16b, v4.16b\n" - "str s22, [x24, x10]\n" - "sqrdmulh v3.4s, v3.4s, v20.4s\n" - "ldr q22, [SP, #0x20]\n" - "srshl v25.4s, v25.4s, v4.4s\n" - "sshr v16.4s, v16.4s, #0x1f\n" - "sqrdmulh v19.4s, v19.4s, v20.4s\n" - "and v17.16b, v3.16b, v4.16b\n" - "add v25.4s, v25.4s, v14.4s\n" - "sqadd v11.4s, v11.4s, v16.4s\n" + "sqrdmulh v28.4s, v28.4s, v21.4s\n" + "sqrdmulh v29.4s, v29.4s, v21.4s\n" + "sqrdmulh v30.4s, v30.4s, v21.4s\n" + "sqrdmulh v31.4s, v31.4s, v21.4s\n" + "sqadd v24.4s, v24.4s, v19.4s\n" + "sqadd v25.4s, v25.4s, v18.4s\n" + "sqadd v26.4s, v26.4s, v17.4s\n" + "sqadd v27.4s, v27.4s, v16.4s\n" + "and v19.16b, v28.16b, v20.16b\n" + "and v18.16b, v29.16b, v20.16b\n" + "and v17.16b, v30.16b, v20.16b\n" + "and v16.16b, v31.16b, v20.16b\n" + "sshr v19.4s, v19.4s, #0x1f\n" + "sshr v18.4s, v18.4s, #0x1f\n" "sshr v17.4s, v17.4s, #0x1f\n" - "smin v25.4s, v25.4s, v23.4s\n" - "and v16.16b, v19.16b, v4.16b\n" - "srshl v11.4s, v11.4s, v4.4s\n" - "smax v25.4s, v25.4s, v27.4s\n" - "sqadd v3.4s, v3.4s, v17.4s\n" "sshr v16.4s, v16.4s, #0x1f\n" + "sqadd v28.4s, v28.4s, v19.4s\n" + "sqadd v29.4s, v29.4s, v18.4s\n" + "sqadd v30.4s, v30.4s, v17.4s\n" + "sqadd v31.4s, v31.4s, v16.4s\n" + "srshl v24.4s, v24.4s, v20.4s\n" + "srshl v25.4s, v25.4s, v20.4s\n" + "srshl v26.4s, v26.4s, v20.4s\n" + "srshl v27.4s, v27.4s, v20.4s\n" + "srshl v28.4s, v28.4s, v20.4s\n" + "srshl v29.4s, v29.4s, v20.4s\n" + "srshl v30.4s, v30.4s, v20.4s\n" + "srshl v31.4s, v31.4s, v20.4s\n" + "add v24.4s, v24.4s, v12.4s\n" + "add v25.4s, v25.4s, v12.4s\n" + "add v26.4s, v26.4s, v12.4s\n" + "add v27.4s, v27.4s, v12.4s\n" + "add v28.4s, v28.4s, v12.4s\n" + "add v29.4s, v29.4s, v12.4s\n" + "add v30.4s, v30.4s, v12.4s\n" + "add v31.4s, v31.4s, v12.4s\n" + "smin v24.4s, v24.4s, v10.4s\n" + "smin v25.4s, v25.4s, v10.4s\n" + "smin v26.4s, v26.4s, v10.4s\n" + "smin v27.4s, v27.4s, v10.4s\n" + "smin v28.4s, v28.4s, v10.4s\n" + "smin v29.4s, v29.4s, v10.4s\n" + "smin v30.4s, v30.4s, v10.4s\n" + "smin v31.4s, v31.4s, v10.4s\n" + "smax v24.4s, v24.4s, v11.4s\n" + "smax v25.4s, v25.4s, v11.4s\n" + "smax v26.4s, v26.4s, v11.4s\n" + "smax v27.4s, v27.4s, v11.4s\n" + "smax v28.4s, v28.4s, v11.4s\n" + "smax v29.4s, v29.4s, v11.4s\n" + "smax v30.4s, v30.4s, v11.4s\n" + "smax v31.4s, v31.4s, v11.4s\n" + "uzp1 v24.16b, v24.16b, v24.16b\n" "uzp1 v25.16b, v25.16b, v25.16b\n" - "add v11.4s, v11.4s, v14.4s\n" + "uzp1 v26.16b, v26.16b, v26.16b\n" + "uzp1 v27.16b, v27.16b, v27.16b\n" + "uzp1 v28.16b, v28.16b, v28.16b\n" + "uzp1 v29.16b, v29.16b, v29.16b\n" + "uzp1 v30.16b, v30.16b, v30.16b\n" + "uzp1 v31.16b, v31.16b, v31.16b\n" + "uzp1 v24.16b, v24.16b, v24.16b\n" "uzp1 v25.16b, v25.16b, v25.16b\n" - "str s25, [x23, x10]\n" - "smin v11.4s, v11.4s, v23.4s\n" - "srshl v3.4s, v3.4s, v4.4s\n" - "ldr q25, [SP, #0x30]\n" - "sqadd v19.4s, v19.4s, v16.4s\n" - "sqrdmulh v21.4s, v21.4s, v20.4s\n" - "smax v11.4s, v11.4s, v27.4s\n" - "add v3.4s, v3.4s, v14.4s\n" - "srshl v19.4s, v19.4s, v4.4s\n" - "uzp1 v11.16b, v11.16b, v11.16b\n" - "smin v3.4s, v3.4s, v23.4s\n" - "uzp1 v11.16b, v11.16b, v11.16b\n" - "str s11, [x22, x10]\n" - "smax v3.4s, v3.4s, v27.4s\n" - "add v19.4s, v19.4s, v14.4s\n" - "ldr q11, [SP, #0x40]\n" - "and v16.16b, v21.16b, v4.16b\n" - "add v26.4s, v26.4s, v6.4s\n" - "uzp1 v3.16b, v3.16b, v3.16b\n" - "smin v19.4s, v19.4s, v23.4s\n" - "uzp1 v3.16b, v3.16b, v3.16b\n" - "str s3, [x21, x10]\n" - "smax v19.4s, v19.4s, v27.4s\n" - "sshr v16.4s, v16.4s, #0x1f\n" - "ldr q3, [SP, #0x50]\n" - "add v1.4s, v1.4s, v6.4s\n" - "add v22.4s, v22.4s, v6.4s\n" - "uzp1 v19.16b, v19.16b, v19.16b\n" - "sqadd v21.4s, v21.4s, v16.4s\n" - "uzp1 v19.16b, v19.16b, v19.16b\n" - "str s19, [x20, x10]\n" - "add v25.4s, v25.4s, v6.4s\n" - "add v11.4s, v11.4s, v6.4s\n" - "ldr q19, [SP, #0x60]\n" - "srshl v21.4s, v21.4s, v4.4s\n" - "add v3.4s, v3.4s, v6.4s\n" - "add v21.4s, v21.4s, v14.4s\n" - "add v19.4s, v19.4s, v6.4s\n" - "smin v21.4s, v21.4s, v23.4s\n" - "smax v21.4s, v21.4s, v27.4s\n" - "uzp1 v21.16b, v21.16b, v21.16b\n" - "uzp1 v21.16b, v21.16b, v21.16b\n" - "str s21, [x19, x10]\n" - "add x10, x10, #0x4\n" - "ldr q21, [SP, #0x70]\n" - "add v21.4s, v21.4s, v6.4s\n" + "str s24, [x27, x28]\n" + "uzp1 v26.16b, v26.16b, v26.16b\n" + "uzp1 v27.16b, v27.16b, v27.16b\n" + "str s25, [x26, x28]\n" + "uzp1 v28.16b, v28.16b, v28.16b\n" + "uzp1 v29.16b, v29.16b, v29.16b\n" + "str s26, [x25, x28]\n" + "uzp1 v30.16b, v30.16b, v30.16b\n" + "uzp1 v31.16b, v31.16b, v31.16b\n" + "str s27, [x24, x28]\n" + "str s28, [x23, x28]\n" + "dup v24.4s, v22.s[0]\n" + "dup v25.4s, v22.s[1]\n" + "str s29, [x22, x28]\n" + "dup v26.4s, v22.s[2]\n" + "dup v27.4s, v22.s[3]\n" + "str s30, [x21, x28]\n" + "dup v28.4s, v23.s[0]\n" + "dup v29.4s, v23.s[1]\n" + "str s31, [x20, x28]\n" + "dup v30.4s, v23.s[2]\n" + "dup v31.4s, v23.s[3]\n" + "add x28, x28, #0x4\n" + "add v24.4s, v24.4s, v14.4s\n" + "add v25.4s, v25.4s, v14.4s\n" + "add v26.4s, v26.4s, v14.4s\n" + "add v27.4s, v27.4s, v14.4s\n" + "add v28.4s, v28.4s, v14.4s\n" + "add v29.4s, v29.4s, v14.4s\n" + "add v30.4s, v30.4s, v14.4s\n" + "add v31.4s, v31.4s, v14.4s\n" "bgt 1b\n" "2:" // Tail - ".inst 0x4f8fe11a // sdot v26.4s, v8.16b, v15.4b[0]\n" - "ldr q20, [%x[params], #0x0]\n" - "add x26, x26, x10\n" - ".inst 0x4fafe101 // sdot v1.4s, v8.16b, v15.4b[1]\n" - "ldr q4, [%x[params], #0x10]\n" - "add x25, x25, x10\n" - ".inst 0x4f8fe916 // sdot v22.4s, v8.16b, v15.4b[2]\n" - "add x24, x24, x10\n" - ".inst 0x4fafe919 // sdot v25.4s, v8.16b, v15.4b[3]\n" - "add x23, x23, x10\n" - ".inst 0x4f80e10b // sdot v11.4s, v8.16b, v0.4b[0]\n" - "add x22, x22, x10\n" - ".inst 0x4fa0e103 // sdot v3.4s, v8.16b, v0.4b[1]\n" - "add x21, x21, x10\n" - ".inst 0x4f80e913 // sdot v19.4s, v8.16b, v0.4b[2]\n" - "add x20, x20, x10\n" - ".inst 0x4fa0e915 // sdot v21.4s, v8.16b, v0.4b[3]\n" - "add x19, x19, x10\n" - ".inst 0x4f9de13a // sdot v26.4s, v9.16b, v29.4b[0]\n" + "ldr q21, [%x[params], #0x0]\n" + "ldr q20, [%x[params], #0x10]\n" + ".inst 0x4f80e0b8 // sdot v24.4s, v5.16b, v0.4b[0]\n" + ".inst 0x4fa0e0b9 // sdot v25.4s, v5.16b, v0.4b[1]\n" + ".inst 0x4f80e8ba // sdot v26.4s, v5.16b, v0.4b[2]\n" + ".inst 0x4fa0e8bb // sdot v27.4s, v5.16b, v0.4b[3]\n" "cmp %x[n_channels], #0x4\n" - ".inst 0x4fbde121 // sdot v1.4s, v9.16b, v29.4b[1]\n" + "add x27, x27, x28\n" + ".inst 0x4f81e0d8 // sdot v24.4s, v6.16b, v1.4b[0]\n" + ".inst 0x4fa1e0d9 // sdot v25.4s, v6.16b, v1.4b[1]\n" + "add x26, x26, x28\n" + "add x25, x25, x28\n" + ".inst 0x4f81e8da // sdot v26.4s, v6.16b, v1.4b[2]\n" + ".inst 0x4fa1e8db // sdot v27.4s, v6.16b, v1.4b[3]\n" + "add x24, x24, x28\n" + "add x23, x23, x28\n" + ".inst 0x4f82e0bc // sdot v28.4s, v5.16b, v2.4b[0]\n" + ".inst 0x4fa2e0bd // sdot v29.4s, v5.16b, v2.4b[1]\n" + "add x22, x22, x28\n" + "add x21, x21, x28\n" + ".inst 0x4f82e8be // sdot v30.4s, v5.16b, v2.4b[2]\n" + ".inst 0x4fa2e8bf // sdot v31.4s, v5.16b, v2.4b[3]\n" + "add x20, x20, x28\n" "add %x[params], %x[params], #0x20\n" - ".inst 0x4f9de936 // sdot v22.4s, v9.16b, v29.4b[2]\n" - ".inst 0x4fbde939 // sdot v25.4s, v9.16b, v29.4b[3]\n" - ".inst 0x4f82e12b // sdot v11.4s, v9.16b, v2.4b[0]\n" - ".inst 0x4fa2e123 // sdot v3.4s, v9.16b, v2.4b[1]\n" - ".inst 0x4f82e933 // sdot v19.4s, v9.16b, v2.4b[2]\n" - ".inst 0x4fa2e935 // sdot v21.4s, v9.16b, v2.4b[3]\n" - ".inst 0x4f80e15a // sdot v26.4s, v10.16b, v0.4b[0]\n" - ".inst 0x4fa0e141 // sdot v1.4s, v10.16b, v0.4b[1]\n" - ".inst 0x4f80e956 // sdot v22.4s, v10.16b, v0.4b[2]\n" - ".inst 0x4fa0e959 // sdot v25.4s, v10.16b, v0.4b[3]\n" - ".inst 0x4f87e14b // sdot v11.4s, v10.16b, v7.4b[0]\n" - ".inst 0x4fa7e143 // sdot v3.4s, v10.16b, v7.4b[1]\n" - ".inst 0x4f87e953 // sdot v19.4s, v10.16b, v7.4b[2]\n" - ".inst 0x4fa7e955 // sdot v21.4s, v10.16b, v7.4b[3]\n" - "sqrdmulh v26.4s, v26.4s, v20.4s\n" - "sqrdmulh v1.4s, v1.4s, v20.4s\n" - "sqrdmulh v22.4s, v22.4s, v20.4s\n" - "sqrdmulh v25.4s, v25.4s, v20.4s\n" - "and v30.16b, v26.16b, v4.16b\n" - "and v17.16b, v1.16b, v4.16b\n" - "and v16.16b, v22.16b, v4.16b\n" - "sshr v30.4s, v30.4s, #0x1f\n" + ".inst 0x4f82e0f8 // sdot v24.4s, v7.16b, v2.4b[0]\n" + ".inst 0x4fa2e0f9 // sdot v25.4s, v7.16b, v2.4b[1]\n" + "sqrdmulh v24.4s, v24.4s, v21.4s\n" + ".inst 0x4f82e8fa // sdot v26.4s, v7.16b, v2.4b[2]\n" + ".inst 0x4fa2e8fb // sdot v27.4s, v7.16b, v2.4b[3]\n" + "sqrdmulh v25.4s, v25.4s, v21.4s\n" + ".inst 0x4f83e0dc // sdot v28.4s, v6.16b, v3.4b[0]\n" + ".inst 0x4fa3e0dd // sdot v29.4s, v6.16b, v3.4b[1]\n" + "sqrdmulh v26.4s, v26.4s, v21.4s\n" + ".inst 0x4f83e8de // sdot v30.4s, v6.16b, v3.4b[2]\n" + ".inst 0x4fa3e8df // sdot v31.4s, v6.16b, v3.4b[3]\n" + "sqrdmulh v27.4s, v27.4s, v21.4s\n" + ".inst 0x4f84e0fc // sdot v28.4s, v7.16b, v4.4b[0]\n" + ".inst 0x4fa4e0fd // sdot v29.4s, v7.16b, v4.4b[1]\n" + "and v19.16b, v24.16b, v20.16b\n" + ".inst 0x4f84e8fe // sdot v30.4s, v7.16b, v4.4b[2]\n" + ".inst 0x4fa4e8ff // sdot v31.4s, v7.16b, v4.4b[3]\n" + "and v18.16b, v25.16b, v20.16b\n" + "and v17.16b, v26.16b, v20.16b\n" + "and v16.16b, v27.16b, v20.16b\n" + "sshr v19.4s, v19.4s, #0x1f\n" + "sshr v18.4s, v18.4s, #0x1f\n" "sshr v17.4s, v17.4s, #0x1f\n" "sshr v16.4s, v16.4s, #0x1f\n" - "sqadd v26.4s, v26.4s, v30.4s\n" - "sqadd v1.4s, v1.4s, v17.4s\n" - "sqadd v22.4s, v22.4s, v16.4s\n" - "and v16.16b, v25.16b, v4.16b\n" - "srshl v26.4s, v26.4s, v4.4s\n" - "srshl v1.4s, v1.4s, v4.4s\n" - "srshl v22.4s, v22.4s, v4.4s\n" - "sshr v16.4s, v16.4s, #0x1f\n" - "add v26.4s, v26.4s, v14.4s\n" - "add v1.4s, v1.4s, v14.4s\n" - "add v22.4s, v22.4s, v14.4s\n" - "smin v26.4s, v26.4s, v23.4s\n" - "smin v1.4s, v1.4s, v23.4s\n" - "smin v22.4s, v22.4s, v23.4s\n" - "smax v26.4s, v26.4s, v27.4s\n" - "smax v1.4s, v1.4s, v27.4s\n" - "smax v22.4s, v22.4s, v27.4s\n" - "uzp1 v26.16b, v26.16b, v26.16b\n" - "uzp1 v1.16b, v1.16b, v1.16b\n" - "uzp1 v26.16b, v26.16b, v26.16b\n" - "uzp1 v1.16b, v1.16b, v1.16b\n" - "uzp1 v22.16b, v22.16b, v22.16b\n" - "sqadd v25.4s, v25.4s, v16.4s\n" - "uzp1 v22.16b, v22.16b, v22.16b\n" - "sqrdmulh v11.4s, v11.4s, v20.4s\n" - "sqrdmulh v3.4s, v3.4s, v20.4s\n" - "srshl v25.4s, v25.4s, v4.4s\n" - "sqrdmulh v19.4s, v19.4s, v20.4s\n" - "and v16.16b, v11.16b, v4.16b\n" - "and v17.16b, v3.16b, v4.16b\n" - "add v25.4s, v25.4s, v14.4s\n" - "sshr v16.4s, v16.4s, #0x1f\n" + "sqrdmulh v28.4s, v28.4s, v21.4s\n" + "sqrdmulh v29.4s, v29.4s, v21.4s\n" + "sqrdmulh v30.4s, v30.4s, v21.4s\n" + "sqrdmulh v31.4s, v31.4s, v21.4s\n" + "sqadd v24.4s, v24.4s, v19.4s\n" + "sqadd v25.4s, v25.4s, v18.4s\n" + "sqadd v26.4s, v26.4s, v17.4s\n" + "sqadd v27.4s, v27.4s, v16.4s\n" + "and v19.16b, v28.16b, v20.16b\n" + "and v18.16b, v29.16b, v20.16b\n" + "and v17.16b, v30.16b, v20.16b\n" + "and v16.16b, v31.16b, v20.16b\n" + "sshr v19.4s, v19.4s, #0x1f\n" + "sshr v18.4s, v18.4s, #0x1f\n" "sshr v17.4s, v17.4s, #0x1f\n" - "smin v25.4s, v25.4s, v23.4s\n" - "sqadd v11.4s, v11.4s, v16.4s\n" - "sqadd v3.4s, v3.4s, v17.4s\n" - "smax v25.4s, v25.4s, v27.4s\n" - "and v16.16b, v19.16b, v4.16b\n" - "srshl v11.4s, v11.4s, v4.4s\n" + "sshr v16.4s, v16.4s, #0x1f\n" + "sqadd v28.4s, v28.4s, v19.4s\n" + "sqadd v29.4s, v29.4s, v18.4s\n" + "sqadd v30.4s, v30.4s, v17.4s\n" + "sqadd v31.4s, v31.4s, v16.4s\n" + "srshl v24.4s, v24.4s, v20.4s\n" + "srshl v25.4s, v25.4s, v20.4s\n" + "srshl v26.4s, v26.4s, v20.4s\n" + "srshl v27.4s, v27.4s, v20.4s\n" + "srshl v28.4s, v28.4s, v20.4s\n" + "srshl v29.4s, v29.4s, v20.4s\n" + "srshl v30.4s, v30.4s, v20.4s\n" + "srshl v31.4s, v31.4s, v20.4s\n" + "add v24.4s, v24.4s, v12.4s\n" + "add v25.4s, v25.4s, v12.4s\n" + "add v26.4s, v26.4s, v12.4s\n" + "add v27.4s, v27.4s, v12.4s\n" + "add v28.4s, v28.4s, v12.4s\n" + "add v29.4s, v29.4s, v12.4s\n" + "add v30.4s, v30.4s, v12.4s\n" + "add v31.4s, v31.4s, v12.4s\n" + "smin v24.4s, v24.4s, v10.4s\n" + "smin v25.4s, v25.4s, v10.4s\n" + "smin v26.4s, v26.4s, v10.4s\n" + "smin v27.4s, v27.4s, v10.4s\n" + "smin v28.4s, v28.4s, v10.4s\n" + "smin v29.4s, v29.4s, v10.4s\n" + "smin v30.4s, v30.4s, v10.4s\n" + "smin v31.4s, v31.4s, v10.4s\n" + "smax v24.4s, v24.4s, v11.4s\n" + "smax v25.4s, v25.4s, v11.4s\n" + "smax v26.4s, v26.4s, v11.4s\n" + "smax v27.4s, v27.4s, v11.4s\n" + "smax v28.4s, v28.4s, v11.4s\n" + "smax v29.4s, v29.4s, v11.4s\n" + "smax v30.4s, v30.4s, v11.4s\n" + "smax v31.4s, v31.4s, v11.4s\n" + "uzp1 v24.16b, v24.16b, v24.16b\n" "uzp1 v25.16b, v25.16b, v25.16b\n" - "srshl v3.4s, v3.4s, v4.4s\n" + "uzp1 v26.16b, v26.16b, v26.16b\n" + "uzp1 v27.16b, v27.16b, v27.16b\n" + "uzp1 v28.16b, v28.16b, v28.16b\n" + "uzp1 v29.16b, v29.16b, v29.16b\n" + "uzp1 v30.16b, v30.16b, v30.16b\n" + "uzp1 v31.16b, v31.16b, v31.16b\n" + "uzp1 v24.16b, v24.16b, v24.16b\n" "uzp1 v25.16b, v25.16b, v25.16b\n" - "add v11.4s, v11.4s, v14.4s\n" - "sshr v16.4s, v16.4s, #0x1f\n" - "add v3.4s, v3.4s, v14.4s\n" - "smin v11.4s, v11.4s, v23.4s\n" - "sqadd v19.4s, v19.4s, v16.4s\n" - "smin v3.4s, v3.4s, v23.4s\n" - "smax v11.4s, v11.4s, v27.4s\n" - "sqrdmulh v21.4s, v21.4s, v20.4s\n" - "smax v3.4s, v3.4s, v27.4s\n" - "uzp1 v11.16b, v11.16b, v11.16b\n" - "srshl v19.4s, v19.4s, v4.4s\n" - "uzp1 v11.16b, v11.16b, v11.16b\n" - "uzp1 v3.16b, v3.16b, v3.16b\n" - "and v16.16b, v21.16b, v4.16b\n" - "uzp1 v3.16b, v3.16b, v3.16b\n" - "add v19.4s, v19.4s, v14.4s\n" - "sshr v16.4s, v16.4s, #0x1f\n" - "smin v19.4s, v19.4s, v23.4s\n" - "sqadd v21.4s, v21.4s, v16.4s\n" - "smax v19.4s, v19.4s, v27.4s\n" - "srshl v21.4s, v21.4s, v4.4s\n" - "uzp1 v19.16b, v19.16b, v19.16b\n" - "uzp1 v19.16b, v19.16b, v19.16b\n" - "add v21.4s, v21.4s, v14.4s\n" - "smin v21.4s, v21.4s, v23.4s\n" - "smax v21.4s, v21.4s, v27.4s\n" - "uzp1 v21.16b, v21.16b, v21.16b\n" - "uzp1 v21.16b, v21.16b, v21.16b\n" + "uzp1 v26.16b, v26.16b, v26.16b\n" + "uzp1 v27.16b, v27.16b, v27.16b\n" + "uzp1 v28.16b, v28.16b, v28.16b\n" + "uzp1 v29.16b, v29.16b, v29.16b\n" + "uzp1 v30.16b, v30.16b, v30.16b\n" + "uzp1 v31.16b, v31.16b, v31.16b\n" "blt 3f\n" - "str s26, [x26, #0x0]\n" - "str s1, [x25, #0x0]\n" - "str s22, [x24, #0x0]\n" - "str s25, [x23, #0x0]\n" - "str s11, [x22, #0x0]\n" - "str s3, [x21, #0x0]\n" - "str s19, [x20, #0x0]\n" - "str s21, [x19, #0x0]\n" + "str s24, [x27, #0x0]\n" + "str s25, [x26, #0x0]\n" + "str s26, [x25, #0x0]\n" + "str s27, [x24, #0x0]\n" + "str s28, [x23, #0x0]\n" + "str s29, [x22, #0x0]\n" + "str s30, [x21, #0x0]\n" + "str s31, [x20, #0x0]\n" "b 4f\n" "3:" // Tail: Oddments - "st1 { v26.b }[0], [x26], #0x1\n" "subs %x[n_channels], %x[n_channels], #0x1\n" - "st1 { v1.b }[0], [x25], #0x1\n" - "st1 { v22.b }[0], [x24], #0x1\n" - "st1 { v25.b }[0], [x23], #0x1\n" - "st1 { v11.b }[0], [x22], #0x1\n" - "st1 { v3.b }[0], [x21], #0x1\n" - "st1 { v19.b }[0], [x20], #0x1\n" - "st1 { v21.b }[0], [x19], #0x1\n" + "st1 { v24.b }[0], [x27], #0x1\n" + "st1 { v25.b }[0], [x26], #0x1\n" + "st1 { v26.b }[0], [x25], #0x1\n" + "st1 { v27.b }[0], [x24], #0x1\n" + "st1 { v28.b }[0], [x23], #0x1\n" + "st1 { v29.b }[0], [x22], #0x1\n" + "st1 { v30.b }[0], [x21], #0x1\n" + "st1 { v31.b }[0], [x20], #0x1\n" "beq 4f\n" - "st1 { v26.b }[1], [x26], #0x1\n" "subs %x[n_channels], %x[n_channels], #0x1\n" - "st1 { v1.b }[1], [x25], #0x1\n" - "st1 { v22.b }[1], [x24], #0x1\n" - "st1 { v25.b }[1], [x23], #0x1\n" - "st1 { v11.b }[1], [x22], #0x1\n" - "st1 { v3.b }[1], [x21], #0x1\n" - "st1 { v19.b }[1], [x20], #0x1\n" - "st1 { v21.b }[1], [x19], #0x1\n" + "st1 { v24.b }[1], [x27], #0x1\n" + "st1 { v25.b }[1], [x26], #0x1\n" + "st1 { v26.b }[1], [x25], #0x1\n" + "st1 { v27.b }[1], [x24], #0x1\n" + "st1 { v28.b }[1], [x23], #0x1\n" + "st1 { v29.b }[1], [x22], #0x1\n" + "st1 { v30.b }[1], [x21], #0x1\n" + "st1 { v31.b }[1], [x20], #0x1\n" "beq 4f\n" - "st1 { v26.b }[2], [x26], #0x1\n" "subs %x[n_channels], %x[n_channels], #0x1\n" - "st1 { v1.b }[2], [x25], #0x1\n" - "st1 { v22.b }[2], [x24], #0x1\n" - "st1 { v25.b }[2], [x23], #0x1\n" - "st1 { v11.b }[2], [x22], #0x1\n" - "st1 { v3.b }[2], [x21], #0x1\n" - "st1 { v19.b }[2], [x20], #0x1\n" - "st1 { v21.b }[2], [x19], #0x1\n" + "st1 { v24.b }[2], [x27], #0x1\n" + "st1 { v25.b }[2], [x26], #0x1\n" + "st1 { v26.b }[2], [x25], #0x1\n" + "st1 { v27.b }[2], [x24], #0x1\n" + "st1 { v28.b }[2], [x23], #0x1\n" + "st1 { v29.b }[2], [x22], #0x1\n" + "st1 { v30.b }[2], [x21], #0x1\n" + "st1 { v31.b }[2], [x20], #0x1\n" "beq 4f\n" - "st1 { v26.b }[3], [x26], #0x1\n" + "st1 { v24.b }[3], [x27], #0x1\n" "subs %x[n_channels], %x[n_channels], #0x1\n" - "st1 { v1.b }[3], [x25], #0x1\n" - "st1 { v22.b }[3], [x24], #0x1\n" - "st1 { v25.b }[3], [x23], #0x1\n" - "st1 { v11.b }[3], [x22], #0x1\n" - "st1 { v3.b }[3], [x21], #0x1\n" - "st1 { v19.b }[3], [x20], #0x1\n" - "st1 { v21.b }[3], [x19], #0x1\n" + "st1 { v25.b }[3], [x26], #0x1\n" + "st1 { v26.b }[3], [x25], #0x1\n" + "st1 { v27.b }[3], [x24], #0x1\n" + "st1 { v28.b }[3], [x23], #0x1\n" + "st1 { v29.b }[3], [x22], #0x1\n" + "st1 { v30.b }[3], [x21], #0x1\n" + "st1 { v31.b }[3], [x20], #0x1\n" "4:" // Tail: End - "add SP, SP, #0x80\n" : [n_channels] "+&r" (n_output_channels), [params] "+&r" (params) : [inptrs] "r" (inptrs), [offsetof_Requantize32_b_offset] "I" (offsetof(arm_gemm::Requantize32, b_offset)), [offsetof_Requantize32_c_offset] "I" (offsetof(arm_gemm::Requantize32, c_offset)), [offsetof_Requantize32_maxval] "I" (offsetof(arm_gemm::Requantize32, maxval)), [offsetof_Requantize32_minval] "I" (offsetof(arm_gemm::Requantize32, minval)), [outptrs] "r" (outptrs), [qp] "r" (&qp) - : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" ); } -- cgit v1.2.1