aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
diff options
context:
space:
mode:
authorMichael Tyler <michael.tyler@arm.com>2023-02-01 16:37:07 +0000
committermichael.tyler <michael.tyler@arm.com>2023-02-08 15:33:26 +0000
commit7d9a626aaba9837cb82d189a9c4f0bcef58825bb (patch)
treee3d8cfeb7f8539cca3a8bf2f1a8f412d25d89041 /src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
parent4e2bbbbb23e6f4bd452f7f865e51228e1f51efec (diff)
downloadComputeLibrary-7d9a626aaba9837cb82d189a9c4f0bcef58825bb.tar.gz
Update CPU kernels to remove x19 and w19
Resolves: COMPMID-5805 Change-Id: Idf720bbb136474810086f5089c5ed23b3f79835a Signed-off-by: Michael Tyler <michael.tyler@arm.com> Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9081 Benchmark: Arm Jenkins <bsgcomp@arm.com> Tested-by: Arm Jenkins <bsgcomp@arm.com> Comments-Addressed: Arm Jenkins <bsgcomp@arm.com> Reviewed-by: Gunes Bayir <gunes.bayir@arm.com> Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp1084
1 files changed, 529 insertions, 555 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
index b198eff6ac..9fa38c6efe 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021, 2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -40,622 +40,596 @@ void a64_s8q_packed_to_nhwc_5x5_s1_with_multiplier_output4x2_dot_depthfirst_impl
)
{
__asm__ __volatile__(
- "movi v15.16b, #0x1\n"
- "ldr x21, [%x[inptrs], #0x0]\n"
- "add SP, SP, #-0x80\n"
- "movi v14.4s, #0x1\n"
- "ldr x20, [%x[inptrs], #0x8]\n"
- "add x22, %x[qp], %[offsetof_Requantize32_b_offset]\n"
- "movi v28.4s, #0x0\n"
- "ldr x19, [%x[inptrs], #0x10]\n"
- "mov x11, #0x0\n"
- "movi v27.4s, #0x0\n"
- "ld1 { v13.16b }, [x21]\n"
- "mov x10, #0x0\n"
- "movi v26.4s, #0x0\n"
- "ld1 { v12.16b }, [x20]\n"
- "add x9, %x[qp], %[offsetof_Requantize32_c_offset]\n"
- "movi v25.4s, #0x0\n"
- "ld1 { v7.16b }, [x19]\n"
- "add x28, %x[qp], %[offsetof_Requantize32_minval]\n"
+ "ldr q12, [%x[params], #0x0]\n"
+ "ldr q8, [%x[params], #0x10]\n"
+ "movi v28.16b, #0x1\n"
+ "movi v18.4s, #0x0\n"
+ "ldr q9, [%x[params], #0x20]\n"
+ "ldr q10, [%x[params], #0x30]\n"
+ "movi v31.4s, #0x0\n"
"movi v24.4s, #0x0\n"
- "ldr x21, [%x[inptrs], #0x18]\n"
- "add x27, %x[qp], %[offsetof_Requantize32_maxval]\n"
- "mov v18.16b, v13.16b\n"
+ "ldr q11, [%x[params], #0x40]\n"
+ "ldr x20, [%x[inptrs], #0x18]\n"
+ "movi v30.4s, #0x0\n"
+ "movi v21.4s, #0x0\n"
+ "ld1 { v3.16b }, [x20]\n"
"ldr x20, [%x[inptrs], #0x20]\n"
- "cmp %x[n_channels], #0x4\n"
- "ext v18.16b, v18.16b, v18.16b, #0x1\n"
- "ldr x19, [%x[inptrs], #0x28]\n"
- "mov v17.16b, v12.16b\n"
- "ld1 { v6.16b }, [x21]\n"
- "ext v17.16b, v17.16b, v17.16b, #0x1\n"
- "ld1 { v5.16b }, [x20]\n"
- "mov v16.16b, v7.16b\n"
- "ld1 { v4.16b }, [x19]\n"
+ "mov v16.16b, v3.16b\n"
"ext v16.16b, v16.16b, v16.16b, #0x1\n"
- "ldr x20, [%x[inptrs], #0x30]\n"
- "zip1 v13.2d, v13.2d, v18.2d\n"
- "ldr x19, [%x[inptrs], #0x38]\n"
- "zip1 v12.2d, v12.2d, v17.2d\n"
- "ld1r { v3.4s }, [x22]\n"
- "mov v18.16b, v6.16b\n"
+ "ld1 { v4.16b }, [x20]\n"
+ "ldr x20, [%x[inptrs], #0x10]\n"
+ "mov v15.16b, v4.16b\n"
+ "ext v15.16b, v15.16b, v15.16b, #0x1\n"
"ld1 { v2.16b }, [x20]\n"
- "zip1 v7.2d, v7.2d, v16.2d\n"
- "ld1 { v1.16b }, [x19]\n"
- "ext v18.16b, v18.16b, v18.16b, #0x1\n"
- "ldp x26, x25, [%x[outptrs], #0x0]\n"
- "mov v17.16b, v5.16b\n"
- "ldp x24, x23, [%x[outptrs], #0x10]\n"
- "ext v17.16b, v17.16b, v17.16b, #0x1\n"
- "ldp x22, x21, [%x[outptrs], #0x20]\n"
- "mov v16.16b, v4.16b\n"
- "ldp x20, x19, [%x[outptrs], #0x30]\n"
- "zip1 v6.2d, v6.2d, v18.2d\n"
- "ld1r { v0.4s }, [x9]\n"
- "ext v16.16b, v16.16b, v16.16b, #0x1\n"
- "ld1r { v31.4s }, [x28]\n"
- "zip1 v5.2d, v5.2d, v17.2d\n"
- "ld1r { v30.4s }, [x27]\n"
- "mov v17.16b, v2.16b\n"
- "ldr q29, [%x[params], #0x0]\n"
+ "ldr x20, [%x[inptrs], #0x8]\n"
+ "mov v20.16b, v2.16b\n"
+ "ext v20.16b, v20.16b, v20.16b, #0x1\n"
+ "ld1 { v1.16b }, [x20]\n"
+ "ldr x20, [%x[inptrs], #0x28]\n"
+ "zip1 v3.2d, v3.2d, v16.2d\n"
+ "zip1 v4.2d, v4.2d, v15.2d\n"
+ "ld1 { v5.16b }, [x20]\n"
+ "ldr x20, [%x[inptrs], #0x30]\n"
+ "mov v26.16b, v1.16b\n"
+ "mov v13.16b, v5.16b\n"
+ "ld1 { v6.16b }, [x20]\n"
+ "ldr x20, [%x[inptrs], #0x38]\n"
+ "mov v19.16b, v6.16b\n"
+ "ext v26.16b, v26.16b, v26.16b, #0x1\n"
+ "ld1 { v7.16b }, [x20]\n"
+ "ldr x20, [%x[inptrs], #0x0]\n"
+ "mov v17.16b, v7.16b\n"
+ "zip1 v2.2d, v2.2d, v20.2d\n"
+ "ld1 { v0.16b }, [x20]\n"
+ "ext v13.16b, v13.16b, v13.16b, #0x1\n"
+ "ext v19.16b, v19.16b, v19.16b, #0x1\n"
+ ".inst 0x4f83e392 // sdot v18.4s, v28.16b, v3.4b[0]\n"
"ext v17.16b, v17.16b, v17.16b, #0x1\n"
- "ldr q8, [%x[params], #0x10]\n"
- "zip1 v4.2d, v4.2d, v16.2d\n"
- "ldr q9, [%x[params], #0x20]\n"
- "mov v16.16b, v1.16b\n"
- "ldr q10, [%x[params], #0x30]\n"
+ ".inst 0x4f83eb9f // sdot v31.4s, v28.16b, v3.4b[2]\n"
+ ".inst 0x4f84e398 // sdot v24.4s, v28.16b, v4.4b[0]\n"
+ "add x20, %x[qp], %[offsetof_Requantize32_b_offset]\n"
+ "ld1r { v23.4s }, [x20]\n"
+ ".inst 0x4f84eb9e // sdot v30.4s, v28.16b, v4.4b[2]\n"
+ "mov v16.16b, v0.16b\n"
+ ".inst 0x4f82e395 // sdot v21.4s, v28.16b, v2.4b[0]\n"
+ "movi v20.4s, #0x0\n"
+ "movi v29.4s, #0x1\n"
+ ".inst 0x4f82eb94 // sdot v20.4s, v28.16b, v2.4b[2]\n"
+ "add x20, %x[qp], %[offsetof_Requantize32_c_offset]\n"
+ "ld1r { v14.4s }, [x20]\n"
"ext v16.16b, v16.16b, v16.16b, #0x1\n"
- "ldr q11, [%x[params], #0x40]\n"
- "add %x[params], %x[params], #0x50\n"
- "zip1 v2.2d, v2.2d, v17.2d\n"
- "movi v23.4s, #0x0\n"
+ "zip1 v1.2d, v1.2d, v26.2d\n"
+ ".inst 0x4fa3e3b2 // sdot v18.4s, v29.16b, v3.4b[1]\n"
+ "zip1 v5.2d, v5.2d, v13.2d\n"
+ "zip1 v6.2d, v6.2d, v19.2d\n"
+ ".inst 0x4fa3ebbf // sdot v31.4s, v29.16b, v3.4b[3]\n"
+ "add x20, %x[qp], %[offsetof_Requantize32_minval]\n"
+ "ld1r { v13.4s }, [x20]\n"
+ "zip1 v7.2d, v7.2d, v17.2d\n"
"movi v22.4s, #0x0\n"
- "zip1 v1.2d, v1.2d, v16.2d\n"
- "movi v21.4s, #0x0\n"
- "movi v18.4s, #0x0\n"
- "movi v17.4s, #0x0\n"
- "movi v16.4s, #0x0\n"
- "movi v20.4s, #0x0\n"
+ ".inst 0x4fa4e3b8 // sdot v24.4s, v29.16b, v4.4b[1]\n"
+ "movi v26.4s, #0x0\n"
+ ".inst 0x4fa4ebbe // sdot v30.4s, v29.16b, v4.4b[3]\n"
+ ".inst 0x4f81e396 // sdot v22.4s, v28.16b, v1.4b[0]\n"
+ "add x20, %x[qp], %[offsetof_Requantize32_maxval]\n"
+ "ld1r { v15.4s }, [x20]\n"
+ "movi v25.4s, #0x0\n"
+ "movi v27.4s, #0x0\n"
+ ".inst 0x4f81eb9a // sdot v26.4s, v28.16b, v1.4b[2]\n"
+ "zip1 v0.2d, v0.2d, v16.2d\n"
"movi v19.4s, #0x0\n"
- ".inst 0x4f8de1fc // sdot v28.4s, v15.16b, v13.4b[0]\n"
- ".inst 0x4f8de9fb // sdot v27.4s, v15.16b, v13.4b[2]\n"
- ".inst 0x4f8ce1fa // sdot v26.4s, v15.16b, v12.4b[0]\n"
- ".inst 0x4f8ce9f9 // sdot v25.4s, v15.16b, v12.4b[2]\n"
- ".inst 0x4fade1dc // sdot v28.4s, v14.16b, v13.4b[1]\n"
- ".inst 0x4fade9db // sdot v27.4s, v14.16b, v13.4b[3]\n"
- ".inst 0x4face1da // sdot v26.4s, v14.16b, v12.4b[1]\n"
- ".inst 0x4face9d9 // sdot v25.4s, v14.16b, v12.4b[3]\n"
- ".inst 0x4f87e1f8 // sdot v24.4s, v15.16b, v7.4b[0]\n"
- ".inst 0x4f87e9f7 // sdot v23.4s, v15.16b, v7.4b[2]\n"
- ".inst 0x4f86e1f6 // sdot v22.4s, v15.16b, v6.4b[0]\n"
- ".inst 0x4f86e9f5 // sdot v21.4s, v15.16b, v6.4b[2]\n"
- ".inst 0x4fa7e1d8 // sdot v24.4s, v14.16b, v7.4b[1]\n"
- ".inst 0x4fa7e9d7 // sdot v23.4s, v14.16b, v7.4b[3]\n"
- ".inst 0x4fa6e1d6 // sdot v22.4s, v14.16b, v6.4b[1]\n"
- ".inst 0x4fa6e9d5 // sdot v21.4s, v14.16b, v6.4b[3]\n"
- ".inst 0x4f85e1f2 // sdot v18.4s, v15.16b, v5.4b[0]\n"
- ".inst 0x4f85e9f1 // sdot v17.4s, v15.16b, v5.4b[2]\n"
- ".inst 0x4f84e1f0 // sdot v16.4s, v15.16b, v4.4b[0]\n"
- ".inst 0x4f84e9f4 // sdot v20.4s, v15.16b, v4.4b[2]\n"
- ".inst 0x4fa5e1d2 // sdot v18.4s, v14.16b, v5.4b[1]\n"
- ".inst 0x4fa5e9d1 // sdot v17.4s, v14.16b, v5.4b[3]\n"
- ".inst 0x4fa4e1d0 // sdot v16.4s, v14.16b, v4.4b[1]\n"
- ".inst 0x4fa4e9d4 // sdot v20.4s, v14.16b, v4.4b[3]\n"
- ".inst 0x4f82e1f3 // sdot v19.4s, v15.16b, v2.4b[0]\n"
- "mov v28.16b, v28.16b\n"
- "mov v27.16b, v27.16b\n"
- "add v28.4s, v28.4s, v26.4s\n"
- ".inst 0x4fa2e1d3 // sdot v19.4s, v14.16b, v2.4b[1]\n"
- "add v27.4s, v27.4s, v25.4s\n"
- "add v28.4s, v28.4s, v24.4s\n"
- "mov v26.16b, v26.16b\n"
- "add v27.4s, v27.4s, v23.4s\n"
- "add v28.4s, v28.4s, v22.4s\n"
- "mov v25.16b, v25.16b\n"
- "add v27.4s, v27.4s, v21.4s\n"
- "add v28.4s, v28.4s, v18.4s\n"
- "add v26.4s, v26.4s, v24.4s\n"
- "add v27.4s, v27.4s, v17.4s\n"
- "add v25.4s, v25.4s, v23.4s\n"
- "add v26.4s, v26.4s, v22.4s\n"
- "mov v24.16b, v24.16b\n"
- "add v25.4s, v25.4s, v21.4s\n"
- "add v26.4s, v26.4s, v18.4s\n"
- "mov v23.16b, v23.16b\n"
- "add v25.4s, v25.4s, v17.4s\n"
- "add v26.4s, v26.4s, v16.4s\n"
- "add v24.4s, v24.4s, v22.4s\n"
- "add v25.4s, v25.4s, v20.4s\n"
- "add v23.4s, v23.4s, v21.4s\n"
- "add v24.4s, v24.4s, v18.4s\n"
- "mov v22.16b, v22.16b\n"
- "add v23.4s, v23.4s, v17.4s\n"
- "add v24.4s, v24.4s, v16.4s\n"
- "mov v21.16b, v21.16b\n"
- "add v23.4s, v23.4s, v20.4s\n"
- "add v24.4s, v24.4s, v19.4s\n"
- "add v22.4s, v22.4s, v18.4s\n"
+ ".inst 0x4f85e399 // sdot v25.4s, v28.16b, v5.4b[0]\n"
+ "cmp %x[n_channels], #0x4\n"
+ ".inst 0x4f85eb9b // sdot v27.4s, v28.16b, v5.4b[2]\n"
+ ".inst 0x4f86e393 // sdot v19.4s, v28.16b, v6.4b[0]\n"
+ "add v24.4s, v18.4s, v24.4s\n"
+ "mov x9, #0x0\n"
"movi v18.4s, #0x0\n"
- ".inst 0x4f82e9f2 // sdot v18.4s, v15.16b, v2.4b[2]\n"
- "add v21.4s, v21.4s, v17.4s\n"
- "movi v17.4s, #0x0\n"
- ".inst 0x4f81e1f1 // sdot v17.4s, v15.16b, v1.4b[0]\n"
- ".inst 0x4fa2e9d2 // sdot v18.4s, v14.16b, v2.4b[3]\n"
- "add v22.4s, v22.4s, v16.4s\n"
+ ".inst 0x4f86eb92 // sdot v18.4s, v28.16b, v6.4b[2]\n"
+ ".inst 0x4fa2e3b5 // sdot v21.4s, v29.16b, v2.4b[1]\n"
+ "mov x28, #0x0\n"
+ ".inst 0x4fa2ebb4 // sdot v20.4s, v29.16b, v2.4b[3]\n"
+ "add v17.4s, v31.4s, v30.4s\n"
+ ".inst 0x4fa1e3b6 // sdot v22.4s, v29.16b, v1.4b[1]\n"
+ "ldp x27, x26, [%x[outptrs], #0x0]\n"
"movi v16.4s, #0x0\n"
- ".inst 0x4fa1e1d1 // sdot v17.4s, v14.16b, v1.4b[1]\n"
- ".inst 0x4f81e9f0 // sdot v16.4s, v15.16b, v1.4b[2]\n"
- "add v23.4s, v23.4s, v18.4s\n"
- "add v21.4s, v21.4s, v20.4s\n"
- "add v22.4s, v22.4s, v19.4s\n"
- ".inst 0x4fa1e9d0 // sdot v16.4s, v14.16b, v1.4b[3]\n"
- "add v21.4s, v21.4s, v18.4s\n"
- "add v22.4s, v22.4s, v17.4s\n"
- "neg v3.4s, v3.4s\n"
- "add v21.4s, v21.4s, v16.4s\n"
- "mul v28.4s, v28.4s, v3.4s\n"
- "str q28, [SP, #0x0]\n"
- "mul v27.4s, v27.4s, v3.4s\n"
- "mul v26.4s, v26.4s, v3.4s\n"
- "str q27, [SP, #0x10]\n"
- "mul v25.4s, v25.4s, v3.4s\n"
- "mul v24.4s, v24.4s, v3.4s\n"
- "str q26, [SP, #0x20]\n"
- "mul v23.4s, v23.4s, v3.4s\n"
- "str q25, [SP, #0x30]\n"
- "mul v22.4s, v22.4s, v3.4s\n"
- "mul v21.4s, v21.4s, v3.4s\n"
- "str q24, [SP, #0x40]\n"
- "add v28.4s, v28.4s, v29.4s\n"
- "str q23, [SP, #0x50]\n"
- "add v27.4s, v27.4s, v29.4s\n"
- "str q22, [SP, #0x60]\n"
- "add v26.4s, v26.4s, v29.4s\n"
- "add v25.4s, v25.4s, v29.4s\n"
- "str q21, [SP, #0x70]\n"
- "add v24.4s, v24.4s, v29.4s\n"
- "add v23.4s, v23.4s, v29.4s\n"
- "add v22.4s, v22.4s, v29.4s\n"
- "add v21.4s, v21.4s, v29.4s\n"
+ ".inst 0x4f87e390 // sdot v16.4s, v28.16b, v7.4b[0]\n"
+ ".inst 0x4fa1ebba // sdot v26.4s, v29.16b, v1.4b[3]\n"
+ "ldp x25, x24, [%x[outptrs], #0x10]\n"
+ ".inst 0x4fa5e3b9 // sdot v25.4s, v29.16b, v5.4b[1]\n"
+ ".inst 0x4fa5ebbb // sdot v27.4s, v29.16b, v5.4b[3]\n"
+ "add v30.4s, v21.4s, v24.4s\n"
+ "ldp x23, x22, [%x[outptrs], #0x20]\n"
+ ".inst 0x4fa6e3b3 // sdot v19.4s, v29.16b, v6.4b[1]\n"
+ ".inst 0x4fa6ebb2 // sdot v18.4s, v29.16b, v6.4b[3]\n"
+ "add v31.4s, v20.4s, v17.4s\n"
+ "ldp x21, x20, [%x[outptrs], #0x30]\n"
+ ".inst 0x4fa7e3b0 // sdot v16.4s, v29.16b, v7.4b[1]\n"
+ "add v22.4s, v22.4s, v30.4s\n"
+ "add %x[params], %x[params], #0x50\n"
+ "add v21.4s, v26.4s, v31.4s\n"
+ "add v20.4s, v25.4s, v19.4s\n"
+ "add v19.4s, v27.4s, v18.4s\n"
+ "add v18.4s, v16.4s, v24.4s\n"
+ "movi v16.4s, #0x0\n"
+ ".inst 0x4f87eb90 // sdot v16.4s, v28.16b, v7.4b[2]\n"
+ ".inst 0x4fa7ebb0 // sdot v16.4s, v29.16b, v7.4b[3]\n"
+ "add v17.4s, v16.4s, v17.4s\n"
+ "movi v16.4s, #0x0\n"
+ ".inst 0x4f80e390 // sdot v16.4s, v28.16b, v0.4b[0]\n"
+ ".inst 0x4fa0e3b0 // sdot v16.4s, v29.16b, v0.4b[1]\n"
+ "add v24.4s, v22.4s, v16.4s\n"
+ "add v26.4s, v22.4s, v25.4s\n"
+ "movi v16.4s, #0x0\n"
+ ".inst 0x4f80eb90 // sdot v16.4s, v28.16b, v0.4b[2]\n"
+ ".inst 0x4fa0ebb0 // sdot v16.4s, v29.16b, v0.4b[3]\n"
+ "add v25.4s, v21.4s, v16.4s\n"
+ "add v27.4s, v21.4s, v27.4s\n"
+ "add v28.4s, v20.4s, v30.4s\n"
+ "add v29.4s, v19.4s, v31.4s\n"
+ "add v30.4s, v18.4s, v20.4s\n"
+ "add v31.4s, v17.4s, v19.4s\n"
+ "neg v23.4s, v23.4s\n"
+ "mul v24.4s, v24.4s, v23.4s\n"
+ "mul v25.4s, v25.4s, v23.4s\n"
+ "mul v26.4s, v26.4s, v23.4s\n"
+ "mul v27.4s, v27.4s, v23.4s\n"
+ "mul v28.4s, v28.4s, v23.4s\n"
+ "mul v29.4s, v29.4s, v23.4s\n"
+ "mul v30.4s, v30.4s, v23.4s\n"
+ "mul v31.4s, v31.4s, v23.4s\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, 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"
"ble 2f\n"
"1:" // Loop
- ".inst 0x4f8de11c // sdot v28.4s, v8.16b, v13.4b[0]\n"
- "ldr q20, [%x[params], #0x60]\n"
- "add x11, x11, #0x10\n"
- ".inst 0x4f8de91b // sdot v27.4s, v8.16b, v13.4b[2]\n"
- "ldr q19, [%x[params], #0x70]\n"
+ "ldr q21, [%x[params], #0x60]\n"
+ "ldr q20, [%x[params], #0x70]\n"
+ ".inst 0x4f80e118 // sdot v24.4s, v8.16b, v0.4b[0]\n"
+ ".inst 0x4f80e919 // sdot v25.4s, v8.16b, v0.4b[2]\n"
+ "ldr q12, [%x[params], #0x80]\n"
+ ".inst 0x4f81e11a // sdot v26.4s, v8.16b, v1.4b[0]\n"
+ ".inst 0x4f81e91b // sdot v27.4s, v8.16b, v1.4b[2]\n"
"sub %x[n_channels], %x[n_channels], #0x4\n"
- ".inst 0x4f8ce11a // sdot v26.4s, v8.16b, v12.4b[0]\n"
- "ldr q29, [%x[params], #0x80]\n"
+ ".inst 0x4fa0e138 // sdot v24.4s, v9.16b, v0.4b[1]\n"
+ ".inst 0x4fa0e939 // sdot v25.4s, v9.16b, v0.4b[3]\n"
"cmp %x[n_channels], #0x4\n"
- ".inst 0x4f8ce919 // sdot v25.4s, v8.16b, v12.4b[2]\n"
- ".inst 0x4f87e118 // sdot v24.4s, v8.16b, v7.4b[0]\n"
- ".inst 0x4f87e917 // sdot v23.4s, v8.16b, v7.4b[2]\n"
- ".inst 0x4f86e116 // sdot v22.4s, v8.16b, v6.4b[0]\n"
- ".inst 0x4f86e915 // sdot v21.4s, v8.16b, v6.4b[2]\n"
+ "add x9, x9, #0x10\n"
+ ".inst 0x4fa1e13a // sdot v26.4s, v9.16b, v1.4b[1]\n"
+ ".inst 0x4fa1e93b // sdot v27.4s, v9.16b, v1.4b[3]\n"
+ ".inst 0x4f82e11c // sdot v28.4s, v8.16b, v2.4b[0]\n"
+ ".inst 0x4f82e91d // sdot v29.4s, v8.16b, v2.4b[2]\n"
+ ".inst 0x4f83e11e // sdot v30.4s, v8.16b, v3.4b[0]\n"
+ ".inst 0x4f83e91f // sdot v31.4s, v8.16b, v3.4b[2]\n"
"ldr q8, [%x[params], #0x0]\n"
- ".inst 0x4fade13c // sdot v28.4s, v9.16b, v13.4b[1]\n"
- ".inst 0x4fade93b // sdot v27.4s, v9.16b, v13.4b[3]\n"
- ".inst 0x4face13a // sdot v26.4s, v9.16b, v12.4b[1]\n"
- ".inst 0x4face939 // sdot v25.4s, v9.16b, v12.4b[3]\n"
- ".inst 0x4fa7e138 // sdot v24.4s, v9.16b, v7.4b[1]\n"
- ".inst 0x4fa7e937 // sdot v23.4s, v9.16b, v7.4b[3]\n"
- ".inst 0x4fa6e136 // sdot v22.4s, v9.16b, v6.4b[1]\n"
- ".inst 0x4fa6e935 // sdot v21.4s, v9.16b, v6.4b[3]\n"
+ ".inst 0x4f81e158 // sdot v24.4s, v10.16b, v1.4b[0]\n"
+ ".inst 0x4f81e959 // sdot v25.4s, v10.16b, v1.4b[2]\n"
+ ".inst 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ ".inst 0x4f82e95b // sdot v27.4s, v10.16b, v2.4b[2]\n"
+ ".inst 0x4fa2e13c // sdot v28.4s, v9.16b, v2.4b[1]\n"
+ ".inst 0x4fa2e93d // sdot v29.4s, v9.16b, v2.4b[3]\n"
+ ".inst 0x4fa3e13e // sdot v30.4s, v9.16b, v3.4b[1]\n"
+ ".inst 0x4fa3e93f // sdot v31.4s, v9.16b, v3.4b[3]\n"
"ldr q9, [%x[params], #0x10]\n"
- ".inst 0x4f8ce15c // sdot v28.4s, v10.16b, v12.4b[0]\n"
- ".inst 0x4f8ce95b // sdot v27.4s, v10.16b, v12.4b[2]\n"
- ".inst 0x4f87e15a // sdot v26.4s, v10.16b, v7.4b[0]\n"
- ".inst 0x4f87e959 // sdot v25.4s, v10.16b, v7.4b[2]\n"
- ".inst 0x4f86e158 // sdot v24.4s, v10.16b, v6.4b[0]\n"
- ".inst 0x4f86e957 // sdot v23.4s, v10.16b, v6.4b[2]\n"
- ".inst 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
- ".inst 0x4f85e955 // sdot v21.4s, v10.16b, v5.4b[2]\n"
+ ".inst 0x4fa1e178 // sdot v24.4s, v11.16b, v1.4b[1]\n"
+ ".inst 0x4fa1e979 // sdot v25.4s, v11.16b, v1.4b[3]\n"
+ ".inst 0x4fa2e17a // sdot v26.4s, v11.16b, v2.4b[1]\n"
+ ".inst 0x4fa2e97b // sdot v27.4s, v11.16b, v2.4b[3]\n"
+ ".inst 0x4f83e15c // sdot v28.4s, v10.16b, v3.4b[0]\n"
+ ".inst 0x4f83e95d // sdot v29.4s, v10.16b, v3.4b[2]\n"
+ ".inst 0x4f84e15e // sdot v30.4s, v10.16b, v4.4b[0]\n"
+ ".inst 0x4f84e95f // sdot v31.4s, v10.16b, v4.4b[2]\n"
"ldr q10, [%x[params], #0x20]\n"
- ".inst 0x4face17c // sdot v28.4s, v11.16b, v12.4b[1]\n"
- ".inst 0x4face97b // sdot v27.4s, v11.16b, v12.4b[3]\n"
- ".inst 0x4fa7e17a // sdot v26.4s, v11.16b, v7.4b[1]\n"
- ".inst 0x4fa7e979 // sdot v25.4s, v11.16b, v7.4b[3]\n"
- ".inst 0x4fa6e178 // sdot v24.4s, v11.16b, v6.4b[1]\n"
- ".inst 0x4fa6e977 // sdot v23.4s, v11.16b, v6.4b[3]\n"
- ".inst 0x4fa5e176 // sdot v22.4s, v11.16b, v5.4b[1]\n"
- ".inst 0x4fa5e975 // sdot v21.4s, v11.16b, v5.4b[3]\n"
+ ".inst 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ ".inst 0x4f82e919 // sdot v25.4s, v8.16b, v2.4b[2]\n"
+ ".inst 0x4f83e11a // sdot v26.4s, v8.16b, v3.4b[0]\n"
+ ".inst 0x4f83e91b // sdot v27.4s, v8.16b, v3.4b[2]\n"
+ ".inst 0x4fa3e17c // sdot v28.4s, v11.16b, v3.4b[1]\n"
+ ".inst 0x4fa3e97d // sdot v29.4s, v11.16b, v3.4b[3]\n"
+ ".inst 0x4fa4e17e // sdot v30.4s, v11.16b, v4.4b[1]\n"
+ ".inst 0x4fa4e97f // sdot v31.4s, v11.16b, v4.4b[3]\n"
"ldr q11, [%x[params], #0x30]\n"
- ".inst 0x4f87e11c // sdot v28.4s, v8.16b, v7.4b[0]\n"
- ".inst 0x4f87e91b // sdot v27.4s, v8.16b, v7.4b[2]\n"
- ".inst 0x4f86e11a // sdot v26.4s, v8.16b, v6.4b[0]\n"
- ".inst 0x4f86e919 // sdot v25.4s, v8.16b, v6.4b[2]\n"
- ".inst 0x4f85e118 // sdot v24.4s, v8.16b, v5.4b[0]\n"
- ".inst 0x4f85e917 // sdot v23.4s, v8.16b, v5.4b[2]\n"
- ".inst 0x4f84e116 // sdot v22.4s, v8.16b, v4.4b[0]\n"
- ".inst 0x4f84e915 // sdot v21.4s, v8.16b, v4.4b[2]\n"
+ ".inst 0x4fa2e138 // sdot v24.4s, v9.16b, v2.4b[1]\n"
+ ".inst 0x4fa2e939 // sdot v25.4s, v9.16b, v2.4b[3]\n"
+ ".inst 0x4fa3e13a // sdot v26.4s, v9.16b, v3.4b[1]\n"
+ ".inst 0x4fa3e93b // sdot v27.4s, v9.16b, v3.4b[3]\n"
+ ".inst 0x4f84e11c // sdot v28.4s, v8.16b, v4.4b[0]\n"
+ ".inst 0x4f84e91d // sdot v29.4s, v8.16b, v4.4b[2]\n"
+ ".inst 0x4f85e11e // sdot v30.4s, v8.16b, v5.4b[0]\n"
+ ".inst 0x4f85e91f // sdot v31.4s, v8.16b, v5.4b[2]\n"
"ldr q8, [%x[params], #0x40]\n"
- ".inst 0x4fa7e13c // sdot v28.4s, v9.16b, v7.4b[1]\n"
- ".inst 0x4fa7e93b // sdot v27.4s, v9.16b, v7.4b[3]\n"
- ".inst 0x4fa6e13a // sdot v26.4s, v9.16b, v6.4b[1]\n"
- ".inst 0x4fa6e939 // sdot v25.4s, v9.16b, v6.4b[3]\n"
- ".inst 0x4fa5e138 // sdot v24.4s, v9.16b, v5.4b[1]\n"
- ".inst 0x4fa5e937 // sdot v23.4s, v9.16b, v5.4b[3]\n"
- ".inst 0x4fa4e136 // sdot v22.4s, v9.16b, v4.4b[1]\n"
- ".inst 0x4fa4e935 // sdot v21.4s, v9.16b, v4.4b[3]\n"
+ ".inst 0x4f83e158 // sdot v24.4s, v10.16b, v3.4b[0]\n"
+ ".inst 0x4f83e959 // sdot v25.4s, v10.16b, v3.4b[2]\n"
+ ".inst 0x4f84e15a // sdot v26.4s, v10.16b, v4.4b[0]\n"
+ ".inst 0x4f84e95b // sdot v27.4s, v10.16b, v4.4b[2]\n"
+ ".inst 0x4fa4e13c // sdot v28.4s, v9.16b, v4.4b[1]\n"
+ ".inst 0x4fa4e93d // sdot v29.4s, v9.16b, v4.4b[3]\n"
+ ".inst 0x4fa5e13e // sdot v30.4s, v9.16b, v5.4b[1]\n"
+ ".inst 0x4fa5e93f // sdot v31.4s, v9.16b, v5.4b[3]\n"
"ldr q9, [%x[params], #0x50]\n"
- ".inst 0x4f86e15c // sdot v28.4s, v10.16b, v6.4b[0]\n"
- ".inst 0x4f86e95b // sdot v27.4s, v10.16b, v6.4b[2]\n"
- ".inst 0x4f85e15a // sdot v26.4s, v10.16b, v5.4b[0]\n"
- ".inst 0x4f85e959 // sdot v25.4s, v10.16b, v5.4b[2]\n"
- ".inst 0x4f84e158 // sdot v24.4s, v10.16b, v4.4b[0]\n"
- ".inst 0x4f84e957 // sdot v23.4s, v10.16b, v4.4b[2]\n"
- ".inst 0x4f82e156 // sdot v22.4s, v10.16b, v2.4b[0]\n"
- ".inst 0x4f82e955 // sdot v21.4s, v10.16b, v2.4b[2]\n"
+ ".inst 0x4fa3e178 // sdot v24.4s, v11.16b, v3.4b[1]\n"
+ ".inst 0x4fa3e979 // sdot v25.4s, v11.16b, v3.4b[3]\n"
+ ".inst 0x4fa4e17a // sdot v26.4s, v11.16b, v4.4b[1]\n"
+ ".inst 0x4fa4e97b // sdot v27.4s, v11.16b, v4.4b[3]\n"
+ ".inst 0x4f85e15c // sdot v28.4s, v10.16b, v5.4b[0]\n"
+ ".inst 0x4f85e95d // sdot v29.4s, v10.16b, v5.4b[2]\n"
+ ".inst 0x4f86e15e // sdot v30.4s, v10.16b, v6.4b[0]\n"
+ ".inst 0x4f86e95f // sdot v31.4s, v10.16b, v6.4b[2]\n"
"ldr q10, [%x[params], #0xb0]\n"
- ".inst 0x4fa6e17c // sdot v28.4s, v11.16b, v6.4b[1]\n"
- ".inst 0x4fa6e97b // sdot v27.4s, v11.16b, v6.4b[3]\n"
- ".inst 0x4fa5e17a // sdot v26.4s, v11.16b, v5.4b[1]\n"
- ".inst 0x4fa5e979 // sdot v25.4s, v11.16b, v5.4b[3]\n"
- ".inst 0x4fa4e178 // sdot v24.4s, v11.16b, v4.4b[1]\n"
- ".inst 0x4fa4e977 // sdot v23.4s, v11.16b, v4.4b[3]\n"
- ".inst 0x4fa2e176 // sdot v22.4s, v11.16b, v2.4b[1]\n"
- ".inst 0x4fa2e975 // sdot v21.4s, v11.16b, v2.4b[3]\n"
- "ldr q11, [%x[params], #0xc0]\n"
- ".inst 0x4f85e11c // sdot v28.4s, v8.16b, v5.4b[0]\n"
- ".inst 0x4f85e91b // sdot v27.4s, v8.16b, v5.4b[2]\n"
- ".inst 0x4f84e11a // sdot v26.4s, v8.16b, v4.4b[0]\n"
+ ".inst 0x4f84e118 // sdot v24.4s, v8.16b, v4.4b[0]\n"
".inst 0x4f84e919 // sdot v25.4s, v8.16b, v4.4b[2]\n"
- ".inst 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
- ".inst 0x4f82e917 // sdot v23.4s, v8.16b, v2.4b[2]\n"
- ".inst 0x4f81e116 // sdot v22.4s, v8.16b, v1.4b[0]\n"
- ".inst 0x4f81e915 // sdot v21.4s, v8.16b, v1.4b[2]\n"
- "ldr q8, [%x[params], #0x90]\n"
- ".inst 0x4fa5e13c // sdot v28.4s, v9.16b, v5.4b[1]\n"
- ".inst 0x4fa5e93b // sdot v27.4s, v9.16b, v5.4b[3]\n"
- ".inst 0x4fa4e13a // sdot v26.4s, v9.16b, v4.4b[1]\n"
+ ".inst 0x4f85e11a // sdot v26.4s, v8.16b, v5.4b[0]\n"
+ ".inst 0x4f85e91b // sdot v27.4s, v8.16b, v5.4b[2]\n"
+ ".inst 0x4fa5e17c // sdot v28.4s, v11.16b, v5.4b[1]\n"
+ ".inst 0x4fa5e97d // sdot v29.4s, v11.16b, v5.4b[3]\n"
+ ".inst 0x4fa6e17e // sdot v30.4s, v11.16b, v6.4b[1]\n"
+ ".inst 0x4fa6e97f // sdot v31.4s, v11.16b, v6.4b[3]\n"
+ "ldr q11, [%x[params], #0xc0]\n"
+ ".inst 0x4fa4e138 // sdot v24.4s, v9.16b, v4.4b[1]\n"
".inst 0x4fa4e939 // sdot v25.4s, v9.16b, v4.4b[3]\n"
- ".inst 0x4fa2e138 // sdot v24.4s, v9.16b, v2.4b[1]\n"
- ".inst 0x4fa2e937 // sdot v23.4s, v9.16b, v2.4b[3]\n"
- ".inst 0x4fa1e136 // sdot v22.4s, v9.16b, v1.4b[1]\n"
- ".inst 0x4fa1e935 // sdot v21.4s, v9.16b, v1.4b[3]\n"
+ "sqrdmulh v24.4s, v24.4s, v21.4s\n"
+ ".inst 0x4fa5e13a // sdot v26.4s, v9.16b, v5.4b[1]\n"
+ ".inst 0x4fa5e93b // sdot v27.4s, v9.16b, v5.4b[3]\n"
+ "sqrdmulh v25.4s, v25.4s, v21.4s\n"
+ ".inst 0x4f86e11c // sdot v28.4s, v8.16b, v6.4b[0]\n"
+ ".inst 0x4f86e91d // sdot v29.4s, v8.16b, v6.4b[2]\n"
+ "sqrdmulh v26.4s, v26.4s, v21.4s\n"
+ ".inst 0x4f87e11e // sdot v30.4s, v8.16b, v7.4b[0]\n"
+ ".inst 0x4f87e91f // sdot v31.4s, v8.16b, v7.4b[2]\n"
+ "ldr q8, [%x[params], #0x90]\n"
+ "sqrdmulh v27.4s, v27.4s, v21.4s\n"
+ ".inst 0x4fa6e13c // sdot v28.4s, v9.16b, v6.4b[1]\n"
+ ".inst 0x4fa6e93d // sdot v29.4s, v9.16b, v6.4b[3]\n"
+ "and v19.16b, v24.16b, v20.16b\n"
+ ".inst 0x4fa7e13e // sdot v30.4s, v9.16b, v7.4b[1]\n"
+ ".inst 0x4fa7e93f // sdot v31.4s, v9.16b, v7.4b[3]\n"
"ldr q9, [%x[params], #0xa0]\n"
- "add %x[params], %x[params], #0xd0\n"
- "sqrdmulh v28.4s, v28.4s, v20.4s\n"
- "sqrdmulh v27.4s, v27.4s, v20.4s\n"
- "sqrdmulh v26.4s, v26.4s, v20.4s\n"
- "sqrdmulh v25.4s, v25.4s, v20.4s\n"
- "sqrdmulh v24.4s, v24.4s, v20.4s\n"
- "and v18.16b, v28.16b, v19.16b\n"
- "and v17.16b, v27.16b, v19.16b\n"
- "and v16.16b, v26.16b, v19.16b\n"
+ "and v18.16b, v25.16b, v20.16b\n"
+ "sshr v19.4s, v19.4s, #0x1f\n"
"sshr v18.4s, v18.4s, #0x1f\n"
+ "add %x[params], %x[params], #0xd0\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"
+ "and v17.16b, v26.16b, v20.16b\n"
"sshr v17.4s, v17.4s, #0x1f\n"
+ "sqadd v24.4s, v24.4s, v19.4s\n"
+ "and v16.16b, v27.16b, v20.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "sqadd v28.4s, v28.4s, v18.4s\n"
- "sqadd v27.4s, v27.4s, v17.4s\n"
- "sqadd v26.4s, v26.4s, v16.4s\n"
- "and v16.16b, v25.16b, v19.16b\n"
- "srshl v28.4s, v28.4s, v19.4s\n"
- "srshl v27.4s, v27.4s, v19.4s\n"
- "srshl v26.4s, v26.4s, v19.4s\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "add v28.4s, v28.4s, v0.4s\n"
- "add v27.4s, v27.4s, v0.4s\n"
- "add v26.4s, v26.4s, v0.4s\n"
- "smin v28.4s, v28.4s, v30.4s\n"
- "smin v27.4s, v27.4s, v30.4s\n"
- "smin v26.4s, v26.4s, v30.4s\n"
- "smax v28.4s, v28.4s, v31.4s\n"
- "smax v27.4s, v27.4s, v31.4s\n"
- "smax v26.4s, v26.4s, v31.4s\n"
- "uzp1 v28.16b, v28.16b, v28.16b\n"
- "uzp1 v27.16b, v27.16b, v27.16b\n"
- "uzp1 v28.16b, v28.16b, v28.16b\n"
- "str s28, [x26, x10]\n"
- "uzp1 v27.16b, v27.16b, v27.16b\n"
- "uzp1 v26.16b, v26.16b, v26.16b\n"
- "ldr q28, [SP, #0x0]\n"
- "sqadd v25.4s, v25.4s, v16.4s\n"
- "str s27, [x25, x10]\n"
- "uzp1 v26.16b, v26.16b, v26.16b\n"
- "ldr q27, [SP, #0x10]\n"
- "and v16.16b, v24.16b, v19.16b\n"
- "str s26, [x24, x10]\n"
- "sqrdmulh v23.4s, v23.4s, v20.4s\n"
- "ldr q26, [SP, #0x20]\n"
- "srshl v25.4s, v25.4s, v19.4s\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sqrdmulh v22.4s, v22.4s, v20.4s\n"
- "and v17.16b, v23.16b, v19.16b\n"
- "add v25.4s, v25.4s, v0.4s\n"
- "sqadd v24.4s, v24.4s, v16.4s\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"
+ "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, v30.4s\n"
- "and v16.16b, v22.16b, v19.16b\n"
- "srshl v24.4s, v24.4s, v19.4s\n"
- "smax v25.4s, v25.4s, v31.4s\n"
- "sqadd v23.4s, v23.4s, v17.4s\n"
+ "sqadd v28.4s, v28.4s, v19.4s\n"
+ "and v16.16b, v31.16b, v20.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "add v24.4s, v24.4s, v0.4s\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "str s25, [x23, x10]\n"
- "smin v24.4s, v24.4s, v30.4s\n"
- "srshl v23.4s, v23.4s, v19.4s\n"
- "ldr q25, [SP, #0x30]\n"
- "sqadd v22.4s, v22.4s, v16.4s\n"
- "sqrdmulh v21.4s, v21.4s, v20.4s\n"
- "smax v24.4s, v24.4s, v31.4s\n"
- "add v23.4s, v23.4s, v0.4s\n"
- "srshl v22.4s, v22.4s, 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, 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"
+ "smin v24.4s, v24.4s, v15.4s\n"
+ "smin v25.4s, v25.4s, v15.4s\n"
+ "smin v26.4s, v26.4s, v15.4s\n"
+ "smin v27.4s, v27.4s, v15.4s\n"
+ "smin v28.4s, v28.4s, v15.4s\n"
+ "smin v29.4s, v29.4s, v15.4s\n"
+ "smin v30.4s, v30.4s, v15.4s\n"
+ "smin v31.4s, v31.4s, v15.4s\n"
+ "smax v24.4s, v24.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v26.4s, v26.4s, v13.4s\n"
+ "smax v27.4s, v27.4s, v13.4s\n"
+ "smax v28.4s, v28.4s, v13.4s\n"
+ "smax v29.4s, v29.4s, v13.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v31.4s, v31.4s, v13.4s\n"
"uzp1 v24.16b, v24.16b, v24.16b\n"
- "smin v23.4s, v23.4s, v30.4s\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\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"
"uzp1 v24.16b, v24.16b, v24.16b\n"
- "str s24, [x22, x10]\n"
- "smax v23.4s, v23.4s, v31.4s\n"
- "add v22.4s, v22.4s, v0.4s\n"
- "ldr q24, [SP, #0x40]\n"
- "and v16.16b, v21.16b, v19.16b\n"
- "add v28.4s, v28.4s, v29.4s\n"
- "uzp1 v23.16b, v23.16b, v23.16b\n"
- "smin v22.4s, v22.4s, v30.4s\n"
- "uzp1 v23.16b, v23.16b, v23.16b\n"
- "str s23, [x21, x10]\n"
- "smax v22.4s, v22.4s, v31.4s\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "ldr q23, [SP, #0x50]\n"
- "add v27.4s, v27.4s, v29.4s\n"
- "add v26.4s, v26.4s, v29.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "sqadd v21.4s, v21.4s, v16.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "str s22, [x20, x10]\n"
- "add v25.4s, v25.4s, v29.4s\n"
- "add v24.4s, v24.4s, v29.4s\n"
- "ldr q22, [SP, #0x60]\n"
- "srshl v21.4s, v21.4s, v19.4s\n"
- "add v23.4s, v23.4s, v29.4s\n"
- "add v21.4s, v21.4s, v0.4s\n"
- "add v22.4s, v22.4s, v29.4s\n"
- "smin v21.4s, v21.4s, v30.4s\n"
- "smax v21.4s, v21.4s, v31.4s\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, v29.4s\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\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, 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"
"bgt 1b\n"
"2:" // Tail
- ".inst 0x4f8de11c // sdot v28.4s, v8.16b, v13.4b[0]\n"
- "ldr q20, [%x[params], #0x60]\n"
- "add x26, x26, x10\n"
- ".inst 0x4f8de91b // sdot v27.4s, v8.16b, v13.4b[2]\n"
- "ldr q19, [%x[params], #0x70]\n"
- "add x25, x25, x10\n"
- ".inst 0x4f8ce11a // sdot v26.4s, v8.16b, v12.4b[0]\n"
- "add x24, x24, x10\n"
- ".inst 0x4f8ce919 // sdot v25.4s, v8.16b, v12.4b[2]\n"
- "add x23, x23, x10\n"
- ".inst 0x4f87e118 // sdot v24.4s, v8.16b, v7.4b[0]\n"
- "add x22, x22, x10\n"
- ".inst 0x4f87e917 // sdot v23.4s, v8.16b, v7.4b[2]\n"
- "add x21, x21, x10\n"
- ".inst 0x4f86e116 // sdot v22.4s, v8.16b, v6.4b[0]\n"
- "add x20, x20, x10\n"
- ".inst 0x4f86e915 // sdot v21.4s, v8.16b, v6.4b[2]\n"
- "ldr q8, [%x[params], #0x0]\n"
- "add x19, x19, x10\n"
- ".inst 0x4fade13c // sdot v28.4s, v9.16b, v13.4b[1]\n"
+ "ldr q21, [%x[params], #0x60]\n"
+ "ldr q20, [%x[params], #0x70]\n"
+ ".inst 0x4f80e118 // sdot v24.4s, v8.16b, v0.4b[0]\n"
+ ".inst 0x4f80e919 // sdot v25.4s, v8.16b, v0.4b[2]\n"
+ ".inst 0x4f81e11a // sdot v26.4s, v8.16b, v1.4b[0]\n"
+ ".inst 0x4f81e91b // sdot v27.4s, v8.16b, v1.4b[2]\n"
"cmp %x[n_channels], #0x4\n"
- ".inst 0x4fade93b // sdot v27.4s, v9.16b, v13.4b[3]\n"
- ".inst 0x4face13a // sdot v26.4s, v9.16b, v12.4b[1]\n"
- ".inst 0x4face939 // sdot v25.4s, v9.16b, v12.4b[3]\n"
- ".inst 0x4fa7e138 // sdot v24.4s, v9.16b, v7.4b[1]\n"
- ".inst 0x4fa7e937 // sdot v23.4s, v9.16b, v7.4b[3]\n"
- ".inst 0x4fa6e136 // sdot v22.4s, v9.16b, v6.4b[1]\n"
- ".inst 0x4fa6e935 // sdot v21.4s, v9.16b, v6.4b[3]\n"
+ "add x27, x27, x28\n"
+ ".inst 0x4fa0e138 // sdot v24.4s, v9.16b, v0.4b[1]\n"
+ ".inst 0x4fa0e939 // sdot v25.4s, v9.16b, v0.4b[3]\n"
+ "add x26, x26, x28\n"
+ "add x25, x25, x28\n"
+ ".inst 0x4fa1e13a // sdot v26.4s, v9.16b, v1.4b[1]\n"
+ ".inst 0x4fa1e93b // sdot v27.4s, v9.16b, v1.4b[3]\n"
+ "add x24, x24, x28\n"
+ "add x23, x23, x28\n"
+ ".inst 0x4f82e11c // sdot v28.4s, v8.16b, v2.4b[0]\n"
+ ".inst 0x4f82e91d // sdot v29.4s, v8.16b, v2.4b[2]\n"
+ "add x22, x22, x28\n"
+ "add x21, x21, x28\n"
+ ".inst 0x4f83e11e // sdot v30.4s, v8.16b, v3.4b[0]\n"
+ ".inst 0x4f83e91f // sdot v31.4s, v8.16b, v3.4b[2]\n"
+ "ldr q8, [%x[params], #0x0]\n"
+ "add x20, x20, x28\n"
+ ".inst 0x4f81e158 // sdot v24.4s, v10.16b, v1.4b[0]\n"
+ ".inst 0x4f81e959 // sdot v25.4s, v10.16b, v1.4b[2]\n"
+ ".inst 0x4f82e15a // sdot v26.4s, v10.16b, v2.4b[0]\n"
+ ".inst 0x4f82e95b // sdot v27.4s, v10.16b, v2.4b[2]\n"
+ ".inst 0x4fa2e13c // sdot v28.4s, v9.16b, v2.4b[1]\n"
+ ".inst 0x4fa2e93d // sdot v29.4s, v9.16b, v2.4b[3]\n"
+ ".inst 0x4fa3e13e // sdot v30.4s, v9.16b, v3.4b[1]\n"
+ ".inst 0x4fa3e93f // sdot v31.4s, v9.16b, v3.4b[3]\n"
"ldr q9, [%x[params], #0x10]\n"
- ".inst 0x4f8ce15c // sdot v28.4s, v10.16b, v12.4b[0]\n"
- ".inst 0x4f8ce95b // sdot v27.4s, v10.16b, v12.4b[2]\n"
- ".inst 0x4f87e15a // sdot v26.4s, v10.16b, v7.4b[0]\n"
- ".inst 0x4f87e959 // sdot v25.4s, v10.16b, v7.4b[2]\n"
- ".inst 0x4f86e158 // sdot v24.4s, v10.16b, v6.4b[0]\n"
- ".inst 0x4f86e957 // sdot v23.4s, v10.16b, v6.4b[2]\n"
- ".inst 0x4f85e156 // sdot v22.4s, v10.16b, v5.4b[0]\n"
- ".inst 0x4f85e955 // sdot v21.4s, v10.16b, v5.4b[2]\n"
+ ".inst 0x4fa1e178 // sdot v24.4s, v11.16b, v1.4b[1]\n"
+ ".inst 0x4fa1e979 // sdot v25.4s, v11.16b, v1.4b[3]\n"
+ ".inst 0x4fa2e17a // sdot v26.4s, v11.16b, v2.4b[1]\n"
+ ".inst 0x4fa2e97b // sdot v27.4s, v11.16b, v2.4b[3]\n"
+ ".inst 0x4f83e15c // sdot v28.4s, v10.16b, v3.4b[0]\n"
+ ".inst 0x4f83e95d // sdot v29.4s, v10.16b, v3.4b[2]\n"
+ ".inst 0x4f84e15e // sdot v30.4s, v10.16b, v4.4b[0]\n"
+ ".inst 0x4f84e95f // sdot v31.4s, v10.16b, v4.4b[2]\n"
"ldr q10, [%x[params], #0x20]\n"
- ".inst 0x4face17c // sdot v28.4s, v11.16b, v12.4b[1]\n"
- ".inst 0x4face97b // sdot v27.4s, v11.16b, v12.4b[3]\n"
- ".inst 0x4fa7e17a // sdot v26.4s, v11.16b, v7.4b[1]\n"
- ".inst 0x4fa7e979 // sdot v25.4s, v11.16b, v7.4b[3]\n"
- ".inst 0x4fa6e178 // sdot v24.4s, v11.16b, v6.4b[1]\n"
- ".inst 0x4fa6e977 // sdot v23.4s, v11.16b, v6.4b[3]\n"
- ".inst 0x4fa5e176 // sdot v22.4s, v11.16b, v5.4b[1]\n"
- ".inst 0x4fa5e975 // sdot v21.4s, v11.16b, v5.4b[3]\n"
+ ".inst 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
+ ".inst 0x4f82e919 // sdot v25.4s, v8.16b, v2.4b[2]\n"
+ ".inst 0x4f83e11a // sdot v26.4s, v8.16b, v3.4b[0]\n"
+ ".inst 0x4f83e91b // sdot v27.4s, v8.16b, v3.4b[2]\n"
+ ".inst 0x4fa3e17c // sdot v28.4s, v11.16b, v3.4b[1]\n"
+ ".inst 0x4fa3e97d // sdot v29.4s, v11.16b, v3.4b[3]\n"
+ ".inst 0x4fa4e17e // sdot v30.4s, v11.16b, v4.4b[1]\n"
+ ".inst 0x4fa4e97f // sdot v31.4s, v11.16b, v4.4b[3]\n"
"ldr q11, [%x[params], #0x30]\n"
- ".inst 0x4f87e11c // sdot v28.4s, v8.16b, v7.4b[0]\n"
- ".inst 0x4f87e91b // sdot v27.4s, v8.16b, v7.4b[2]\n"
- ".inst 0x4f86e11a // sdot v26.4s, v8.16b, v6.4b[0]\n"
- ".inst 0x4f86e919 // sdot v25.4s, v8.16b, v6.4b[2]\n"
- ".inst 0x4f85e118 // sdot v24.4s, v8.16b, v5.4b[0]\n"
- ".inst 0x4f85e917 // sdot v23.4s, v8.16b, v5.4b[2]\n"
- ".inst 0x4f84e116 // sdot v22.4s, v8.16b, v4.4b[0]\n"
- ".inst 0x4f84e915 // sdot v21.4s, v8.16b, v4.4b[2]\n"
+ ".inst 0x4fa2e138 // sdot v24.4s, v9.16b, v2.4b[1]\n"
+ ".inst 0x4fa2e939 // sdot v25.4s, v9.16b, v2.4b[3]\n"
+ ".inst 0x4fa3e13a // sdot v26.4s, v9.16b, v3.4b[1]\n"
+ ".inst 0x4fa3e93b // sdot v27.4s, v9.16b, v3.4b[3]\n"
+ ".inst 0x4f84e11c // sdot v28.4s, v8.16b, v4.4b[0]\n"
+ ".inst 0x4f84e91d // sdot v29.4s, v8.16b, v4.4b[2]\n"
+ ".inst 0x4f85e11e // sdot v30.4s, v8.16b, v5.4b[0]\n"
+ ".inst 0x4f85e91f // sdot v31.4s, v8.16b, v5.4b[2]\n"
"ldr q8, [%x[params], #0x40]\n"
- ".inst 0x4fa7e13c // sdot v28.4s, v9.16b, v7.4b[1]\n"
- ".inst 0x4fa7e93b // sdot v27.4s, v9.16b, v7.4b[3]\n"
- ".inst 0x4fa6e13a // sdot v26.4s, v9.16b, v6.4b[1]\n"
- ".inst 0x4fa6e939 // sdot v25.4s, v9.16b, v6.4b[3]\n"
- ".inst 0x4fa5e138 // sdot v24.4s, v9.16b, v5.4b[1]\n"
- ".inst 0x4fa5e937 // sdot v23.4s, v9.16b, v5.4b[3]\n"
- ".inst 0x4fa4e136 // sdot v22.4s, v9.16b, v4.4b[1]\n"
- ".inst 0x4fa4e935 // sdot v21.4s, v9.16b, v4.4b[3]\n"
+ ".inst 0x4f83e158 // sdot v24.4s, v10.16b, v3.4b[0]\n"
+ ".inst 0x4f83e959 // sdot v25.4s, v10.16b, v3.4b[2]\n"
+ ".inst 0x4f84e15a // sdot v26.4s, v10.16b, v4.4b[0]\n"
+ ".inst 0x4f84e95b // sdot v27.4s, v10.16b, v4.4b[2]\n"
+ ".inst 0x4fa4e13c // sdot v28.4s, v9.16b, v4.4b[1]\n"
+ ".inst 0x4fa4e93d // sdot v29.4s, v9.16b, v4.4b[3]\n"
+ ".inst 0x4fa5e13e // sdot v30.4s, v9.16b, v5.4b[1]\n"
+ ".inst 0x4fa5e93f // sdot v31.4s, v9.16b, v5.4b[3]\n"
"ldr q9, [%x[params], #0x50]\n"
"add %x[params], %x[params], #0x80\n"
- ".inst 0x4f86e15c // sdot v28.4s, v10.16b, v6.4b[0]\n"
- ".inst 0x4f86e95b // sdot v27.4s, v10.16b, v6.4b[2]\n"
- ".inst 0x4f85e15a // sdot v26.4s, v10.16b, v5.4b[0]\n"
- ".inst 0x4f85e959 // sdot v25.4s, v10.16b, v5.4b[2]\n"
- ".inst 0x4f84e158 // sdot v24.4s, v10.16b, v4.4b[0]\n"
- ".inst 0x4f84e957 // sdot v23.4s, v10.16b, v4.4b[2]\n"
- ".inst 0x4f82e156 // sdot v22.4s, v10.16b, v2.4b[0]\n"
- ".inst 0x4f82e955 // sdot v21.4s, v10.16b, v2.4b[2]\n"
- ".inst 0x4fa6e17c // sdot v28.4s, v11.16b, v6.4b[1]\n"
- ".inst 0x4fa6e97b // sdot v27.4s, v11.16b, v6.4b[3]\n"
- ".inst 0x4fa5e17a // sdot v26.4s, v11.16b, v5.4b[1]\n"
- ".inst 0x4fa5e979 // sdot v25.4s, v11.16b, v5.4b[3]\n"
- ".inst 0x4fa4e178 // sdot v24.4s, v11.16b, v4.4b[1]\n"
- ".inst 0x4fa4e977 // sdot v23.4s, v11.16b, v4.4b[3]\n"
- ".inst 0x4fa2e176 // sdot v22.4s, v11.16b, v2.4b[1]\n"
- ".inst 0x4fa2e975 // sdot v21.4s, v11.16b, v2.4b[3]\n"
- ".inst 0x4f85e11c // sdot v28.4s, v8.16b, v5.4b[0]\n"
- ".inst 0x4f85e91b // sdot v27.4s, v8.16b, v5.4b[2]\n"
- ".inst 0x4f84e11a // sdot v26.4s, v8.16b, v4.4b[0]\n"
+ ".inst 0x4fa3e178 // sdot v24.4s, v11.16b, v3.4b[1]\n"
+ ".inst 0x4fa3e979 // sdot v25.4s, v11.16b, v3.4b[3]\n"
+ ".inst 0x4fa4e17a // sdot v26.4s, v11.16b, v4.4b[1]\n"
+ ".inst 0x4fa4e97b // sdot v27.4s, v11.16b, v4.4b[3]\n"
+ ".inst 0x4f85e15c // sdot v28.4s, v10.16b, v5.4b[0]\n"
+ ".inst 0x4f85e95d // sdot v29.4s, v10.16b, v5.4b[2]\n"
+ ".inst 0x4f86e15e // sdot v30.4s, v10.16b, v6.4b[0]\n"
+ ".inst 0x4f86e95f // sdot v31.4s, v10.16b, v6.4b[2]\n"
+ ".inst 0x4f84e118 // sdot v24.4s, v8.16b, v4.4b[0]\n"
".inst 0x4f84e919 // sdot v25.4s, v8.16b, v4.4b[2]\n"
- ".inst 0x4f82e118 // sdot v24.4s, v8.16b, v2.4b[0]\n"
- ".inst 0x4f82e917 // sdot v23.4s, v8.16b, v2.4b[2]\n"
- ".inst 0x4f81e116 // sdot v22.4s, v8.16b, v1.4b[0]\n"
- ".inst 0x4f81e915 // sdot v21.4s, v8.16b, v1.4b[2]\n"
- ".inst 0x4fa5e13c // sdot v28.4s, v9.16b, v5.4b[1]\n"
- ".inst 0x4fa5e93b // sdot v27.4s, v9.16b, v5.4b[3]\n"
- ".inst 0x4fa4e13a // sdot v26.4s, v9.16b, v4.4b[1]\n"
+ ".inst 0x4f85e11a // sdot v26.4s, v8.16b, v5.4b[0]\n"
+ ".inst 0x4f85e91b // sdot v27.4s, v8.16b, v5.4b[2]\n"
+ ".inst 0x4fa5e17c // sdot v28.4s, v11.16b, v5.4b[1]\n"
+ ".inst 0x4fa5e97d // sdot v29.4s, v11.16b, v5.4b[3]\n"
+ ".inst 0x4fa6e17e // sdot v30.4s, v11.16b, v6.4b[1]\n"
+ ".inst 0x4fa6e97f // sdot v31.4s, v11.16b, v6.4b[3]\n"
+ ".inst 0x4fa4e138 // sdot v24.4s, v9.16b, v4.4b[1]\n"
".inst 0x4fa4e939 // sdot v25.4s, v9.16b, v4.4b[3]\n"
- ".inst 0x4fa2e138 // sdot v24.4s, v9.16b, v2.4b[1]\n"
- ".inst 0x4fa2e937 // sdot v23.4s, v9.16b, v2.4b[3]\n"
- ".inst 0x4fa1e136 // sdot v22.4s, v9.16b, v1.4b[1]\n"
- ".inst 0x4fa1e935 // sdot v21.4s, v9.16b, v1.4b[3]\n"
- "sqrdmulh v28.4s, v28.4s, v20.4s\n"
- "sqrdmulh v27.4s, v27.4s, v20.4s\n"
- "sqrdmulh v26.4s, v26.4s, v20.4s\n"
- "sqrdmulh v25.4s, v25.4s, v20.4s\n"
- "and v18.16b, v28.16b, v19.16b\n"
- "and v17.16b, v27.16b, v19.16b\n"
- "and v16.16b, v26.16b, v19.16b\n"
+ "sqrdmulh v24.4s, v24.4s, v21.4s\n"
+ ".inst 0x4fa5e13a // sdot v26.4s, v9.16b, v5.4b[1]\n"
+ ".inst 0x4fa5e93b // sdot v27.4s, v9.16b, v5.4b[3]\n"
+ "sqrdmulh v25.4s, v25.4s, v21.4s\n"
+ ".inst 0x4f86e11c // sdot v28.4s, v8.16b, v6.4b[0]\n"
+ ".inst 0x4f86e91d // sdot v29.4s, v8.16b, v6.4b[2]\n"
+ "sqrdmulh v26.4s, v26.4s, v21.4s\n"
+ ".inst 0x4f87e11e // sdot v30.4s, v8.16b, v7.4b[0]\n"
+ ".inst 0x4f87e91f // sdot v31.4s, v8.16b, v7.4b[2]\n"
+ "sqrdmulh v27.4s, v27.4s, v21.4s\n"
+ ".inst 0x4fa6e13c // sdot v28.4s, v9.16b, v6.4b[1]\n"
+ ".inst 0x4fa6e93d // sdot v29.4s, v9.16b, v6.4b[3]\n"
+ "and v19.16b, v24.16b, v20.16b\n"
+ ".inst 0x4fa7e13e // sdot v30.4s, v9.16b, v7.4b[1]\n"
+ ".inst 0x4fa7e93f // sdot v31.4s, v9.16b, v7.4b[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 v28.4s, v28.4s, v18.4s\n"
- "sqadd v27.4s, v27.4s, v17.4s\n"
- "sqadd v26.4s, v26.4s, v16.4s\n"
- "and v16.16b, v25.16b, v19.16b\n"
- "srshl v28.4s, v28.4s, v19.4s\n"
- "srshl v27.4s, v27.4s, v19.4s\n"
- "srshl v26.4s, v26.4s, v19.4s\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "add v28.4s, v28.4s, v0.4s\n"
- "add v27.4s, v27.4s, v0.4s\n"
- "add v26.4s, v26.4s, v0.4s\n"
- "smin v28.4s, v28.4s, v30.4s\n"
- "smin v27.4s, v27.4s, v30.4s\n"
- "smin v26.4s, v26.4s, v30.4s\n"
- "smax v28.4s, v28.4s, v31.4s\n"
- "smax v27.4s, v27.4s, v31.4s\n"
- "smax v26.4s, v26.4s, v31.4s\n"
- "uzp1 v28.16b, v28.16b, v28.16b\n"
- "uzp1 v27.16b, v27.16b, v27.16b\n"
- "uzp1 v28.16b, v28.16b, v28.16b\n"
- "uzp1 v27.16b, v27.16b, v27.16b\n"
- "uzp1 v26.16b, v26.16b, v26.16b\n"
- "sqadd v25.4s, v25.4s, v16.4s\n"
- "uzp1 v26.16b, v26.16b, v26.16b\n"
- "sqrdmulh v24.4s, v24.4s, v20.4s\n"
- "sqrdmulh v23.4s, v23.4s, v20.4s\n"
- "srshl v25.4s, v25.4s, v19.4s\n"
- "sqrdmulh v22.4s, v22.4s, v20.4s\n"
- "and v16.16b, v24.16b, v19.16b\n"
- "and v17.16b, v23.16b, v19.16b\n"
- "add v25.4s, v25.4s, v0.4s\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, v30.4s\n"
- "sqadd v24.4s, v24.4s, v16.4s\n"
- "sqadd v23.4s, v23.4s, v17.4s\n"
- "smax v25.4s, v25.4s, v31.4s\n"
- "and v16.16b, v22.16b, v19.16b\n"
- "srshl v24.4s, v24.4s, v19.4s\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "srshl v23.4s, v23.4s, v19.4s\n"
- "uzp1 v25.16b, v25.16b, v25.16b\n"
- "add v24.4s, v24.4s, v0.4s\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "add v23.4s, v23.4s, v0.4s\n"
- "smin v24.4s, v24.4s, v30.4s\n"
- "sqadd v22.4s, v22.4s, v16.4s\n"
- "smin v23.4s, v23.4s, v30.4s\n"
- "smax v24.4s, v24.4s, v31.4s\n"
- "sqrdmulh v21.4s, v21.4s, v20.4s\n"
- "smax v23.4s, v23.4s, v31.4s\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, 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"
+ "smin v24.4s, v24.4s, v15.4s\n"
+ "smin v25.4s, v25.4s, v15.4s\n"
+ "smin v26.4s, v26.4s, v15.4s\n"
+ "smin v27.4s, v27.4s, v15.4s\n"
+ "smin v28.4s, v28.4s, v15.4s\n"
+ "smin v29.4s, v29.4s, v15.4s\n"
+ "smin v30.4s, v30.4s, v15.4s\n"
+ "smin v31.4s, v31.4s, v15.4s\n"
+ "smax v24.4s, v24.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v26.4s, v26.4s, v13.4s\n"
+ "smax v27.4s, v27.4s, v13.4s\n"
+ "smax v28.4s, v28.4s, v13.4s\n"
+ "smax v29.4s, v29.4s, v13.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v31.4s, v31.4s, v13.4s\n"
"uzp1 v24.16b, v24.16b, v24.16b\n"
- "srshl v22.4s, v22.4s, v19.4s\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\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"
"uzp1 v24.16b, v24.16b, v24.16b\n"
- "uzp1 v23.16b, v23.16b, v23.16b\n"
- "and v16.16b, v21.16b, v19.16b\n"
- "uzp1 v23.16b, v23.16b, v23.16b\n"
- "add v22.4s, v22.4s, v0.4s\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "smin v22.4s, v22.4s, v30.4s\n"
- "sqadd v21.4s, v21.4s, v16.4s\n"
- "smax v22.4s, v22.4s, v31.4s\n"
- "srshl v21.4s, v21.4s, v19.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "add v21.4s, v21.4s, v0.4s\n"
- "smin v21.4s, v21.4s, v30.4s\n"
- "smax v21.4s, v21.4s, v31.4s\n"
- "uzp1 v21.16b, v21.16b, v21.16b\n"
- "uzp1 v21.16b, v21.16b, v21.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\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 s28, [x26, #0x0]\n"
- "str s27, [x25, #0x0]\n"
- "str s26, [x24, #0x0]\n"
- "str s25, [x23, #0x0]\n"
- "str s24, [x22, #0x0]\n"
- "str s23, [x21, #0x0]\n"
- "str s22, [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 { v28.b }[0], [x26], #0x1\n"
"subs %x[n_channels], %x[n_channels], #0x1\n"
- "st1 { v27.b }[0], [x25], #0x1\n"
- "st1 { v26.b }[0], [x24], #0x1\n"
- "st1 { v25.b }[0], [x23], #0x1\n"
- "st1 { v24.b }[0], [x22], #0x1\n"
- "st1 { v23.b }[0], [x21], #0x1\n"
- "st1 { v22.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 { v28.b }[1], [x26], #0x1\n"
"subs %x[n_channels], %x[n_channels], #0x1\n"
- "st1 { v27.b }[1], [x25], #0x1\n"
- "st1 { v26.b }[1], [x24], #0x1\n"
- "st1 { v25.b }[1], [x23], #0x1\n"
- "st1 { v24.b }[1], [x22], #0x1\n"
- "st1 { v23.b }[1], [x21], #0x1\n"
- "st1 { v22.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 { v28.b }[2], [x26], #0x1\n"
"subs %x[n_channels], %x[n_channels], #0x1\n"
- "st1 { v27.b }[2], [x25], #0x1\n"
- "st1 { v26.b }[2], [x24], #0x1\n"
- "st1 { v25.b }[2], [x23], #0x1\n"
- "st1 { v24.b }[2], [x22], #0x1\n"
- "st1 { v23.b }[2], [x21], #0x1\n"
- "st1 { v22.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 { v28.b }[3], [x26], #0x1\n"
+ "st1 { v24.b }[3], [x27], #0x1\n"
"subs %x[n_channels], %x[n_channels], #0x1\n"
- "st1 { v27.b }[3], [x25], #0x1\n"
- "st1 { v26.b }[3], [x24], #0x1\n"
- "st1 { v25.b }[3], [x23], #0x1\n"
- "st1 { v24.b }[3], [x22], #0x1\n"
- "st1 { v23.b }[3], [x21], #0x1\n"
- "st1 { v22.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"
);
}