aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp
diff options
context:
space:
mode:
authorMichael Tyler <michael.tyler@arm.com>2022-12-15 12:39:29 +0000
committermichael.tyler <michael.tyler@arm.com>2023-01-16 09:31:00 +0000
commitba209750abc1ac7e42bab9fef5db284384d70fb3 (patch)
tree1065f242db9a9a5e48bd4a9f2fd68aef1924827a /src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp
parent8094f9dd5307c55f545b2cb41ec80a739a9b4d6f (diff)
downloadComputeLibrary-ba209750abc1ac7e42bab9fef5db284384d70fb3.tar.gz
Update CPU kernels to remove x19
Resolves: COMPMID-5805 Signed-off-by: Michael Tyler <michael.tyler@arm.com> Change-Id: I250f64531e209625e4ff176dd5a552c1c34bc484 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8909 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> Benchmark: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp2738
1 files changed, 1539 insertions, 1199 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp
index f8245fc5d9..fda88f94bb 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021-2022 Arm Limited.
+ * Copyright (c) 2021-2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -41,1282 +41,1622 @@ void a64_s8q_nhwc_3x3_s1_output2x2_dot_depthfirst_impl(
)
{
__asm__ __volatile__(
- "ldp x13, x12, [%x[inptrs], #0x0]\n"
- "add SP, SP, #-0x80\n"
- "ldp x11, x10, [%x[inptrs], #0x10]\n"
- "mov x19, #0x1\n"
- "ldp x9, x28, [%x[inptrs], #0x20]\n"
- "orr x19, x19, #0x100\n"
- "ldp x27, x26, [%x[inptrs], #0x30]\n"
- "orr x19, x19, #0x10000\n"
- "dup v11.4s, w19\n"
- "ldp x25, x24, [%x[outptrs], #0x0]\n"
- "mov x23, #0x0\n"
+ "mov x20, #0x1\n"
+ "orr x20, x20, #0x100\n"
+ "ldp x15, x14, [%x[inptrs], #0x0]\n"
+ "ldp x13, x12, [%x[inptrs], #0x10]\n"
+ "orr x20, x20, #0x10000\n"
+ "lsr x11, %x[n_channels], #0x4\n"
+ "dup v14.4s, w20\n"
+ "ldp x10, x9, [%x[inptrs], #0x20]\n"
+ "add x20, %x[qp], %[offsetof_Requantize32_minval]\n"
+ "ld1r { v13.4s }, [x20]\n"
+ "add x20, %x[qp], %[offsetof_Requantize32_maxval]\n"
+ "ld1r { v12.4s }, [x20]\n"
+ "add x20, %x[qp], %[offsetof_Requantize32_b_offset]\n"
+ "ld1r { v11.4s }, [x20]\n"
+ "add x20, %x[qp], %[offsetof_Requantize32_c_offset]\n"
+ "ld1r { v10.4s }, [x20]\n"
+ "mov x28, #0x0\n"
+ "mov x27, #0x0\n"
+ "ldp x26, x25, [%x[inptrs], #0x30]\n"
+ "ldp x24, x23, [%x[outptrs], #0x0]\n"
"ldp x22, x21, [%x[outptrs], #0x10]\n"
- "lsr x20, %x[n_channels], #0x4\n"
- "add x19, %x[qp], %[offsetof_Requantize32_minval]\n"
- "ld1r { v9.4s }, [x19]\n"
- "add x19, %x[qp], %[offsetof_Requantize32_maxval]\n"
- "ld1r { v12.4s }, [x19]\n"
- "add x19, %x[qp], %[offsetof_Requantize32_b_offset]\n"
- "ld1r { v14.4s }, [x19]\n"
- "add x19, %x[qp], %[offsetof_Requantize32_c_offset]\n"
- "ld1r { v13.4s }, [x19]\n"
- "cbz x20, 2f\n"
+ "cbz x11, 3f\n"
+ "ldr q9, [x15, x28]\n"
+ "ldr q8, [x14, x28]\n"
+ "subs x11, x11, #0x1\n"
+ "ldr q7, [x13, x28]\n"
+ "ldr q6, [x12, x28]\n"
+ "zip2 v5.16b, v9.16b, v7.16b\n"
+ "zip1 v9.16b, v9.16b, v7.16b\n"
+ "ldr q4, [x10, x28]\n"
+ "ldr q3, [x9, x28]\n"
+ "zip1 v7.16b, v8.16b, v6.16b\n"
+ "zip2 v6.16b, v8.16b, v6.16b\n"
+ "ldr q2, [x26, x28]\n"
+ "ldr q1, [x25, x28]\n"
+ "zip2 v8.16b, v9.16b, v7.16b\n"
+ "zip1 v9.16b, v9.16b, v7.16b\n"
+ "ldr q0, [%x[params], #0x10]\n"
+ "ldr q16, [%x[params], #0x20]\n"
+ "zip1 v7.16b, v5.16b, v6.16b\n"
+ "zip2 v6.16b, v5.16b, v6.16b\n"
+ "ldr q5, [%x[params], #0x0]\n"
+ "ldr q31, [%x[params], #0x30]\n"
+ "zip2 v30.16b, v4.16b, v2.16b\n"
+ "zip1 v4.16b, v4.16b, v2.16b\n"
+ "ldp x15, x14, [%x[inptrs], #0x40]\n"
+ "ldr q29, [x15, x28]\n"
+ "zip1 v2.16b, v3.16b, v1.16b\n"
+ "zip2 v1.16b, v3.16b, v1.16b\n"
+ "ldr q28, [x14, x28]\n"
+ "ldp x13, x12, [%x[inptrs], #0x50]\n"
+ "zip2 v3.16b, v4.16b, v2.16b\n"
+ "zip1 v4.16b, v4.16b, v2.16b\n"
+ "ldr q27, [x13, x28]\n"
+ "ldr q26, [x12, x28]\n"
+ "zip2 v25.16b, v29.16b, v27.16b\n"
+ "zip1 v29.16b, v29.16b, v27.16b\n"
+ "ldp x10, x9, [%x[inptrs], #0x60]\n"
+ "ldr q24, [x10, x28]\n"
+ "zip1 v27.16b, v28.16b, v26.16b\n"
+ "zip2 v26.16b, v28.16b, v26.16b\n"
+ "ldr q23, [x9, x28]\n"
+ "ldp x26, x25, [%x[inptrs], #0x70]\n"
+ "zip1 v2.16b, v30.16b, v1.16b\n"
+ "zip2 v1.16b, v30.16b, v1.16b\n"
+ "ldr q22, [x26, x28]\n"
+ "ldr q21, [x25, x28]\n"
+ "zip2 v20.16b, v24.16b, v22.16b\n"
+ "zip1 v24.16b, v24.16b, v22.16b\n"
+ "zip1 v22.16b, v23.16b, v21.16b\n"
+ "zip2 v21.16b, v23.16b, v21.16b\n"
+ "ldp x15, x14, [%x[inptrs], #0x0]\n"
+ "ldp x13, x12, [%x[inptrs], #0x10]\n"
+ "ldp x10, x9, [%x[inptrs], #0x20]\n"
+ "ldp x26, x25, [%x[inptrs], #0x30]\n"
+ "zip2 v28.16b, v29.16b, v27.16b\n"
+ "zip1 v29.16b, v29.16b, v27.16b\n"
+ "zip1 v27.16b, v25.16b, v26.16b\n"
+ "zip2 v26.16b, v25.16b, v26.16b\n"
+ "add %x[params], %x[params], #0x40\n"
+ "zip2 v23.16b, v24.16b, v22.16b\n"
+ "zip1 v24.16b, v24.16b, v22.16b\n"
+ "zip1 v22.16b, v20.16b, v21.16b\n"
+ "zip2 v21.16b, v20.16b, v21.16b\n"
+ "mov v30.16b, v5.16b\n"
+ "mov v25.16b, v5.16b\n"
+ "mov v20.16b, v5.16b\n"
+ "beq 2f\n"
"1:" // Loop
- "movi v15.4s, #0x0\n"
- "ldr q27, [x13, x23]\n"
- "subs x20, x20, #0x1\n"
- "movi v10.4s, #0x0\n"
- "ldr q1, [x12, x23]\n"
- "ldp x13, x12, [%x[inptrs], #0x40]\n"
- "ldr q25, [x11, x23]\n"
- "zip1 v7.16b, v27.16b, v25.16b\n"
- "ldr q23, [x10, x23]\n"
- "zip2 v5.16b, v27.16b, v25.16b\n"
- "ldp x11, x10, [%x[inptrs], #0x50]\n"
- "ldr q31, [x9, x23]\n"
- "zip1 v8.16b, v1.16b, v23.16b\n"
- "ldr q28, [x28, x23]\n"
- "zip2 v3.16b, v1.16b, v23.16b\n"
- "ldp x9, x28, [%x[inptrs], #0x60]\n"
- "zip1 v6.16b, v7.16b, v8.16b\n"
- "ldr q21, [x27, x23]\n"
- "zip2 v8.16b, v7.16b, v8.16b\n"
- "ldr q26, [x26, x23]\n"
- "zip1 v7.16b, v5.16b, v3.16b\n"
- "ldp x27, x26, [%x[inptrs], #0x70]\n"
- "zip2 v5.16b, v5.16b, v3.16b\n"
- "ldr q24, [x13, x23]\n"
- "ldr q22, [x12, x23]\n"
- "zip1 v2.16b, v31.16b, v21.16b\n"
- "zip2 v4.16b, v31.16b, v21.16b\n"
- "ldp x13, x12, [%x[inptrs], #0x0]\n"
- "zip1 v1.16b, v28.16b, v26.16b\n"
- "ldr q20, [x11, x23]\n"
- "zip2 v31.16b, v28.16b, v26.16b\n"
- "ldr q16, [x10, x23]\n"
- "zip1 v3.16b, v2.16b, v1.16b\n"
- "ldp x11, x10, [%x[inptrs], #0x10]\n"
- "zip2 v2.16b, v2.16b, v1.16b\n"
- "ldr q19, [x9, x23]\n"
- "zip1 v1.16b, v4.16b, v31.16b\n"
- "ldr q0, [x28, x23]\n"
- "zip1 v28.16b, v24.16b, v20.16b\n"
- "ldp x9, x28, [%x[inptrs], #0x20]\n"
- "zip2 v26.16b, v24.16b, v20.16b\n"
- "ldr q18, [x27, x23]\n"
- "zip1 v24.16b, v22.16b, v16.16b\n"
- "ldr q17, [x26, x23]\n"
- "zip2 v22.16b, v22.16b, v16.16b\n"
- "ldp x27, x26, [%x[inptrs], #0x30]\n"
- "zip2 v16.16b, v4.16b, v31.16b\n"
- "str q7, [SP, #0x0]\n"
- "zip1 v31.16b, v28.16b, v24.16b\n"
- "str q5, [SP, #0x10]\n"
- "zip1 v20.16b, v19.16b, v18.16b\n"
- "str q1, [SP, #0x20]\n"
- "zip2 v19.16b, v19.16b, v18.16b\n"
- "str q16, [SP, #0x30]\n"
- "zip1 v18.16b, v0.16b, v17.16b\n"
- "ldr q30, [%x[params], #0x0]\n"
- "zip2 v17.16b, v0.16b, v17.16b\n"
- "ldr q29, [%x[params], #0x10]\n"
- "zip2 v28.16b, v28.16b, v24.16b\n"
- "ldr q27, [%x[params], #0x20]\n"
- "zip1 v16.16b, v26.16b, v22.16b\n"
- "str q16, [SP, #0x40]\n"
- "zip2 v16.16b, v26.16b, v22.16b\n"
- "str q16, [SP, #0x50]\n"
- "zip1 v26.16b, v20.16b, v18.16b\n"
- "ldr q25, [%x[params], #0x30]\n"
- "zip2 v24.16b, v20.16b, v18.16b\n"
- "ldr q23, [%x[params], #0x40]\n"
- "zip1 v16.16b, v19.16b, v17.16b\n"
- "str q16, [SP, #0x60]\n"
- "zip2 v16.16b, v19.16b, v17.16b\n"
- "str q16, [SP, #0x70]\n"
- "mov v22.16b, v30.16b\n"
- "ldr q21, [%x[params], #0x50]\n"
- "mov v20.16b, v30.16b\n"
- "mov v19.16b, v30.16b\n"
- ".inst 0x4e8697be // sdot v30.4s, v29.16b, v6.16b\n"
- ".inst 0x4e8397b4 // sdot v20.4s, v29.16b, v3.16b\n"
- ".inst 0x4e83956f // sdot v15.4s, v11.16b, v3.16b\n"
- ".inst 0x4e83977e // sdot v30.4s, v27.16b, v3.16b\n"
- "ext v3.16b, v3.16b, v3.16b, #0x1\n"
- ".inst 0x4e9f9774 // sdot v20.4s, v27.16b, v31.16b\n"
- ".inst 0x4e9f956f // sdot v15.4s, v11.16b, v31.16b\n"
- ".inst 0x4e9f973e // sdot v30.4s, v25.16b, v31.16b\n"
- "ext v31.16b, v31.16b, v31.16b, #0x1\n"
- ".inst 0x4e9a9734 // sdot v20.4s, v25.16b, v26.16b\n"
- "mov v17.16b, v15.16b\n"
- ".inst 0x4e86956f // sdot v15.4s, v11.16b, v6.16b\n"
- "mls v30.4s, v15.4s, v14.4s\n"
- ".inst 0x4e9a9571 // sdot v17.4s, v11.16b, v26.16b\n"
- "ext v6.16b, v6.16b, v6.16b, #0x1\n"
- "mls v20.4s, v17.4s, v14.4s\n"
- "ext v26.16b, v26.16b, v26.16b, #0x1\n"
- ".inst 0x4e8697b6 // sdot v22.4s, v29.16b, v6.16b\n"
- ".inst 0x4e8397b3 // sdot v19.4s, v29.16b, v3.16b\n"
- "ldr q29, [%x[params], #0x70]\n"
- ".inst 0x4e83956a // sdot v10.4s, v11.16b, v3.16b\n"
- "sqrdmulh v30.4s, v30.4s, v23.4s\n"
- ".inst 0x4e839776 // sdot v22.4s, v27.16b, v3.16b\n"
- "ldr q3, [SP, #0x20]\n"
- ".inst 0x4e9f9773 // sdot v19.4s, v27.16b, v31.16b\n"
- "ldr q27, [%x[params], #0x80]\n"
- ".inst 0x4e9f956a // sdot v10.4s, v11.16b, v31.16b\n"
- "and v18.16b, v30.16b, v21.16b\n"
- "sshr v18.4s, v18.4s, #0x1f\n"
- ".inst 0x4e9f9736 // sdot v22.4s, v25.16b, v31.16b\n"
- "ldr q31, [SP, #0x40]\n"
- ".inst 0x4e9a9733 // sdot v19.4s, v25.16b, v26.16b\n"
- "ldr q25, [%x[params], #0x90]\n"
- "mov v17.16b, v10.16b\n"
- ".inst 0x4e86956a // sdot v10.4s, v11.16b, v6.16b\n"
- "ldr q6, [SP, #0x0]\n"
- "mls v22.4s, v10.4s, v14.4s\n"
- ".inst 0x4e9a9571 // sdot v17.4s, v11.16b, v26.16b\n"
- "ldr q26, [SP, #0x60]\n"
- "sqadd v30.4s, v30.4s, v18.4s\n"
- "mls v19.4s, v17.4s, v14.4s\n"
- "sqrdmulh v20.4s, v20.4s, v23.4s\n"
- "movi v15.4s, #0x0\n"
- "srshl v30.4s, v30.4s, v21.4s\n"
- "sqrdmulh v22.4s, v22.4s, v23.4s\n"
- ".inst 0x4e82956f // sdot v15.4s, v11.16b, v2.16b\n"
- "and v16.16b, v20.16b, v21.16b\n"
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8495d3 // sdot v19.4s, v14.16b, v4.16b\n"
+ ".inst 0x4e899405 // sdot v5.4s, v0.16b, v9.16b\n"
+ "add x28, x28, #0x10\n"
+ ".inst 0x4e9d95d3 // sdot v19.4s, v14.16b, v29.16b\n"
+ ".inst 0x4e849419 // sdot v25.4s, v0.16b, v4.16b\n"
+ "subs x11, x11, #0x1\n"
+ ".inst 0x4e849605 // sdot v5.4s, v16.16b, v4.16b\n"
+ "ext v4.16b, v4.16b, v4.16b, #0x1\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9895d2 // sdot v18.4s, v14.16b, v24.16b\n"
+ ".inst 0x4e8995d3 // sdot v19.4s, v14.16b, v9.16b\n"
+ "ext v9.16b, v9.16b, v9.16b, #0x1\n"
+ ".inst 0x4e9d9619 // sdot v25.4s, v16.16b, v29.16b\n"
+ ".inst 0x4e9d97e5 // sdot v5.4s, v31.16b, v29.16b\n"
+ "ext v29.16b, v29.16b, v29.16b, #0x1\n"
+ ".inst 0x4e89941e // sdot v30.4s, v0.16b, v9.16b\n"
+ ".inst 0x4e849414 // sdot v20.4s, v0.16b, v4.16b\n"
+ "movi v17.4s, #0x0\n"
+ ".inst 0x4e8495d1 // sdot v17.4s, v14.16b, v4.16b\n"
+ ".inst 0x4e9d95d1 // sdot v17.4s, v14.16b, v29.16b\n"
+ ".inst 0x4e9897f9 // sdot v25.4s, v31.16b, v24.16b\n"
+ "ext v24.16b, v24.16b, v24.16b, #0x1\n"
+ ".inst 0x4e84961e // sdot v30.4s, v16.16b, v4.16b\n"
+ "ldr q4, [%x[params], #0x10]\n"
+ ".inst 0x4e9d9614 // sdot v20.4s, v16.16b, v29.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9895d0 // sdot v16.4s, v14.16b, v24.16b\n"
+ ".inst 0x4e8995d1 // sdot v17.4s, v14.16b, v9.16b\n"
+ "ldr q9, [%x[params], #0x0]\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ ".inst 0x4e9d97fe // sdot v30.4s, v31.16b, v29.16b\n"
+ ".inst 0x4e9897f4 // sdot v20.4s, v31.16b, v24.16b\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "ldr q9, [%x[params], #0x60]\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "add v30.4s, v30.4s, v13.4s\n"
- "and v17.16b, v22.16b, v21.16b\n"
- "sshr v17.4s, v17.4s, #0x1f\n"
- "smax v30.4s, v30.4s, v9.4s\n"
- "sqrdmulh v19.4s, v19.4s, v23.4s\n"
- "ldr q23, [%x[params], #0xa0]\n"
- ".inst 0x4e9c956f // sdot v15.4s, v11.16b, v28.16b\n"
- "sqadd v20.4s, v20.4s, v16.4s\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "ldr q16, [%x[params], #0x40]\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "ldr q31, [%x[params], #0x50]\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "ldr q0, [%x[params], #0x30]\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "ldr q4, [%x[params], #0x70]\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
"smin v30.4s, v30.4s, v12.4s\n"
- "and v16.16b, v19.16b, v21.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sqadd v22.4s, v22.4s, v17.4s\n"
- "srshl v20.4s, v20.4s, v21.4s\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "mov v17.16b, v15.16b\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "str s30, [x25, x23]\n"
- "srshl v22.4s, v22.4s, v21.4s\n"
- "add v20.4s, v20.4s, v13.4s\n"
- "ldr q30, [%x[params], #0x60]\n"
- "sqadd v19.4s, v19.4s, v16.4s\n"
- ".inst 0x4e88956f // sdot v15.4s, v11.16b, v8.16b\n"
- "smax v20.4s, v20.4s, v9.4s\n"
- "add v22.4s, v22.4s, v13.4s\n"
- "srshl v19.4s, v19.4s, v21.4s\n"
- "ldr q21, [%x[params], #0xb0]\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
"smin v20.4s, v20.4s, v12.4s\n"
- "smax v22.4s, v22.4s, v9.4s\n"
- ".inst 0x4e989571 // sdot v17.4s, v11.16b, v24.16b\n"
- "add v19.4s, v19.4s, v13.4s\n"
- "smin v22.4s, v22.4s, v12.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8395d3 // sdot v19.4s, v14.16b, v3.16b\n"
+ ".inst 0x4e9c95d3 // sdot v19.4s, v14.16b, v28.16b\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "str s5, [x24, x27]\n"
+ "ldr q5, [%x[params], #0x20]\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "smax v19.4s, v19.4s, v9.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9795d2 // sdot v18.4s, v14.16b, v23.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "str s30, [x23, x27]\n"
+ ".inst 0x4e8895d3 // sdot v19.4s, v14.16b, v8.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "str s20, [x22, x23]\n"
- "smin v19.4s, v19.4s, v12.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "str s22, [x24, x23]\n"
- "mov v22.16b, v30.16b\n"
- "mov v20.16b, v30.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- ".inst 0x4e8297b4 // sdot v20.4s, v29.16b, v2.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "str s19, [x21, x23]\n"
- "mov v19.16b, v30.16b\n"
- "add x23, x23, #0x4\n"
- ".inst 0x4e8897be // sdot v30.4s, v29.16b, v8.16b\n"
- ".inst 0x4e9c9774 // sdot v20.4s, v27.16b, v28.16b\n"
+ "str s25, [x22, x27]\n"
+ "mov v30.16b, v5.16b\n"
+ "str s20, [x21, x27]\n"
+ "mov v25.16b, v5.16b\n"
+ "mov v20.16b, v5.16b\n"
+ ".inst 0x4e889405 // sdot v5.4s, v0.16b, v8.16b\n"
+ ".inst 0x4e839419 // sdot v25.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e839605 // sdot v5.4s, v16.16b, v3.16b\n"
"ext v8.16b, v8.16b, v8.16b, #0x1\n"
- "movi v10.4s, #0x0\n"
- ".inst 0x4e82977e // sdot v30.4s, v27.16b, v2.16b\n"
- ".inst 0x4e989734 // sdot v20.4s, v25.16b, v24.16b\n"
- "mls v20.4s, v17.4s, v14.4s\n"
- ".inst 0x4e9c973e // sdot v30.4s, v25.16b, v28.16b\n"
- "ext v2.16b, v2.16b, v2.16b, #0x1\n"
- "mls v30.4s, v15.4s, v14.4s\n"
+ "add x27, x27, #0x4\n"
+ "ext v3.16b, v3.16b, v3.16b, #0x1\n"
+ "movi v17.4s, #0x0\n"
+ ".inst 0x4e88941e // sdot v30.4s, v0.16b, v8.16b\n"
+ ".inst 0x4e839414 // sdot v20.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e8395d1 // sdot v17.4s, v14.16b, v3.16b\n"
+ ".inst 0x4e9c9619 // sdot v25.4s, v16.16b, v28.16b\n"
+ ".inst 0x4e9c97e5 // sdot v5.4s, v31.16b, v28.16b\n"
"ext v28.16b, v28.16b, v28.16b, #0x1\n"
- "ext v24.16b, v24.16b, v24.16b, #0x1\n"
- ".inst 0x4e8897b6 // sdot v22.4s, v29.16b, v8.16b\n"
- ".inst 0x4e8297b3 // sdot v19.4s, v29.16b, v2.16b\n"
- "ldr q29, [%x[params], #0xd0]\n"
- ".inst 0x4e82956a // sdot v10.4s, v11.16b, v2.16b\n"
- "sqrdmulh v20.4s, v20.4s, v23.4s\n"
- ".inst 0x4e829776 // sdot v22.4s, v27.16b, v2.16b\n"
- "ldr q2, [SP, #0x30]\n"
- ".inst 0x4e9c9773 // sdot v19.4s, v27.16b, v28.16b\n"
- "ldr q27, [%x[params], #0xe0]\n"
- ".inst 0x4e9c956a // sdot v10.4s, v11.16b, v28.16b\n"
- "sqrdmulh v30.4s, v30.4s, v23.4s\n"
- ".inst 0x4e9c9736 // sdot v22.4s, v25.16b, v28.16b\n"
- "ldr q28, [SP, #0x50]\n"
- ".inst 0x4e989733 // sdot v19.4s, v25.16b, v24.16b\n"
- "ldr q25, [%x[params], #0xf0]\n"
- "mov v17.16b, v10.16b\n"
- ".inst 0x4e88956a // sdot v10.4s, v11.16b, v8.16b\n"
- "ldr q8, [SP, #0x10]\n"
- "mls v22.4s, v10.4s, v14.4s\n"
- ".inst 0x4e989571 // sdot v17.4s, v11.16b, v24.16b\n"
- "ldr q24, [SP, #0x70]\n"
- "and v18.16b, v30.16b, v21.16b\n"
- "mls v19.4s, v17.4s, v14.4s\n"
- "and v16.16b, v20.16b, v21.16b\n"
- "movi v15.4s, #0x0\n"
- "sshr v18.4s, v18.4s, #0x1f\n"
- "sqrdmulh v22.4s, v22.4s, v23.4s\n"
+ ".inst 0x4e83961e // sdot v30.4s, v16.16b, v3.16b\n"
+ "ldr q3, [x9, x28]\n"
+ ".inst 0x4e9c9614 // sdot v20.4s, v16.16b, v28.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ ".inst 0x4e9c95d1 // sdot v17.4s, v14.16b, v28.16b\n"
+ ".inst 0x4e9797f9 // sdot v25.4s, v31.16b, v23.16b\n"
+ "ext v23.16b, v23.16b, v23.16b, #0x1\n"
+ ".inst 0x4e9c97fe // sdot v30.4s, v31.16b, v28.16b\n"
+ ".inst 0x4e9797f4 // sdot v20.4s, v31.16b, v23.16b\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9795d0 // sdot v16.4s, v14.16b, v23.16b\n"
+ ".inst 0x4e8895d1 // sdot v17.4s, v14.16b, v8.16b\n"
+ "ldr q8, [x14, x28]\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "ldr q9, [%x[params], #0xc0]\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- ".inst 0x4e83956f // sdot v15.4s, v11.16b, v3.16b\n"
- "movi v10.4s, #0x0\n"
- "and v17.16b, v22.16b, v21.16b\n"
- "sshr v17.4s, v17.4s, #0x1f\n"
- "sqadd v30.4s, v30.4s, v18.4s\n"
- "sqadd v20.4s, v20.4s, v16.4s\n"
- "sqrdmulh v19.4s, v19.4s, v23.4s\n"
- "ldr q23, [%x[params], #0x100]\n"
- ".inst 0x4e9f956f // sdot v15.4s, v11.16b, v31.16b\n"
- "srshl v30.4s, v30.4s, v21.4s\n"
- "srshl v20.4s, v20.4s, v21.4s\n"
- "sqadd v22.4s, v22.4s, v17.4s\n"
- "and v16.16b, v19.16b, v21.16b\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "ldr q16, [%x[params], #0xa0]\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "ldr q31, [%x[params], #0xb0]\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "ldr q0, [%x[params], #0x90]\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "ldr q4, [%x[params], #0xd0]\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
+ "smin v30.4s, v30.4s, v12.4s\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
+ "smin v20.4s, v20.4s, v12.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8295d3 // sdot v19.4s, v14.16b, v2.16b\n"
+ ".inst 0x4e9b95d3 // sdot v19.4s, v14.16b, v27.16b\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "str s5, [x24, x27]\n"
+ "ldr q5, [%x[params], #0x80]\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9695d2 // sdot v18.4s, v14.16b, v22.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "str s30, [x23, x27]\n"
+ ".inst 0x4e8795d3 // sdot v19.4s, v14.16b, v7.16b\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "str s25, [x22, x27]\n"
+ "mov v30.16b, v5.16b\n"
+ "str s20, [x21, x27]\n"
+ "mov v25.16b, v5.16b\n"
+ "mov v20.16b, v5.16b\n"
+ ".inst 0x4e879405 // sdot v5.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e829419 // sdot v25.4s, v0.16b, v2.16b\n"
+ ".inst 0x4e829605 // sdot v5.4s, v16.16b, v2.16b\n"
+ "ext v7.16b, v7.16b, v7.16b, #0x1\n"
+ "add x27, x27, #0x4\n"
+ "ext v2.16b, v2.16b, v2.16b, #0x1\n"
+ "movi v17.4s, #0x0\n"
+ ".inst 0x4e87941e // sdot v30.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e829414 // sdot v20.4s, v0.16b, v2.16b\n"
+ ".inst 0x4e8295d1 // sdot v17.4s, v14.16b, v2.16b\n"
+ ".inst 0x4e9b9619 // sdot v25.4s, v16.16b, v27.16b\n"
+ ".inst 0x4e9b97e5 // sdot v5.4s, v31.16b, v27.16b\n"
+ "ext v27.16b, v27.16b, v27.16b, #0x1\n"
+ ".inst 0x4e82961e // sdot v30.4s, v16.16b, v2.16b\n"
+ "ldr q2, [x26, x28]\n"
+ ".inst 0x4e9b9614 // sdot v20.4s, v16.16b, v27.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ ".inst 0x4e9b95d1 // sdot v17.4s, v14.16b, v27.16b\n"
+ ".inst 0x4e9697f9 // sdot v25.4s, v31.16b, v22.16b\n"
+ "ext v22.16b, v22.16b, v22.16b, #0x1\n"
+ ".inst 0x4e9b97fe // sdot v30.4s, v31.16b, v27.16b\n"
+ ".inst 0x4e9697f4 // sdot v20.4s, v31.16b, v22.16b\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9695d0 // sdot v16.4s, v14.16b, v22.16b\n"
+ ".inst 0x4e8795d1 // sdot v17.4s, v14.16b, v7.16b\n"
+ "ldr q7, [x13, x28]\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "ldr q9, [%x[params], #0x120]\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "add v30.4s, v30.4s, v13.4s\n"
- "srshl v22.4s, v22.4s, v21.4s\n"
- "add v20.4s, v20.4s, v13.4s\n"
- "mov v17.16b, v15.16b\n"
- "smax v30.4s, v30.4s, v9.4s\n"
- "add v22.4s, v22.4s, v13.4s\n"
- "smax v20.4s, v20.4s, v9.4s\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "ldr q16, [%x[params], #0x100]\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "ldr q31, [%x[params], #0x110]\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "ldr q0, [%x[params], #0xf0]\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "ldr q4, [%x[params], #0x130]\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
"smin v30.4s, v30.4s, v12.4s\n"
- "smax v22.4s, v22.4s, v9.4s\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
"smin v20.4s, v20.4s, v12.4s\n"
- "sqadd v19.4s, v19.4s, v16.4s\n"
- "smin v22.4s, v22.4s, v12.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8195d3 // sdot v19.4s, v14.16b, v1.16b\n"
+ ".inst 0x4e9a95d3 // sdot v19.4s, v14.16b, v26.16b\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
+ "str s5, [x24, x27]\n"
+ "ldr q5, [%x[params], #0xe0]\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "srshl v19.4s, v19.4s, v21.4s\n"
- "ldr q21, [%x[params], #0x110]\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9595d2 // sdot v18.4s, v14.16b, v21.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "str s30, [x25, x23]\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "str s30, [x23, x27]\n"
+ ".inst 0x4e8695d3 // sdot v19.4s, v14.16b, v6.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "ldr q30, [%x[params], #0xc0]\n"
- "add v19.4s, v19.4s, v13.4s\n"
- "str s20, [x22, x23]\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "str s22, [x24, x23]\n"
- "smax v19.4s, v19.4s, v9.4s\n"
- ".inst 0x4e86956f // sdot v15.4s, v11.16b, v6.16b\n"
- "mov v22.16b, v30.16b\n"
- "mov v20.16b, v30.16b\n"
- "smin v19.4s, v19.4s, v12.4s\n"
- ".inst 0x4e8397b4 // sdot v20.4s, v29.16b, v3.16b\n"
- ".inst 0x4e9a9571 // sdot v17.4s, v11.16b, v26.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "str s19, [x21, x23]\n"
- "mov v19.16b, v30.16b\n"
- "add x23, x23, #0x4\n"
- ".inst 0x4e8697be // sdot v30.4s, v29.16b, v6.16b\n"
- ".inst 0x4e9f9774 // sdot v20.4s, v27.16b, v31.16b\n"
+ "str s25, [x22, x27]\n"
+ "mov v30.16b, v5.16b\n"
+ "str s20, [x21, x27]\n"
+ "mov v25.16b, v5.16b\n"
+ "mov v20.16b, v5.16b\n"
+ ".inst 0x4e869405 // sdot v5.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e819419 // sdot v25.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e819605 // sdot v5.4s, v16.16b, v1.16b\n"
"ext v6.16b, v6.16b, v6.16b, #0x1\n"
- ".inst 0x4e83977e // sdot v30.4s, v27.16b, v3.16b\n"
- ".inst 0x4e9a9734 // sdot v20.4s, v25.16b, v26.16b\n"
- "mls v20.4s, v17.4s, v14.4s\n"
- ".inst 0x4e9f973e // sdot v30.4s, v25.16b, v31.16b\n"
- "ext v3.16b, v3.16b, v3.16b, #0x1\n"
- "mls v30.4s, v15.4s, v14.4s\n"
- "ext v31.16b, v31.16b, v31.16b, #0x1\n"
+ "add x27, x27, #0x4\n"
+ "ext v1.16b, v1.16b, v1.16b, #0x1\n"
+ "movi v17.4s, #0x0\n"
+ ".inst 0x4e86941e // sdot v30.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e819414 // sdot v20.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e8195d1 // sdot v17.4s, v14.16b, v1.16b\n"
+ ".inst 0x4e9a9619 // sdot v25.4s, v16.16b, v26.16b\n"
+ ".inst 0x4e9a97e5 // sdot v5.4s, v31.16b, v26.16b\n"
"ext v26.16b, v26.16b, v26.16b, #0x1\n"
- ".inst 0x4e8697b6 // sdot v22.4s, v29.16b, v6.16b\n"
- ".inst 0x4e8397b3 // sdot v19.4s, v29.16b, v3.16b\n"
- "ldr q29, [%x[params], #0x130]\n"
- ".inst 0x4e83956a // sdot v10.4s, v11.16b, v3.16b\n"
- "sqrdmulh v20.4s, v20.4s, v23.4s\n"
- ".inst 0x4e839776 // sdot v22.4s, v27.16b, v3.16b\n"
- ".inst 0x4e9f9773 // sdot v19.4s, v27.16b, v31.16b\n"
- "ldr q27, [%x[params], #0x140]\n"
- ".inst 0x4e9f956a // sdot v10.4s, v11.16b, v31.16b\n"
- "sqrdmulh v30.4s, v30.4s, v23.4s\n"
- ".inst 0x4e9f9736 // sdot v22.4s, v25.16b, v31.16b\n"
- ".inst 0x4e9a9733 // sdot v19.4s, v25.16b, v26.16b\n"
- "ldr q25, [%x[params], #0x150]\n"
- "mov v17.16b, v10.16b\n"
- ".inst 0x4e86956a // sdot v10.4s, v11.16b, v6.16b\n"
- "mls v22.4s, v10.4s, v14.4s\n"
- ".inst 0x4e9a9571 // sdot v17.4s, v11.16b, v26.16b\n"
- "and v18.16b, v30.16b, v21.16b\n"
- "sshr v18.4s, v18.4s, #0x1f\n"
- "and v16.16b, v20.16b, v21.16b\n"
- "movi v15.4s, #0x0\n"
- "mls v19.4s, v17.4s, v14.4s\n"
- ".inst 0x4e82956f // sdot v15.4s, v11.16b, v2.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sqrdmulh v22.4s, v22.4s, v23.4s\n"
- "movi v10.4s, #0x0\n"
- "sqadd v30.4s, v30.4s, v18.4s\n"
- ".inst 0x4e9c956f // sdot v15.4s, v11.16b, v28.16b\n"
- "sqrdmulh v19.4s, v19.4s, v23.4s\n"
- "ldr q23, [%x[params], #0x160]\n"
- "and v17.16b, v22.16b, v21.16b\n"
- "sshr v17.4s, v17.4s, #0x1f\n"
- "srshl v30.4s, v30.4s, v21.4s\n"
- "sqadd v20.4s, v20.4s, v16.4s\n"
- "and v16.16b, v19.16b, v21.16b\n"
+ ".inst 0x4e81961e // sdot v30.4s, v16.16b, v1.16b\n"
+ "ldr q1, [x25, x28]\n"
+ ".inst 0x4e9a9614 // sdot v20.4s, v16.16b, v26.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ ".inst 0x4e9a95d1 // sdot v17.4s, v14.16b, v26.16b\n"
+ ".inst 0x4e9597f9 // sdot v25.4s, v31.16b, v21.16b\n"
+ "ext v21.16b, v21.16b, v21.16b, #0x1\n"
+ ".inst 0x4e9a97fe // sdot v30.4s, v31.16b, v26.16b\n"
+ ".inst 0x4e9597f4 // sdot v20.4s, v31.16b, v21.16b\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9595d0 // sdot v16.4s, v14.16b, v21.16b\n"
+ ".inst 0x4e8695d1 // sdot v17.4s, v14.16b, v6.16b\n"
+ "ldr q6, [x12, x28]\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "ldr q9, [x15, x28]\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "ldp x15, x14, [%x[inptrs], #0x40]\n"
+ "ldr q29, [x15, x28]\n"
+ "ldr q28, [x14, x28]\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "add v30.4s, v30.4s, v13.4s\n"
- "srshl v20.4s, v20.4s, v21.4s\n"
- "sqadd v22.4s, v22.4s, v17.4s\n"
- "mov v17.16b, v15.16b\n"
- "smax v30.4s, v30.4s, v9.4s\n"
- "add v20.4s, v20.4s, v13.4s\n"
- "srshl v22.4s, v22.4s, v21.4s\n"
+ "ldp x13, x12, [%x[inptrs], #0x50]\n"
+ "ldr q27, [x13, x28]\n"
+ "ldr q26, [x12, x28]\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "ldr q16, [%x[params], #0x160]\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "ldr q31, [%x[params], #0x170]\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "ldr q0, [%x[params], #0x150]\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "ldr q4, [x10, x28]\n"
+ "ldp x10, x9, [%x[inptrs], #0x60]\n"
+ "ldr q24, [x10, x28]\n"
+ "ldr q23, [x9, x28]\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "ldp x26, x25, [%x[inptrs], #0x70]\n"
+ "ldr q22, [x26, x28]\n"
+ "ldr q21, [x25, x28]\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "ldp x15, x14, [%x[inptrs], #0x0]\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
+ "ldp x13, x12, [%x[inptrs], #0x10]\n"
+ "ldp x10, x9, [%x[inptrs], #0x20]\n"
"smin v30.4s, v30.4s, v12.4s\n"
- "smax v20.4s, v20.4s, v9.4s\n"
- "sqadd v19.4s, v19.4s, v16.4s\n"
- "add v22.4s, v22.4s, v13.4s\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
+ "ldp x26, x25, [%x[inptrs], #0x30]\n"
"smin v20.4s, v20.4s, v12.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "smax v22.4s, v22.4s, v9.4s\n"
- "srshl v19.4s, v19.4s, v21.4s\n"
- "ldr q21, [%x[params], #0x170]\n"
+ "str s5, [x24, x27]\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "zip2 v5.16b, v9.16b, v7.16b\n"
+ "zip1 v9.16b, v9.16b, v7.16b\n"
+ "zip1 v7.16b, v8.16b, v6.16b\n"
+ "zip2 v6.16b, v8.16b, v6.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "str s30, [x25, x23]\n"
- "smin v22.4s, v22.4s, v12.4s\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "str s30, [x23, x27]\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "ldr q30, [%x[params], #0x120]\n"
+ "str s25, [x22, x27]\n"
+ "zip2 v8.16b, v9.16b, v7.16b\n"
+ "str s20, [x21, x27]\n"
+ "zip1 v9.16b, v9.16b, v7.16b\n"
+ "zip1 v7.16b, v5.16b, v6.16b\n"
+ "add x27, x27, #0x4\n"
+ "zip2 v6.16b, v5.16b, v6.16b\n"
+ "ldr q5, [%x[params], #0x140]\n"
+ "zip2 v30.16b, v4.16b, v2.16b\n"
"add %x[params], %x[params], #0x180\n"
- "add v19.4s, v19.4s, v13.4s\n"
+ "zip1 v4.16b, v4.16b, v2.16b\n"
+ "zip1 v2.16b, v3.16b, v1.16b\n"
+ "zip2 v1.16b, v3.16b, v1.16b\n"
+ "zip2 v25.16b, v29.16b, v27.16b\n"
+ "zip1 v29.16b, v29.16b, v27.16b\n"
+ "zip1 v27.16b, v28.16b, v26.16b\n"
+ "zip2 v26.16b, v28.16b, v26.16b\n"
+ "zip2 v20.16b, v24.16b, v22.16b\n"
+ "zip1 v24.16b, v24.16b, v22.16b\n"
+ "zip1 v22.16b, v23.16b, v21.16b\n"
+ "zip2 v21.16b, v23.16b, v21.16b\n"
+ "zip2 v3.16b, v4.16b, v2.16b\n"
+ "zip1 v4.16b, v4.16b, v2.16b\n"
+ "zip1 v2.16b, v30.16b, v1.16b\n"
+ "zip2 v1.16b, v30.16b, v1.16b\n"
+ "zip2 v28.16b, v29.16b, v27.16b\n"
+ "zip1 v29.16b, v29.16b, v27.16b\n"
+ "zip1 v27.16b, v25.16b, v26.16b\n"
+ "zip2 v26.16b, v25.16b, v26.16b\n"
+ "zip2 v23.16b, v24.16b, v22.16b\n"
+ "zip1 v24.16b, v24.16b, v22.16b\n"
+ "zip1 v22.16b, v20.16b, v21.16b\n"
+ "zip2 v21.16b, v20.16b, v21.16b\n"
+ "mov v30.16b, v5.16b\n"
+ "mov v25.16b, v5.16b\n"
+ "mov v20.16b, v5.16b\n"
+ "bgt 1b\n"
+ "2:" // Detached iteration
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8495d3 // sdot v19.4s, v14.16b, v4.16b\n"
+ ".inst 0x4e899405 // sdot v5.4s, v0.16b, v9.16b\n"
+ "tst %x[n_channels], #0xf\n"
+ ".inst 0x4e9d95d3 // sdot v19.4s, v14.16b, v29.16b\n"
+ ".inst 0x4e849419 // sdot v25.4s, v0.16b, v4.16b\n"
+ "add x28, x28, #0x10\n"
+ ".inst 0x4e849605 // sdot v5.4s, v16.16b, v4.16b\n"
+ "ext v4.16b, v4.16b, v4.16b, #0x1\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9895d2 // sdot v18.4s, v14.16b, v24.16b\n"
+ ".inst 0x4e8995d3 // sdot v19.4s, v14.16b, v9.16b\n"
+ "ext v9.16b, v9.16b, v9.16b, #0x1\n"
+ ".inst 0x4e9d9619 // sdot v25.4s, v16.16b, v29.16b\n"
+ ".inst 0x4e9d97e5 // sdot v5.4s, v31.16b, v29.16b\n"
+ "ext v29.16b, v29.16b, v29.16b, #0x1\n"
+ ".inst 0x4e89941e // sdot v30.4s, v0.16b, v9.16b\n"
+ ".inst 0x4e849414 // sdot v20.4s, v0.16b, v4.16b\n"
+ "movi v17.4s, #0x0\n"
+ ".inst 0x4e8495d1 // sdot v17.4s, v14.16b, v4.16b\n"
+ ".inst 0x4e9d95d1 // sdot v17.4s, v14.16b, v29.16b\n"
+ ".inst 0x4e9897f9 // sdot v25.4s, v31.16b, v24.16b\n"
+ "ext v24.16b, v24.16b, v24.16b, #0x1\n"
+ ".inst 0x4e84961e // sdot v30.4s, v16.16b, v4.16b\n"
+ "ldr q4, [%x[params], #0x10]\n"
+ ".inst 0x4e9d9614 // sdot v20.4s, v16.16b, v29.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9895d0 // sdot v16.4s, v14.16b, v24.16b\n"
+ ".inst 0x4e8995d1 // sdot v17.4s, v14.16b, v9.16b\n"
+ "ldr q9, [%x[params], #0x0]\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ ".inst 0x4e9d97fe // sdot v30.4s, v31.16b, v29.16b\n"
+ ".inst 0x4e9897f4 // sdot v20.4s, v31.16b, v24.16b\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "ldr q9, [%x[params], #0x60]\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "ldr q16, [%x[params], #0x40]\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "ldr q31, [%x[params], #0x50]\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "ldr q0, [%x[params], #0x30]\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "ldr q4, [%x[params], #0x70]\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
+ "smin v30.4s, v30.4s, v12.4s\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
+ "smin v20.4s, v20.4s, v12.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8395d3 // sdot v19.4s, v14.16b, v3.16b\n"
+ ".inst 0x4e9c95d3 // sdot v19.4s, v14.16b, v28.16b\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "str s5, [x24, x27]\n"
+ "ldr q5, [%x[params], #0x20]\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9795d2 // sdot v18.4s, v14.16b, v23.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "str s30, [x23, x27]\n"
+ ".inst 0x4e8895d3 // sdot v19.4s, v14.16b, v8.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "str s20, [x22, x23]\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- ".inst 0x4e88956f // sdot v15.4s, v11.16b, v8.16b\n"
- "smax v19.4s, v19.4s, v9.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "str s22, [x24, x23]\n"
- "smin v19.4s, v19.4s, v12.4s\n"
- "mov v22.16b, v30.16b\n"
- "mov v20.16b, v30.16b\n"
- ".inst 0x4e8297b4 // sdot v20.4s, v29.16b, v2.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- ".inst 0x4e989571 // sdot v17.4s, v11.16b, v24.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "str s19, [x21, x23]\n"
- "mov v19.16b, v30.16b\n"
- "add x23, x23, #0x4\n"
- ".inst 0x4e8897be // sdot v30.4s, v29.16b, v8.16b\n"
- ".inst 0x4e9c9774 // sdot v20.4s, v27.16b, v28.16b\n"
+ "str s25, [x22, x27]\n"
+ "mov v30.16b, v5.16b\n"
+ "str s20, [x21, x27]\n"
+ "mov v25.16b, v5.16b\n"
+ "mov v20.16b, v5.16b\n"
+ ".inst 0x4e889405 // sdot v5.4s, v0.16b, v8.16b\n"
+ ".inst 0x4e839419 // sdot v25.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e839605 // sdot v5.4s, v16.16b, v3.16b\n"
"ext v8.16b, v8.16b, v8.16b, #0x1\n"
- ".inst 0x4e82977e // sdot v30.4s, v27.16b, v2.16b\n"
- ".inst 0x4e989734 // sdot v20.4s, v25.16b, v24.16b\n"
- "mls v20.4s, v17.4s, v14.4s\n"
- ".inst 0x4e9c973e // sdot v30.4s, v25.16b, v28.16b\n"
- "ext v2.16b, v2.16b, v2.16b, #0x1\n"
- "mls v30.4s, v15.4s, v14.4s\n"
+ "add x27, x27, #0x4\n"
+ "ext v3.16b, v3.16b, v3.16b, #0x1\n"
+ "movi v17.4s, #0x0\n"
+ ".inst 0x4e88941e // sdot v30.4s, v0.16b, v8.16b\n"
+ ".inst 0x4e839414 // sdot v20.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e8395d1 // sdot v17.4s, v14.16b, v3.16b\n"
+ ".inst 0x4e9c9619 // sdot v25.4s, v16.16b, v28.16b\n"
+ ".inst 0x4e9c97e5 // sdot v5.4s, v31.16b, v28.16b\n"
"ext v28.16b, v28.16b, v28.16b, #0x1\n"
- "ext v24.16b, v24.16b, v24.16b, #0x1\n"
- ".inst 0x4e8897b6 // sdot v22.4s, v29.16b, v8.16b\n"
- ".inst 0x4e8297b3 // sdot v19.4s, v29.16b, v2.16b\n"
- ".inst 0x4e82956a // sdot v10.4s, v11.16b, v2.16b\n"
- "sqrdmulh v20.4s, v20.4s, v23.4s\n"
- ".inst 0x4e829776 // sdot v22.4s, v27.16b, v2.16b\n"
- ".inst 0x4e9c9773 // sdot v19.4s, v27.16b, v28.16b\n"
- ".inst 0x4e9c956a // sdot v10.4s, v11.16b, v28.16b\n"
- "sqrdmulh v30.4s, v30.4s, v23.4s\n"
- ".inst 0x4e9c9736 // sdot v22.4s, v25.16b, v28.16b\n"
- ".inst 0x4e989733 // sdot v19.4s, v25.16b, v24.16b\n"
- "mov v17.16b, v10.16b\n"
- ".inst 0x4e88956a // sdot v10.4s, v11.16b, v8.16b\n"
- "mls v22.4s, v10.4s, v14.4s\n"
- ".inst 0x4e989571 // sdot v17.4s, v11.16b, v24.16b\n"
- "and v18.16b, v30.16b, v21.16b\n"
- "sshr v18.4s, v18.4s, #0x1f\n"
- "and v16.16b, v20.16b, v21.16b\n"
- "mls v19.4s, v17.4s, v14.4s\n"
- "sqrdmulh v22.4s, v22.4s, v23.4s\n"
+ ".inst 0x4e83961e // sdot v30.4s, v16.16b, v3.16b\n"
+ ".inst 0x4e9c9614 // sdot v20.4s, v16.16b, v28.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ ".inst 0x4e9c95d1 // sdot v17.4s, v14.16b, v28.16b\n"
+ ".inst 0x4e9797f9 // sdot v25.4s, v31.16b, v23.16b\n"
+ "ext v23.16b, v23.16b, v23.16b, #0x1\n"
+ ".inst 0x4e9c97fe // sdot v30.4s, v31.16b, v28.16b\n"
+ ".inst 0x4e9797f4 // sdot v20.4s, v31.16b, v23.16b\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9795d0 // sdot v16.4s, v14.16b, v23.16b\n"
+ ".inst 0x4e8895d1 // sdot v17.4s, v14.16b, v8.16b\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "ldr q9, [%x[params], #0xc0]\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "sqadd v30.4s, v30.4s, v18.4s\n"
- "and v17.16b, v22.16b, v21.16b\n"
- "sshr v17.4s, v17.4s, #0x1f\n"
- "srshl v30.4s, v30.4s, v21.4s\n"
- "sqadd v20.4s, v20.4s, v16.4s\n"
- "sqrdmulh v19.4s, v19.4s, v23.4s\n"
- "add v30.4s, v30.4s, v13.4s\n"
- "srshl v20.4s, v20.4s, v21.4s\n"
- "sqadd v22.4s, v22.4s, v17.4s\n"
- "and v16.16b, v19.16b, v21.16b\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "ldr q16, [%x[params], #0xa0]\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "ldr q31, [%x[params], #0xb0]\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "ldr q0, [%x[params], #0x90]\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "ldr q4, [%x[params], #0xd0]\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
+ "smin v30.4s, v30.4s, v12.4s\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
+ "smin v20.4s, v20.4s, v12.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8295d3 // sdot v19.4s, v14.16b, v2.16b\n"
+ ".inst 0x4e9b95d3 // sdot v19.4s, v14.16b, v27.16b\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "str s5, [x24, x27]\n"
+ "ldr q5, [%x[params], #0x80]\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9695d2 // sdot v18.4s, v14.16b, v22.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "str s30, [x23, x27]\n"
+ ".inst 0x4e8795d3 // sdot v19.4s, v14.16b, v7.16b\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "str s25, [x22, x27]\n"
+ "mov v30.16b, v5.16b\n"
+ "str s20, [x21, x27]\n"
+ "mov v25.16b, v5.16b\n"
+ "mov v20.16b, v5.16b\n"
+ ".inst 0x4e879405 // sdot v5.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e829419 // sdot v25.4s, v0.16b, v2.16b\n"
+ ".inst 0x4e829605 // sdot v5.4s, v16.16b, v2.16b\n"
+ "ext v7.16b, v7.16b, v7.16b, #0x1\n"
+ "add x27, x27, #0x4\n"
+ "ext v2.16b, v2.16b, v2.16b, #0x1\n"
+ "movi v17.4s, #0x0\n"
+ ".inst 0x4e87941e // sdot v30.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e829414 // sdot v20.4s, v0.16b, v2.16b\n"
+ ".inst 0x4e8295d1 // sdot v17.4s, v14.16b, v2.16b\n"
+ ".inst 0x4e9b9619 // sdot v25.4s, v16.16b, v27.16b\n"
+ ".inst 0x4e9b97e5 // sdot v5.4s, v31.16b, v27.16b\n"
+ "ext v27.16b, v27.16b, v27.16b, #0x1\n"
+ ".inst 0x4e82961e // sdot v30.4s, v16.16b, v2.16b\n"
+ ".inst 0x4e9b9614 // sdot v20.4s, v16.16b, v27.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ ".inst 0x4e9b95d1 // sdot v17.4s, v14.16b, v27.16b\n"
+ ".inst 0x4e9697f9 // sdot v25.4s, v31.16b, v22.16b\n"
+ "ext v22.16b, v22.16b, v22.16b, #0x1\n"
+ ".inst 0x4e9b97fe // sdot v30.4s, v31.16b, v27.16b\n"
+ ".inst 0x4e9697f4 // sdot v20.4s, v31.16b, v22.16b\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9695d0 // sdot v16.4s, v14.16b, v22.16b\n"
+ ".inst 0x4e8795d1 // sdot v17.4s, v14.16b, v7.16b\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "ldr q9, [%x[params], #0x120]\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "smax v30.4s, v30.4s, v9.4s\n"
- "srshl v22.4s, v22.4s, v21.4s\n"
- "add v20.4s, v20.4s, v13.4s\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "ldr q16, [%x[params], #0x100]\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "ldr q31, [%x[params], #0x110]\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "ldr q0, [%x[params], #0xf0]\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "ldr q4, [%x[params], #0x130]\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
"smin v30.4s, v30.4s, v12.4s\n"
- "add v22.4s, v22.4s, v13.4s\n"
- "smax v20.4s, v20.4s, v9.4s\n"
- "sqadd v19.4s, v19.4s, v16.4s\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
+ "smin v20.4s, v20.4s, v12.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8195d3 // sdot v19.4s, v14.16b, v1.16b\n"
+ ".inst 0x4e9a95d3 // sdot v19.4s, v14.16b, v26.16b\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "smax v22.4s, v22.4s, v9.4s\n"
+ "str s5, [x24, x27]\n"
+ "ldr q5, [%x[params], #0xe0]\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9595d2 // sdot v18.4s, v14.16b, v21.16b\n"
+ "add %x[params], %x[params], #0x140\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "str s30, [x23, x27]\n"
+ ".inst 0x4e8695d3 // sdot v19.4s, v14.16b, v6.16b\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "str s25, [x22, x27]\n"
+ "mov v30.16b, v5.16b\n"
+ "str s20, [x21, x27]\n"
+ "mov v25.16b, v5.16b\n"
+ "mov v20.16b, v5.16b\n"
+ ".inst 0x4e869405 // sdot v5.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e819419 // sdot v25.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e819605 // sdot v5.4s, v16.16b, v1.16b\n"
+ "ext v6.16b, v6.16b, v6.16b, #0x1\n"
+ "add x27, x27, #0x4\n"
+ "ext v1.16b, v1.16b, v1.16b, #0x1\n"
+ "movi v17.4s, #0x0\n"
+ ".inst 0x4e86941e // sdot v30.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e819414 // sdot v20.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e8195d1 // sdot v17.4s, v14.16b, v1.16b\n"
+ ".inst 0x4e9a9619 // sdot v25.4s, v16.16b, v26.16b\n"
+ ".inst 0x4e9a97e5 // sdot v5.4s, v31.16b, v26.16b\n"
+ "ext v26.16b, v26.16b, v26.16b, #0x1\n"
+ ".inst 0x4e81961e // sdot v30.4s, v16.16b, v1.16b\n"
+ ".inst 0x4e9a9614 // sdot v20.4s, v16.16b, v26.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ ".inst 0x4e9a95d1 // sdot v17.4s, v14.16b, v26.16b\n"
+ ".inst 0x4e9597f9 // sdot v25.4s, v31.16b, v21.16b\n"
+ "ext v21.16b, v21.16b, v21.16b, #0x1\n"
+ ".inst 0x4e9a97fe // sdot v30.4s, v31.16b, v26.16b\n"
+ ".inst 0x4e9597f4 // sdot v20.4s, v31.16b, v21.16b\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9595d0 // sdot v16.4s, v14.16b, v21.16b\n"
+ ".inst 0x4e8695d1 // sdot v17.4s, v14.16b, v6.16b\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
+ "smin v30.4s, v30.4s, v12.4s\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
"smin v20.4s, v20.4s, v12.4s\n"
- "srshl v19.4s, v19.4s, v21.4s\n"
- "smin v22.4s, v22.4s, v12.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
"uzp1 v30.16b, v30.16b, v30.16b\n"
- "str s30, [x25, x23]\n"
- "add v19.4s, v19.4s, v13.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "str s22, [x24, x23]\n"
- "smax v19.4s, v19.4s, v9.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "str s5, [x24, x27]\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
+ "str s30, [x23, x27]\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "str s20, [x22, x23]\n"
- "smin v19.4s, v19.4s, v12.4s\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "str s19, [x21, x23]\n"
- "add x23, x23, #0x4\n"
- "bgt 1b\n"
- "tst %x[n_channels], #0xf\n"
- "beq 34f\n"
- "2:" // Oddments
- "and x19, %x[n_channels], #0xf\n"
- "add x13, x13, x23\n"
- "add x12, x12, x23\n"
- "add x11, x11, x23\n"
- "add x10, x10, x23\n"
- "add x9, x9, x23\n"
- "add x28, x28, x23\n"
- "add x27, x27, x23\n"
- "add x26, x26, x23\n"
- "tbz %x[n_channels], #3, 6f\n"
- "ld1 { v27.d }[0], [x13], #0x8\n"
- "ld1 { v1.d }[0], [x12], #0x8\n"
- "ld1 { v25.d }[0], [x11], #0x8\n"
- "ld1 { v23.d }[0], [x10], #0x8\n"
- "ld1 { v31.d }[0], [x9], #0x8\n"
- "ld1 { v28.d }[0], [x28], #0x8\n"
- "ld1 { v21.d }[0], [x27], #0x8\n"
- "ld1 { v26.d }[0], [x26], #0x8\n"
- "tbz %x[n_channels], #2, 4f\n"
+ "str s25, [x22, x27]\n"
+ "str s20, [x21, x27]\n"
+ "add x27, x27, #0x4\n"
+ "beq 35f\n"
+ "3:" // Oddments
+ "and x20, %x[n_channels], #0xf\n"
+ "add x15, x15, x28\n"
+ "add x14, x14, x28\n"
+ "add x13, x13, x28\n"
+ "add x12, x12, x28\n"
+ "add x10, x10, x28\n"
+ "add x9, x9, x28\n"
+ "add x26, x26, x28\n"
+ "add x25, x25, x28\n"
+ "tbz %x[n_channels], #3, 7f\n"
+ "ldr d9, [x15], #0x8\n"
+ "ldr d8, [x14], #0x8\n"
+ "ldr d7, [x13], #0x8\n"
+ "ldr d6, [x12], #0x8\n"
+ "ldr d4, [x10], #0x8\n"
+ "ldr d3, [x9], #0x8\n"
+ "ldr d2, [x26], #0x8\n"
+ "ldr d1, [x25], #0x8\n"
+ "tbz %x[n_channels], #2, 5f\n"
+ "ld1 { v9.s }[2], [x15], #0x4\n"
+ "ld1 { v8.s }[2], [x14], #0x4\n"
+ "ld1 { v7.s }[2], [x13], #0x4\n"
+ "ld1 { v6.s }[2], [x12], #0x4\n"
+ "ld1 { v4.s }[2], [x10], #0x4\n"
+ "ld1 { v3.s }[2], [x9], #0x4\n"
+ "ld1 { v2.s }[2], [x26], #0x4\n"
+ "ld1 { v1.s }[2], [x25], #0x4\n"
+ "tbz %x[n_channels], #1, 4f\n"
+ "ld1 { v9.h }[6], [x15], #0x2\n"
+ "ld1 { v8.h }[6], [x14], #0x2\n"
+ "ld1 { v7.h }[6], [x13], #0x2\n"
+ "ld1 { v6.h }[6], [x12], #0x2\n"
+ "ld1 { v4.h }[6], [x10], #0x2\n"
+ "ld1 { v3.h }[6], [x9], #0x2\n"
+ "ld1 { v2.h }[6], [x26], #0x2\n"
+ "ld1 { v1.h }[6], [x25], #0x2\n"
+ "tbz %x[n_channels], #0, 11f\n"
+ "ld1 { v9.b }[14], [x15], #0x1\n"
+ "ld1 { v8.b }[14], [x14], #0x1\n"
+ "ld1 { v7.b }[14], [x13], #0x1\n"
+ "ld1 { v6.b }[14], [x12], #0x1\n"
+ "ld1 { v4.b }[14], [x10], #0x1\n"
+ "ld1 { v3.b }[14], [x9], #0x1\n"
+ "ld1 { v2.b }[14], [x26], #0x1\n"
+ "ld1 { v1.b }[14], [x25], #0x1\n"
+ "b 11f\n"
+ "4:" // Oddments: Load (A): Bit 3: Bit 2: Bit 1: Unset
+ "tbz %x[n_channels], #0, 11f\n"
+ "ld1 { v9.b }[12], [x15], #0x1\n"
+ "ld1 { v8.b }[12], [x14], #0x1\n"
+ "ld1 { v7.b }[12], [x13], #0x1\n"
+ "ld1 { v6.b }[12], [x12], #0x1\n"
+ "ld1 { v4.b }[12], [x10], #0x1\n"
+ "ld1 { v3.b }[12], [x9], #0x1\n"
+ "ld1 { v2.b }[12], [x26], #0x1\n"
+ "ld1 { v1.b }[12], [x25], #0x1\n"
+ "b 11f\n"
+ "5:" // Oddments: Load (A): Bit 3: Bit 2: Unset
+ "tbz %x[n_channels], #1, 6f\n"
+ "ld1 { v9.h }[4], [x15], #0x2\n"
+ "ld1 { v8.h }[4], [x14], #0x2\n"
+ "ld1 { v7.h }[4], [x13], #0x2\n"
+ "ld1 { v6.h }[4], [x12], #0x2\n"
+ "ld1 { v4.h }[4], [x10], #0x2\n"
+ "ld1 { v3.h }[4], [x9], #0x2\n"
+ "ld1 { v2.h }[4], [x26], #0x2\n"
+ "ld1 { v1.h }[4], [x25], #0x2\n"
+ "tbz %x[n_channels], #0, 11f\n"
+ "ld1 { v9.b }[10], [x15], #0x1\n"
+ "ld1 { v8.b }[10], [x14], #0x1\n"
+ "ld1 { v7.b }[10], [x13], #0x1\n"
+ "ld1 { v6.b }[10], [x12], #0x1\n"
+ "ld1 { v4.b }[10], [x10], #0x1\n"
+ "ld1 { v3.b }[10], [x9], #0x1\n"
+ "ld1 { v2.b }[10], [x26], #0x1\n"
+ "ld1 { v1.b }[10], [x25], #0x1\n"
+ "b 11f\n"
+ "6:" // Oddments: Load (A): Bit 3: Bit 2: Unset: Bit 1: Unset
+ "tbz %x[n_channels], #0, 11f\n"
+ "ld1 { v9.b }[8], [x15], #0x1\n"
+ "ld1 { v8.b }[8], [x14], #0x1\n"
+ "ld1 { v7.b }[8], [x13], #0x1\n"
+ "ld1 { v6.b }[8], [x12], #0x1\n"
+ "ld1 { v4.b }[8], [x10], #0x1\n"
+ "ld1 { v3.b }[8], [x9], #0x1\n"
+ "ld1 { v2.b }[8], [x26], #0x1\n"
+ "ld1 { v1.b }[8], [x25], #0x1\n"
+ "b 11f\n"
+ "7:" // Oddments: Load (A): Bit 3: Unset
+ "tbz %x[n_channels], #2, 9f\n"
+ "ldr s9, [x15], #0x4\n"
+ "ldr s8, [x14], #0x4\n"
+ "ldr s7, [x13], #0x4\n"
+ "ldr s6, [x12], #0x4\n"
+ "ldr s4, [x10], #0x4\n"
+ "ldr s3, [x9], #0x4\n"
+ "ldr s2, [x26], #0x4\n"
+ "ldr s1, [x25], #0x4\n"
+ "tbz %x[n_channels], #1, 8f\n"
+ "ld1 { v9.h }[2], [x15], #0x2\n"
+ "ld1 { v8.h }[2], [x14], #0x2\n"
+ "ld1 { v7.h }[2], [x13], #0x2\n"
+ "ld1 { v6.h }[2], [x12], #0x2\n"
+ "ld1 { v4.h }[2], [x10], #0x2\n"
+ "ld1 { v3.h }[2], [x9], #0x2\n"
+ "ld1 { v2.h }[2], [x26], #0x2\n"
+ "ld1 { v1.h }[2], [x25], #0x2\n"
+ "tbz %x[n_channels], #0, 11f\n"
+ "ld1 { v9.b }[6], [x15], #0x1\n"
+ "ld1 { v8.b }[6], [x14], #0x1\n"
+ "ld1 { v7.b }[6], [x13], #0x1\n"
+ "ld1 { v6.b }[6], [x12], #0x1\n"
+ "ld1 { v4.b }[6], [x10], #0x1\n"
+ "ld1 { v3.b }[6], [x9], #0x1\n"
+ "ld1 { v2.b }[6], [x26], #0x1\n"
+ "ld1 { v1.b }[6], [x25], #0x1\n"
+ "b 11f\n"
+ "8:" // Oddments: Load (A): Bit 3: Unset: Bit 2: Bit 1: Unset
+ "tbz %x[n_channels], #0, 11f\n"
+ "ld1 { v9.b }[4], [x15], #0x1\n"
+ "ld1 { v8.b }[4], [x14], #0x1\n"
+ "ld1 { v7.b }[4], [x13], #0x1\n"
+ "ld1 { v6.b }[4], [x12], #0x1\n"
+ "ld1 { v4.b }[4], [x10], #0x1\n"
+ "ld1 { v3.b }[4], [x9], #0x1\n"
+ "ld1 { v2.b }[4], [x26], #0x1\n"
+ "ld1 { v1.b }[4], [x25], #0x1\n"
+ "b 11f\n"
+ "9:" // Oddments: Load (A): Bit 3: Unset: Bit 2: Unset
+ "tbz %x[n_channels], #1, 10f\n"
+ "ldr h9, [x15], #0x2\n"
+ "ldr h8, [x14], #0x2\n"
+ "ldr h7, [x13], #0x2\n"
+ "ldr h6, [x12], #0x2\n"
+ "ldr h4, [x10], #0x2\n"
+ "ldr h3, [x9], #0x2\n"
+ "ldr h2, [x26], #0x2\n"
+ "ldr h1, [x25], #0x2\n"
+ "tbz %x[n_channels], #0, 11f\n"
+ "ld1 { v9.b }[2], [x15], #0x1\n"
+ "ld1 { v8.b }[2], [x14], #0x1\n"
+ "ld1 { v7.b }[2], [x13], #0x1\n"
+ "ld1 { v6.b }[2], [x12], #0x1\n"
+ "ld1 { v4.b }[2], [x10], #0x1\n"
+ "ld1 { v3.b }[2], [x9], #0x1\n"
+ "ld1 { v2.b }[2], [x26], #0x1\n"
+ "ld1 { v1.b }[2], [x25], #0x1\n"
+ "b 11f\n"
+ "10:" // Oddments: Load (A): Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
+ "ldr b9, [x15], #0x1\n"
+ "ldr b8, [x14], #0x1\n"
+ "ldr b7, [x13], #0x1\n"
+ "ldr b6, [x12], #0x1\n"
+ "ldr b4, [x10], #0x1\n"
+ "ldr b3, [x9], #0x1\n"
+ "ldr b2, [x26], #0x1\n"
+ "ldr b1, [x25], #0x1\n"
+ "11:" // Oddments: Load (A): Bit 3: End
+ "ldp x15, x14, [%x[inptrs], #0x40]\n"
+ "ldp x13, x12, [%x[inptrs], #0x50]\n"
+ "add x15, x15, x28\n"
+ "add x14, x14, x28\n"
+ "ldp x10, x9, [%x[inptrs], #0x60]\n"
+ "ldp x26, x25, [%x[inptrs], #0x70]\n"
+ "add x13, x13, x28\n"
+ "add x12, x12, x28\n"
+ "add x10, x10, x28\n"
+ "add x9, x9, x28\n"
+ "add x26, x26, x28\n"
+ "add x25, x25, x28\n"
+ "tbz %x[n_channels], #3, 15f\n"
+ "ldr d29, [x15], #0x8\n"
+ "ldr d28, [x14], #0x8\n"
+ "ldr d27, [x13], #0x8\n"
+ "ldr d26, [x12], #0x8\n"
+ "ldr d24, [x10], #0x8\n"
+ "ldr d23, [x9], #0x8\n"
+ "ldr d22, [x26], #0x8\n"
+ "ldr d21, [x25], #0x8\n"
+ "tbz %x[n_channels], #2, 13f\n"
+ "ld1 { v29.s }[2], [x15], #0x4\n"
+ "ld1 { v28.s }[2], [x14], #0x4\n"
"ld1 { v27.s }[2], [x13], #0x4\n"
- "ld1 { v1.s }[2], [x12], #0x4\n"
- "ld1 { v25.s }[2], [x11], #0x4\n"
- "ld1 { v23.s }[2], [x10], #0x4\n"
- "ld1 { v31.s }[2], [x9], #0x4\n"
- "ld1 { v28.s }[2], [x28], #0x4\n"
- "ld1 { v21.s }[2], [x27], #0x4\n"
- "ld1 { v26.s }[2], [x26], #0x4\n"
- "tbz %x[n_channels], #1, 3f\n"
+ "ld1 { v26.s }[2], [x12], #0x4\n"
+ "ld1 { v24.s }[2], [x10], #0x4\n"
+ "ld1 { v23.s }[2], [x9], #0x4\n"
+ "ld1 { v22.s }[2], [x26], #0x4\n"
+ "ld1 { v21.s }[2], [x25], #0x4\n"
+ "tbz %x[n_channels], #1, 12f\n"
+ "ld1 { v29.h }[6], [x15], #0x2\n"
+ "ld1 { v28.h }[6], [x14], #0x2\n"
"ld1 { v27.h }[6], [x13], #0x2\n"
- "ld1 { v1.h }[6], [x12], #0x2\n"
- "ld1 { v25.h }[6], [x11], #0x2\n"
- "ld1 { v23.h }[6], [x10], #0x2\n"
- "ld1 { v31.h }[6], [x9], #0x2\n"
- "ld1 { v28.h }[6], [x28], #0x2\n"
- "ld1 { v21.h }[6], [x27], #0x2\n"
- "ld1 { v26.h }[6], [x26], #0x2\n"
- "tbz %x[n_channels], #0, 10f\n"
+ "ld1 { v26.h }[6], [x12], #0x2\n"
+ "ld1 { v24.h }[6], [x10], #0x2\n"
+ "ld1 { v23.h }[6], [x9], #0x2\n"
+ "ld1 { v22.h }[6], [x26], #0x2\n"
+ "ld1 { v21.h }[6], [x25], #0x2\n"
+ "tbz %x[n_channels], #0, 19f\n"
+ "ld1 { v29.b }[14], [x15], #0x1\n"
+ "ld1 { v28.b }[14], [x14], #0x1\n"
"ld1 { v27.b }[14], [x13], #0x1\n"
- "ld1 { v1.b }[14], [x12], #0x1\n"
- "ld1 { v25.b }[14], [x11], #0x1\n"
- "ld1 { v23.b }[14], [x10], #0x1\n"
- "ld1 { v31.b }[14], [x9], #0x1\n"
- "ld1 { v28.b }[14], [x28], #0x1\n"
- "ld1 { v21.b }[14], [x27], #0x1\n"
- "ld1 { v26.b }[14], [x26], #0x1\n"
- "b 10f\n"
- "3:" // Oddments: Load (A): Bit 3: Bit 2: Bit 1: Unset
- "tbz %x[n_channels], #0, 10f\n"
+ "ld1 { v26.b }[14], [x12], #0x1\n"
+ "ld1 { v24.b }[14], [x10], #0x1\n"
+ "ld1 { v23.b }[14], [x9], #0x1\n"
+ "ld1 { v22.b }[14], [x26], #0x1\n"
+ "ld1 { v21.b }[14], [x25], #0x1\n"
+ "b 19f\n"
+ "12:" // Oddments: Load (B): Bit 3: Bit 2: Bit 1: Unset
+ "tbz %x[n_channels], #0, 19f\n"
+ "ld1 { v29.b }[12], [x15], #0x1\n"
+ "ld1 { v28.b }[12], [x14], #0x1\n"
"ld1 { v27.b }[12], [x13], #0x1\n"
- "ld1 { v1.b }[12], [x12], #0x1\n"
- "ld1 { v25.b }[12], [x11], #0x1\n"
- "ld1 { v23.b }[12], [x10], #0x1\n"
- "ld1 { v31.b }[12], [x9], #0x1\n"
- "ld1 { v28.b }[12], [x28], #0x1\n"
- "ld1 { v21.b }[12], [x27], #0x1\n"
- "ld1 { v26.b }[12], [x26], #0x1\n"
- "b 10f\n"
- "4:" // Oddments: Load (A): Bit 3: Bit 2: Unset
- "tbz %x[n_channels], #1, 5f\n"
+ "ld1 { v26.b }[12], [x12], #0x1\n"
+ "ld1 { v24.b }[12], [x10], #0x1\n"
+ "ld1 { v23.b }[12], [x9], #0x1\n"
+ "ld1 { v22.b }[12], [x26], #0x1\n"
+ "ld1 { v21.b }[12], [x25], #0x1\n"
+ "b 19f\n"
+ "13:" // Oddments: Load (B): Bit 3: Bit 2: Unset
+ "tbz %x[n_channels], #1, 14f\n"
+ "ld1 { v29.h }[4], [x15], #0x2\n"
+ "ld1 { v28.h }[4], [x14], #0x2\n"
"ld1 { v27.h }[4], [x13], #0x2\n"
- "ld1 { v1.h }[4], [x12], #0x2\n"
- "ld1 { v25.h }[4], [x11], #0x2\n"
- "ld1 { v23.h }[4], [x10], #0x2\n"
- "ld1 { v31.h }[4], [x9], #0x2\n"
- "ld1 { v28.h }[4], [x28], #0x2\n"
- "ld1 { v21.h }[4], [x27], #0x2\n"
- "ld1 { v26.h }[4], [x26], #0x2\n"
- "tbz %x[n_channels], #0, 10f\n"
+ "ld1 { v26.h }[4], [x12], #0x2\n"
+ "ld1 { v24.h }[4], [x10], #0x2\n"
+ "ld1 { v23.h }[4], [x9], #0x2\n"
+ "ld1 { v22.h }[4], [x26], #0x2\n"
+ "ld1 { v21.h }[4], [x25], #0x2\n"
+ "tbz %x[n_channels], #0, 19f\n"
+ "ld1 { v29.b }[10], [x15], #0x1\n"
+ "ld1 { v28.b }[10], [x14], #0x1\n"
"ld1 { v27.b }[10], [x13], #0x1\n"
- "ld1 { v1.b }[10], [x12], #0x1\n"
- "ld1 { v25.b }[10], [x11], #0x1\n"
- "ld1 { v23.b }[10], [x10], #0x1\n"
- "ld1 { v31.b }[10], [x9], #0x1\n"
- "ld1 { v28.b }[10], [x28], #0x1\n"
- "ld1 { v21.b }[10], [x27], #0x1\n"
- "ld1 { v26.b }[10], [x26], #0x1\n"
- "b 10f\n"
- "5:" // Oddments: Load (A): Bit 3: Bit 2: Unset: Bit 1: Unset
- "tbz %x[n_channels], #0, 10f\n"
+ "ld1 { v26.b }[10], [x12], #0x1\n"
+ "ld1 { v24.b }[10], [x10], #0x1\n"
+ "ld1 { v23.b }[10], [x9], #0x1\n"
+ "ld1 { v22.b }[10], [x26], #0x1\n"
+ "ld1 { v21.b }[10], [x25], #0x1\n"
+ "b 19f\n"
+ "14:" // Oddments: Load (B): Bit 3: Bit 2: Unset: Bit 1: Unset
+ "tbz %x[n_channels], #0, 19f\n"
+ "ld1 { v29.b }[8], [x15], #0x1\n"
+ "ld1 { v28.b }[8], [x14], #0x1\n"
"ld1 { v27.b }[8], [x13], #0x1\n"
- "ld1 { v1.b }[8], [x12], #0x1\n"
- "ld1 { v25.b }[8], [x11], #0x1\n"
- "ld1 { v23.b }[8], [x10], #0x1\n"
- "ld1 { v31.b }[8], [x9], #0x1\n"
- "ld1 { v28.b }[8], [x28], #0x1\n"
- "ld1 { v21.b }[8], [x27], #0x1\n"
- "ld1 { v26.b }[8], [x26], #0x1\n"
- "b 10f\n"
- "6:" // Oddments: Load (A): Bit 3: Unset
- "tbz %x[n_channels], #2, 8f\n"
- "ld1 { v27.s }[0], [x13], #0x4\n"
- "ld1 { v1.s }[0], [x12], #0x4\n"
- "ld1 { v25.s }[0], [x11], #0x4\n"
- "ld1 { v23.s }[0], [x10], #0x4\n"
- "ld1 { v31.s }[0], [x9], #0x4\n"
- "ld1 { v28.s }[0], [x28], #0x4\n"
- "ld1 { v21.s }[0], [x27], #0x4\n"
- "ld1 { v26.s }[0], [x26], #0x4\n"
- "tbz %x[n_channels], #1, 7f\n"
+ "ld1 { v26.b }[8], [x12], #0x1\n"
+ "ld1 { v24.b }[8], [x10], #0x1\n"
+ "ld1 { v23.b }[8], [x9], #0x1\n"
+ "ld1 { v22.b }[8], [x26], #0x1\n"
+ "ld1 { v21.b }[8], [x25], #0x1\n"
+ "b 19f\n"
+ "15:" // Oddments: Load (B): Bit 3: Unset
+ "tbz %x[n_channels], #2, 17f\n"
+ "ldr s29, [x15], #0x4\n"
+ "ldr s28, [x14], #0x4\n"
+ "ldr s27, [x13], #0x4\n"
+ "ldr s26, [x12], #0x4\n"
+ "ldr s24, [x10], #0x4\n"
+ "ldr s23, [x9], #0x4\n"
+ "ldr s22, [x26], #0x4\n"
+ "ldr s21, [x25], #0x4\n"
+ "tbz %x[n_channels], #1, 16f\n"
+ "ld1 { v29.h }[2], [x15], #0x2\n"
+ "ld1 { v28.h }[2], [x14], #0x2\n"
"ld1 { v27.h }[2], [x13], #0x2\n"
- "ld1 { v1.h }[2], [x12], #0x2\n"
- "ld1 { v25.h }[2], [x11], #0x2\n"
- "ld1 { v23.h }[2], [x10], #0x2\n"
- "ld1 { v31.h }[2], [x9], #0x2\n"
- "ld1 { v28.h }[2], [x28], #0x2\n"
- "ld1 { v21.h }[2], [x27], #0x2\n"
- "ld1 { v26.h }[2], [x26], #0x2\n"
- "tbz %x[n_channels], #0, 10f\n"
+ "ld1 { v26.h }[2], [x12], #0x2\n"
+ "ld1 { v24.h }[2], [x10], #0x2\n"
+ "ld1 { v23.h }[2], [x9], #0x2\n"
+ "ld1 { v22.h }[2], [x26], #0x2\n"
+ "ld1 { v21.h }[2], [x25], #0x2\n"
+ "tbz %x[n_channels], #0, 19f\n"
+ "ld1 { v29.b }[6], [x15], #0x1\n"
+ "ld1 { v28.b }[6], [x14], #0x1\n"
"ld1 { v27.b }[6], [x13], #0x1\n"
- "ld1 { v1.b }[6], [x12], #0x1\n"
- "ld1 { v25.b }[6], [x11], #0x1\n"
- "ld1 { v23.b }[6], [x10], #0x1\n"
- "ld1 { v31.b }[6], [x9], #0x1\n"
- "ld1 { v28.b }[6], [x28], #0x1\n"
- "ld1 { v21.b }[6], [x27], #0x1\n"
- "ld1 { v26.b }[6], [x26], #0x1\n"
- "b 10f\n"
- "7:" // Oddments: Load (A): Bit 3: Unset: Bit 2: Bit 1: Unset
- "tbz %x[n_channels], #0, 10f\n"
+ "ld1 { v26.b }[6], [x12], #0x1\n"
+ "ld1 { v24.b }[6], [x10], #0x1\n"
+ "ld1 { v23.b }[6], [x9], #0x1\n"
+ "ld1 { v22.b }[6], [x26], #0x1\n"
+ "ld1 { v21.b }[6], [x25], #0x1\n"
+ "b 19f\n"
+ "16:" // Oddments: Load (B): Bit 3: Unset: Bit 2: Bit 1: Unset
+ "tbz %x[n_channels], #0, 19f\n"
+ "ld1 { v29.b }[4], [x15], #0x1\n"
+ "ld1 { v28.b }[4], [x14], #0x1\n"
"ld1 { v27.b }[4], [x13], #0x1\n"
- "ld1 { v1.b }[4], [x12], #0x1\n"
- "ld1 { v25.b }[4], [x11], #0x1\n"
- "ld1 { v23.b }[4], [x10], #0x1\n"
- "ld1 { v31.b }[4], [x9], #0x1\n"
- "ld1 { v28.b }[4], [x28], #0x1\n"
- "ld1 { v21.b }[4], [x27], #0x1\n"
- "ld1 { v26.b }[4], [x26], #0x1\n"
- "b 10f\n"
- "8:" // Oddments: Load (A): Bit 3: Unset: Bit 2: Unset
- "tbz %x[n_channels], #1, 9f\n"
- "ld1 { v27.h }[0], [x13], #0x2\n"
- "ld1 { v1.h }[0], [x12], #0x2\n"
- "ld1 { v25.h }[0], [x11], #0x2\n"
- "ld1 { v23.h }[0], [x10], #0x2\n"
- "ld1 { v31.h }[0], [x9], #0x2\n"
- "ld1 { v28.h }[0], [x28], #0x2\n"
- "ld1 { v21.h }[0], [x27], #0x2\n"
- "ld1 { v26.h }[0], [x26], #0x2\n"
- "tbz %x[n_channels], #0, 10f\n"
+ "ld1 { v26.b }[4], [x12], #0x1\n"
+ "ld1 { v24.b }[4], [x10], #0x1\n"
+ "ld1 { v23.b }[4], [x9], #0x1\n"
+ "ld1 { v22.b }[4], [x26], #0x1\n"
+ "ld1 { v21.b }[4], [x25], #0x1\n"
+ "b 19f\n"
+ "17:" // Oddments: Load (B): Bit 3: Unset: Bit 2: Unset
+ "tbz %x[n_channels], #1, 18f\n"
+ "ldr h29, [x15], #0x2\n"
+ "ldr h28, [x14], #0x2\n"
+ "ldr h27, [x13], #0x2\n"
+ "ldr h26, [x12], #0x2\n"
+ "ldr h24, [x10], #0x2\n"
+ "ldr h23, [x9], #0x2\n"
+ "ldr h22, [x26], #0x2\n"
+ "ldr h21, [x25], #0x2\n"
+ "tbz %x[n_channels], #0, 19f\n"
+ "ld1 { v29.b }[2], [x15], #0x1\n"
+ "ld1 { v28.b }[2], [x14], #0x1\n"
"ld1 { v27.b }[2], [x13], #0x1\n"
- "ld1 { v1.b }[2], [x12], #0x1\n"
- "ld1 { v25.b }[2], [x11], #0x1\n"
- "ld1 { v23.b }[2], [x10], #0x1\n"
- "ld1 { v31.b }[2], [x9], #0x1\n"
- "ld1 { v28.b }[2], [x28], #0x1\n"
- "ld1 { v21.b }[2], [x27], #0x1\n"
- "ld1 { v26.b }[2], [x26], #0x1\n"
- "b 10f\n"
- "9:" // Oddments: Load (A): Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
- "tbz %x[n_channels], #0, 10f\n"
- "ld1 { v27.b }[0], [x13], #0x1\n"
- "ld1 { v1.b }[0], [x12], #0x1\n"
- "ld1 { v25.b }[0], [x11], #0x1\n"
- "ld1 { v23.b }[0], [x10], #0x1\n"
- "ld1 { v31.b }[0], [x9], #0x1\n"
- "ld1 { v28.b }[0], [x28], #0x1\n"
- "ld1 { v21.b }[0], [x27], #0x1\n"
- "ld1 { v26.b }[0], [x26], #0x1\n"
- "10:" // Oddments: Load (A): Bit 3: End
- "ldp x13, x12, [%x[inptrs], #0x40]\n"
- "add x13, x13, x23\n"
- "ldp x11, x10, [%x[inptrs], #0x50]\n"
- "ldp x9, x28, [%x[inptrs], #0x60]\n"
- "add x12, x12, x23\n"
- "ldp x27, x26, [%x[inptrs], #0x70]\n"
- "add x11, x11, x23\n"
- "add x10, x10, x23\n"
- "add x9, x9, x23\n"
- "add x28, x28, x23\n"
- "add x27, x27, x23\n"
- "add x26, x26, x23\n"
- "tbz %x[n_channels], #3, 14f\n"
- "ld1 { v24.d }[0], [x13], #0x8\n"
- "ld1 { v22.d }[0], [x12], #0x8\n"
- "ld1 { v20.d }[0], [x11], #0x8\n"
- "ld1 { v16.d }[0], [x10], #0x8\n"
- "ld1 { v19.d }[0], [x9], #0x8\n"
- "ld1 { v0.d }[0], [x28], #0x8\n"
- "ld1 { v18.d }[0], [x27], #0x8\n"
- "ld1 { v17.d }[0], [x26], #0x8\n"
- "tbz %x[n_channels], #2, 12f\n"
- "ld1 { v24.s }[2], [x13], #0x4\n"
- "ld1 { v22.s }[2], [x12], #0x4\n"
- "ld1 { v20.s }[2], [x11], #0x4\n"
- "ld1 { v16.s }[2], [x10], #0x4\n"
- "ld1 { v19.s }[2], [x9], #0x4\n"
- "ld1 { v0.s }[2], [x28], #0x4\n"
- "ld1 { v18.s }[2], [x27], #0x4\n"
- "ld1 { v17.s }[2], [x26], #0x4\n"
- "tbz %x[n_channels], #1, 11f\n"
- "ld1 { v24.h }[6], [x13], #0x2\n"
- "ld1 { v22.h }[6], [x12], #0x2\n"
- "ld1 { v20.h }[6], [x11], #0x2\n"
- "ld1 { v16.h }[6], [x10], #0x2\n"
- "ld1 { v19.h }[6], [x9], #0x2\n"
- "ld1 { v0.h }[6], [x28], #0x2\n"
- "ld1 { v18.h }[6], [x27], #0x2\n"
- "ld1 { v17.h }[6], [x26], #0x2\n"
- "tbz %x[n_channels], #0, 18f\n"
- "ld1 { v24.b }[14], [x13], #0x1\n"
- "ld1 { v22.b }[14], [x12], #0x1\n"
- "ld1 { v20.b }[14], [x11], #0x1\n"
- "ld1 { v16.b }[14], [x10], #0x1\n"
- "ld1 { v19.b }[14], [x9], #0x1\n"
- "ld1 { v0.b }[14], [x28], #0x1\n"
- "ld1 { v18.b }[14], [x27], #0x1\n"
- "ld1 { v17.b }[14], [x26], #0x1\n"
- "b 18f\n"
- "11:" // Oddments: Load (B): Bit 3: Bit 2: Bit 1: Unset
- "tbz %x[n_channels], #0, 18f\n"
- "ld1 { v24.b }[12], [x13], #0x1\n"
- "ld1 { v22.b }[12], [x12], #0x1\n"
- "ld1 { v20.b }[12], [x11], #0x1\n"
- "ld1 { v16.b }[12], [x10], #0x1\n"
- "ld1 { v19.b }[12], [x9], #0x1\n"
- "ld1 { v0.b }[12], [x28], #0x1\n"
- "ld1 { v18.b }[12], [x27], #0x1\n"
- "ld1 { v17.b }[12], [x26], #0x1\n"
- "b 18f\n"
- "12:" // Oddments: Load (B): Bit 3: Bit 2: Unset
- "tbz %x[n_channels], #1, 13f\n"
- "ld1 { v24.h }[4], [x13], #0x2\n"
- "ld1 { v22.h }[4], [x12], #0x2\n"
- "ld1 { v20.h }[4], [x11], #0x2\n"
- "ld1 { v16.h }[4], [x10], #0x2\n"
- "ld1 { v19.h }[4], [x9], #0x2\n"
- "ld1 { v0.h }[4], [x28], #0x2\n"
- "ld1 { v18.h }[4], [x27], #0x2\n"
- "ld1 { v17.h }[4], [x26], #0x2\n"
- "tbz %x[n_channels], #0, 18f\n"
- "ld1 { v24.b }[10], [x13], #0x1\n"
- "ld1 { v22.b }[10], [x12], #0x1\n"
- "ld1 { v20.b }[10], [x11], #0x1\n"
- "ld1 { v16.b }[10], [x10], #0x1\n"
- "ld1 { v19.b }[10], [x9], #0x1\n"
- "ld1 { v0.b }[10], [x28], #0x1\n"
- "ld1 { v18.b }[10], [x27], #0x1\n"
- "ld1 { v17.b }[10], [x26], #0x1\n"
- "b 18f\n"
- "13:" // Oddments: Load (B): Bit 3: Bit 2: Unset: Bit 1: Unset
- "tbz %x[n_channels], #0, 18f\n"
- "ld1 { v24.b }[8], [x13], #0x1\n"
- "ld1 { v22.b }[8], [x12], #0x1\n"
- "ld1 { v20.b }[8], [x11], #0x1\n"
- "ld1 { v16.b }[8], [x10], #0x1\n"
- "ld1 { v19.b }[8], [x9], #0x1\n"
- "ld1 { v0.b }[8], [x28], #0x1\n"
- "ld1 { v18.b }[8], [x27], #0x1\n"
- "ld1 { v17.b }[8], [x26], #0x1\n"
- "b 18f\n"
- "14:" // Oddments: Load (B): Bit 3: Unset
- "tbz %x[n_channels], #2, 16f\n"
- "ld1 { v24.s }[0], [x13], #0x4\n"
- "ld1 { v22.s }[0], [x12], #0x4\n"
- "ld1 { v20.s }[0], [x11], #0x4\n"
- "ld1 { v16.s }[0], [x10], #0x4\n"
- "ld1 { v19.s }[0], [x9], #0x4\n"
- "ld1 { v0.s }[0], [x28], #0x4\n"
- "ld1 { v18.s }[0], [x27], #0x4\n"
- "ld1 { v17.s }[0], [x26], #0x4\n"
- "tbz %x[n_channels], #1, 15f\n"
- "ld1 { v24.h }[2], [x13], #0x2\n"
- "ld1 { v22.h }[2], [x12], #0x2\n"
- "ld1 { v20.h }[2], [x11], #0x2\n"
- "ld1 { v16.h }[2], [x10], #0x2\n"
- "ld1 { v19.h }[2], [x9], #0x2\n"
- "ld1 { v0.h }[2], [x28], #0x2\n"
- "ld1 { v18.h }[2], [x27], #0x2\n"
- "ld1 { v17.h }[2], [x26], #0x2\n"
- "tbz %x[n_channels], #0, 18f\n"
- "ld1 { v24.b }[6], [x13], #0x1\n"
- "ld1 { v22.b }[6], [x12], #0x1\n"
- "ld1 { v20.b }[6], [x11], #0x1\n"
- "ld1 { v16.b }[6], [x10], #0x1\n"
- "ld1 { v19.b }[6], [x9], #0x1\n"
- "ld1 { v0.b }[6], [x28], #0x1\n"
- "ld1 { v18.b }[6], [x27], #0x1\n"
- "ld1 { v17.b }[6], [x26], #0x1\n"
- "b 18f\n"
- "15:" // Oddments: Load (B): Bit 3: Unset: Bit 2: Bit 1: Unset
- "tbz %x[n_channels], #0, 18f\n"
- "ld1 { v24.b }[4], [x13], #0x1\n"
- "ld1 { v22.b }[4], [x12], #0x1\n"
- "ld1 { v20.b }[4], [x11], #0x1\n"
- "ld1 { v16.b }[4], [x10], #0x1\n"
- "ld1 { v19.b }[4], [x9], #0x1\n"
- "ld1 { v0.b }[4], [x28], #0x1\n"
- "ld1 { v18.b }[4], [x27], #0x1\n"
- "ld1 { v17.b }[4], [x26], #0x1\n"
- "b 18f\n"
- "16:" // Oddments: Load (B): Bit 3: Unset: Bit 2: Unset
- "tbz %x[n_channels], #1, 17f\n"
- "ld1 { v24.h }[0], [x13], #0x2\n"
- "ld1 { v22.h }[0], [x12], #0x2\n"
- "ld1 { v20.h }[0], [x11], #0x2\n"
- "ld1 { v16.h }[0], [x10], #0x2\n"
- "ld1 { v19.h }[0], [x9], #0x2\n"
- "ld1 { v0.h }[0], [x28], #0x2\n"
- "ld1 { v18.h }[0], [x27], #0x2\n"
- "ld1 { v17.h }[0], [x26], #0x2\n"
- "tbz %x[n_channels], #0, 18f\n"
- "ld1 { v24.b }[2], [x13], #0x1\n"
- "ld1 { v22.b }[2], [x12], #0x1\n"
- "ld1 { v20.b }[2], [x11], #0x1\n"
- "ld1 { v16.b }[2], [x10], #0x1\n"
- "ld1 { v19.b }[2], [x9], #0x1\n"
- "ld1 { v0.b }[2], [x28], #0x1\n"
- "ld1 { v18.b }[2], [x27], #0x1\n"
- "ld1 { v17.b }[2], [x26], #0x1\n"
- "b 18f\n"
- "17:" // Oddments: Load (B): Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
- "tbz %x[n_channels], #0, 18f\n"
- "ld1 { v24.b }[0], [x13], #0x1\n"
- "ld1 { v22.b }[0], [x12], #0x1\n"
- "ld1 { v20.b }[0], [x11], #0x1\n"
- "ld1 { v16.b }[0], [x10], #0x1\n"
- "ld1 { v19.b }[0], [x9], #0x1\n"
- "ld1 { v0.b }[0], [x28], #0x1\n"
- "ld1 { v18.b }[0], [x27], #0x1\n"
- "ld1 { v17.b }[0], [x26], #0x1\n"
- "18:" // Oddments: Load (B): Bit 3: End
- "zip1 v7.16b, v27.16b, v25.16b\n"
- "ldr q30, [%x[params], #0x0]\n"
- "cmp x19, #0x4\n"
- "zip2 v5.16b, v27.16b, v25.16b\n"
- "ldr q29, [%x[params], #0x10]\n"
- "zip1 v8.16b, v1.16b, v23.16b\n"
- "ldr q27, [%x[params], #0x20]\n"
- "zip2 v3.16b, v1.16b, v23.16b\n"
- "ldr q25, [%x[params], #0x30]\n"
- "zip1 v2.16b, v31.16b, v21.16b\n"
- "ldr q23, [%x[params], #0x40]\n"
- "zip2 v4.16b, v31.16b, v21.16b\n"
- "ldr q21, [%x[params], #0x50]\n"
+ "ld1 { v26.b }[2], [x12], #0x1\n"
+ "ld1 { v24.b }[2], [x10], #0x1\n"
+ "ld1 { v23.b }[2], [x9], #0x1\n"
+ "ld1 { v22.b }[2], [x26], #0x1\n"
+ "ld1 { v21.b }[2], [x25], #0x1\n"
+ "b 19f\n"
+ "18:" // Oddments: Load (B): Bit 3: Unset: Bit 2: Unset: Bit 1: Unset
+ "ldr b29, [x15], #0x1\n"
+ "ldr b28, [x14], #0x1\n"
+ "ldr b27, [x13], #0x1\n"
+ "ldr b26, [x12], #0x1\n"
+ "ldr b24, [x10], #0x1\n"
+ "ldr b23, [x9], #0x1\n"
+ "ldr b22, [x26], #0x1\n"
+ "ldr b21, [x25], #0x1\n"
+ "19:" // Oddments: Load (B): Bit 3: End
+ "ldr q0, [%x[params], #0x10]\n"
+ "ldr q16, [%x[params], #0x20]\n"
+ "zip2 v30.16b, v4.16b, v2.16b\n"
+ "zip1 v4.16b, v4.16b, v2.16b\n"
+ "ldr q31, [%x[params], #0x30]\n"
+ "zip1 v2.16b, v3.16b, v1.16b\n"
+ "zip2 v5.16b, v9.16b, v7.16b\n"
+ "cmp x20, #0x4\n"
+ "zip1 v9.16b, v9.16b, v7.16b\n"
+ "zip1 v7.16b, v8.16b, v6.16b\n"
+ "zip2 v6.16b, v8.16b, v6.16b\n"
+ "zip2 v1.16b, v3.16b, v1.16b\n"
+ "zip2 v3.16b, v4.16b, v2.16b\n"
+ "zip1 v4.16b, v4.16b, v2.16b\n"
+ "zip2 v25.16b, v29.16b, v27.16b\n"
+ "zip1 v29.16b, v29.16b, v27.16b\n"
+ "zip1 v27.16b, v28.16b, v26.16b\n"
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8495d3 // sdot v19.4s, v14.16b, v4.16b\n"
+ "zip2 v8.16b, v9.16b, v7.16b\n"
+ "zip1 v9.16b, v9.16b, v7.16b\n"
+ "zip1 v7.16b, v5.16b, v6.16b\n"
+ "zip2 v6.16b, v5.16b, v6.16b\n"
+ "ldr q5, [%x[params], #0x0]\n"
+ "zip2 v26.16b, v28.16b, v26.16b\n"
+ "zip2 v20.16b, v24.16b, v22.16b\n"
+ "zip1 v24.16b, v24.16b, v22.16b\n"
+ "zip1 v22.16b, v23.16b, v21.16b\n"
+ "zip2 v21.16b, v23.16b, v21.16b\n"
+ "zip2 v28.16b, v29.16b, v27.16b\n"
+ "zip1 v29.16b, v29.16b, v27.16b\n"
+ "zip1 v2.16b, v30.16b, v1.16b\n"
+ ".inst 0x4e9d95d3 // sdot v19.4s, v14.16b, v29.16b\n"
+ "zip2 v1.16b, v30.16b, v1.16b\n"
+ "zip1 v27.16b, v25.16b, v26.16b\n"
+ "zip2 v26.16b, v25.16b, v26.16b\n"
+ "zip2 v23.16b, v24.16b, v22.16b\n"
+ "zip1 v24.16b, v24.16b, v22.16b\n"
+ "zip1 v22.16b, v20.16b, v21.16b\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9895d2 // sdot v18.4s, v14.16b, v24.16b\n"
+ "zip2 v21.16b, v20.16b, v21.16b\n"
+ "mov v30.16b, v5.16b\n"
+ ".inst 0x4e8995d3 // sdot v19.4s, v14.16b, v9.16b\n"
+ "mov v25.16b, v5.16b\n"
+ "mov v20.16b, v5.16b\n"
+ ".inst 0x4e899405 // sdot v5.4s, v0.16b, v9.16b\n"
+ ".inst 0x4e849419 // sdot v25.4s, v0.16b, v4.16b\n"
+ ".inst 0x4e849605 // sdot v5.4s, v16.16b, v4.16b\n"
+ "ext v4.16b, v4.16b, v4.16b, #0x1\n"
+ "ext v9.16b, v9.16b, v9.16b, #0x1\n"
+ ".inst 0x4e9d9619 // sdot v25.4s, v16.16b, v29.16b\n"
+ ".inst 0x4e9d97e5 // sdot v5.4s, v31.16b, v29.16b\n"
+ "ext v29.16b, v29.16b, v29.16b, #0x1\n"
+ ".inst 0x4e89941e // sdot v30.4s, v0.16b, v9.16b\n"
+ ".inst 0x4e849414 // sdot v20.4s, v0.16b, v4.16b\n"
+ "movi v17.4s, #0x0\n"
+ ".inst 0x4e8495d1 // sdot v17.4s, v14.16b, v4.16b\n"
+ ".inst 0x4e9d95d1 // sdot v17.4s, v14.16b, v29.16b\n"
+ ".inst 0x4e9897f9 // sdot v25.4s, v31.16b, v24.16b\n"
+ "ext v24.16b, v24.16b, v24.16b, #0x1\n"
+ ".inst 0x4e84961e // sdot v30.4s, v16.16b, v4.16b\n"
+ "ldr q4, [%x[params], #0x50]\n"
+ ".inst 0x4e9d9614 // sdot v20.4s, v16.16b, v29.16b\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9895d0 // sdot v16.4s, v14.16b, v24.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ ".inst 0x4e8995d1 // sdot v17.4s, v14.16b, v9.16b\n"
+ "ldr q9, [%x[params], #0x40]\n"
+ ".inst 0x4e9d97fe // sdot v30.4s, v31.16b, v29.16b\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ ".inst 0x4e9897f4 // sdot v20.4s, v31.16b, v24.16b\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
"add %x[params], %x[params], #0x60\n"
- "zip1 v1.16b, v28.16b, v26.16b\n"
- "zip2 v31.16b, v28.16b, v26.16b\n"
- "zip1 v28.16b, v24.16b, v20.16b\n"
- "zip2 v26.16b, v24.16b, v20.16b\n"
- "zip1 v24.16b, v22.16b, v16.16b\n"
- "zip2 v22.16b, v22.16b, v16.16b\n"
- "zip1 v20.16b, v19.16b, v18.16b\n"
- "zip2 v19.16b, v19.16b, v18.16b\n"
- "zip1 v18.16b, v0.16b, v17.16b\n"
- "zip2 v17.16b, v0.16b, v17.16b\n"
- "zip1 v6.16b, v7.16b, v8.16b\n"
- "zip2 v8.16b, v7.16b, v8.16b\n"
- "zip1 v7.16b, v5.16b, v3.16b\n"
- "str q7, [SP, #0x0]\n"
- "zip2 v5.16b, v5.16b, v3.16b\n"
- "str q5, [SP, #0x10]\n"
- "zip1 v3.16b, v2.16b, v1.16b\n"
- "zip2 v2.16b, v2.16b, v1.16b\n"
- "zip1 v1.16b, v4.16b, v31.16b\n"
- "str q1, [SP, #0x20]\n"
- "zip2 v16.16b, v4.16b, v31.16b\n"
- "str q16, [SP, #0x30]\n"
- "zip1 v31.16b, v28.16b, v24.16b\n"
- "zip2 v28.16b, v28.16b, v24.16b\n"
- "zip1 v16.16b, v26.16b, v22.16b\n"
- "str q16, [SP, #0x40]\n"
- "zip2 v16.16b, v26.16b, v22.16b\n"
- "str q16, [SP, #0x50]\n"
- "zip1 v26.16b, v20.16b, v18.16b\n"
- "zip2 v24.16b, v20.16b, v18.16b\n"
- "zip1 v16.16b, v19.16b, v17.16b\n"
- "str q16, [SP, #0x60]\n"
- "zip2 v16.16b, v19.16b, v17.16b\n"
- "str q16, [SP, #0x70]\n"
- "mov v22.16b, v30.16b\n"
- "mov v20.16b, v30.16b\n"
- "mov v19.16b, v30.16b\n"
- ".inst 0x4e8697be // sdot v30.4s, v29.16b, v6.16b\n"
- ".inst 0x4e8397b4 // sdot v20.4s, v29.16b, v3.16b\n"
- "movi v15.4s, #0x0\n"
- ".inst 0x4e83956f // sdot v15.4s, v11.16b, v3.16b\n"
- ".inst 0x4e83977e // sdot v30.4s, v27.16b, v3.16b\n"
- ".inst 0x4e9f9774 // sdot v20.4s, v27.16b, v31.16b\n"
- "ext v3.16b, v3.16b, v3.16b, #0x1\n"
- ".inst 0x4e9f956f // sdot v15.4s, v11.16b, v31.16b\n"
- ".inst 0x4e9f973e // sdot v30.4s, v25.16b, v31.16b\n"
- ".inst 0x4e9a9734 // sdot v20.4s, v25.16b, v26.16b\n"
- "ext v31.16b, v31.16b, v31.16b, #0x1\n"
- "mov v17.16b, v15.16b\n"
- ".inst 0x4e86956f // sdot v15.4s, v11.16b, v6.16b\n"
- "mls v30.4s, v15.4s, v14.4s\n"
- ".inst 0x4e9a9571 // sdot v17.4s, v11.16b, v26.16b\n"
- "ext v6.16b, v6.16b, v6.16b, #0x1\n"
- "mls v20.4s, v17.4s, v14.4s\n"
- "ext v26.16b, v26.16b, v26.16b, #0x1\n"
- ".inst 0x4e8697b6 // sdot v22.4s, v29.16b, v6.16b\n"
- ".inst 0x4e8397b3 // sdot v19.4s, v29.16b, v3.16b\n"
- "movi v10.4s, #0x0\n"
- ".inst 0x4e83956a // sdot v10.4s, v11.16b, v3.16b\n"
- ".inst 0x4e839776 // sdot v22.4s, v27.16b, v3.16b\n"
- ".inst 0x4e9f9773 // sdot v19.4s, v27.16b, v31.16b\n"
- "sqrdmulh v30.4s, v30.4s, v23.4s\n"
- ".inst 0x4e9f956a // sdot v10.4s, v11.16b, v31.16b\n"
- ".inst 0x4e9f9736 // sdot v22.4s, v25.16b, v31.16b\n"
- ".inst 0x4e9a9733 // sdot v19.4s, v25.16b, v26.16b\n"
- "and v18.16b, v30.16b, v21.16b\n"
- "sshr v18.4s, v18.4s, #0x1f\n"
- "mov v17.16b, v10.16b\n"
- ".inst 0x4e86956a // sdot v10.4s, v11.16b, v6.16b\n"
- "mls v22.4s, v10.4s, v14.4s\n"
- ".inst 0x4e9a9571 // sdot v17.4s, v11.16b, v26.16b\n"
- "sqrdmulh v20.4s, v20.4s, v23.4s\n"
- "mls v19.4s, v17.4s, v14.4s\n"
- "sqadd v30.4s, v30.4s, v18.4s\n"
- "and v16.16b, v20.16b, v21.16b\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "sqrdmulh v22.4s, v22.4s, v23.4s\n"
- "srshl v30.4s, v30.4s, v21.4s\n"
- "sqrdmulh v19.4s, v19.4s, v23.4s\n"
- "and v17.16b, v22.16b, v21.16b\n"
- "sshr v17.4s, v17.4s, #0x1f\n"
- "add v30.4s, v30.4s, v13.4s\n"
- "sqadd v20.4s, v20.4s, v16.4s\n"
- "and v16.16b, v19.16b, v21.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "smax v30.4s, v30.4s, v9.4s\n"
- "srshl v20.4s, v20.4s, v21.4s\n"
- "sqadd v22.4s, v22.4s, v17.4s\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
"smin v30.4s, v30.4s, v12.4s\n"
- "add v20.4s, v20.4s, v13.4s\n"
- "srshl v22.4s, v22.4s, v21.4s\n"
- "sqadd v19.4s, v19.4s, v16.4s\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "smax v20.4s, v20.4s, v9.4s\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "add v22.4s, v22.4s, v13.4s\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
"smin v20.4s, v20.4s, v12.4s\n"
- "srshl v19.4s, v19.4s, v21.4s\n"
- "smax v22.4s, v22.4s, v9.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "smin v22.4s, v22.4s, v12.4s\n"
- "add v19.4s, v19.4s, v13.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "smax v19.4s, v19.4s, v9.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "smin v19.4s, v19.4s, v12.4s\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "blt 19f\n"
- "str s30, [x25, x23]\n"
- "str s22, [x24, x23]\n"
- "str s20, [x22, x23]\n"
- "str s19, [x21, x23]\n"
+ "blt 20f\n"
+ "str s5, [x24, x27]\n"
+ "str s30, [x23, x27]\n"
+ "str s25, [x22, x27]\n"
+ "str s20, [x21, x27]\n"
+ "b 23f\n"
+ "20:" // Oddments: Unroll 0: Oddment store
+ "add x24, x24, x27\n"
+ "add x23, x23, x27\n"
+ "add x22, x22, x27\n"
+ "add x21, x21, x27\n"
+ "tbz x20, #1, 21f\n"
+ "st1 { v5.h }[0], [x24], #0x2\n"
+ "st1 { v30.h }[0], [x23], #0x2\n"
+ "st1 { v25.h }[0], [x22], #0x2\n"
+ "st1 { v20.h }[0], [x21], #0x2\n"
+ "tbz x20, #0, 22f\n"
+ "st1 { v5.b }[2], [x24], #0x1\n"
+ "st1 { v30.b }[2], [x23], #0x1\n"
+ "st1 { v25.b }[2], [x22], #0x1\n"
+ "st1 { v20.b }[2], [x21], #0x1\n"
"b 22f\n"
- "19:" // Oddments: Unroll 0: Oddment store
- "add x25, x25, x23\n"
- "add x24, x24, x23\n"
- "add x22, x22, x23\n"
- "add x21, x21, x23\n"
- "tbz x19, #1, 20f\n"
- "st1 { v30.h }[0], [x25], #0x2\n"
- "st1 { v22.h }[0], [x24], #0x2\n"
- "st1 { v20.h }[0], [x22], #0x2\n"
- "st1 { v19.h }[0], [x21], #0x2\n"
- "tbz x19, #0, 21f\n"
- "st1 { v30.b }[2], [x25], #0x1\n"
- "st1 { v22.b }[2], [x24], #0x1\n"
- "st1 { v20.b }[2], [x22], #0x1\n"
- "st1 { v19.b }[2], [x21], #0x1\n"
- "b 21f\n"
- "20:" // Oddments: Unroll 0: Oddment store: Bit 1: Unset
- "tbz x19, #0, 21f\n"
- "st1 { v30.b }[0], [x25], #0x1\n"
- "st1 { v22.b }[0], [x24], #0x1\n"
- "st1 { v20.b }[0], [x22], #0x1\n"
- "st1 { v19.b }[0], [x21], #0x1\n"
- "21:" // Oddments: Unroll 0: Oddment store: Bit 1: End
-
- "22:" // Oddments: Unroll 0: After oddment store
- "add x23, x23, #0x4\n"
- "subs x19, x19, #0x4\n"
- "ble 34f\n"
- "movi v15.4s, #0x0\n"
- "ldr q30, [%x[params], #0x0]\n"
- ".inst 0x4e82956f // sdot v15.4s, v11.16b, v2.16b\n"
- "ldr q29, [%x[params], #0x10]\n"
- "cmp x19, #0x4\n"
- "movi v10.4s, #0x0\n"
- "ldr q27, [%x[params], #0x20]\n"
- "ldr q25, [%x[params], #0x30]\n"
- "mov v22.16b, v30.16b\n"
- "ldr q23, [%x[params], #0x40]\n"
- "mov v20.16b, v30.16b\n"
- "ldr q21, [%x[params], #0x50]\n"
+ "21:" // Oddments: Unroll 0: Oddment store: Bit 1: Unset
+ "st1 { v5.b }[0], [x24], #0x1\n"
+ "st1 { v30.b }[0], [x23], #0x1\n"
+ "st1 { v25.b }[0], [x22], #0x1\n"
+ "st1 { v20.b }[0], [x21], #0x1\n"
+ "22:" // Oddments: Unroll 0: Oddment store: Bit 1: End
+ "23:" // Oddments: Unroll 0: After oddment store
+ "subs x20, x20, #0x4\n"
+ "add x27, x27, #0x4\n"
+ "ble 35f\n"
+ "ldr q5, [%x[params], #0x0]\n"
+ "ldr q0, [%x[params], #0x10]\n"
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8395d3 // sdot v19.4s, v14.16b, v3.16b\n"
+ "ldr q16, [%x[params], #0x20]\n"
+ "ldr q31, [%x[params], #0x30]\n"
+ "mov v30.16b, v5.16b\n"
+ "mov v25.16b, v5.16b\n"
+ "ldr q9, [%x[params], #0x40]\n"
+ "ldr q4, [%x[params], #0x50]\n"
+ "mov v20.16b, v5.16b\n"
+ ".inst 0x4e889405 // sdot v5.4s, v0.16b, v8.16b\n"
+ ".inst 0x4e9c95d3 // sdot v19.4s, v14.16b, v28.16b\n"
+ ".inst 0x4e839419 // sdot v25.4s, v0.16b, v3.16b\n"
+ "movi v17.4s, #0x0\n"
+ "cmp x20, #0x4\n"
+ ".inst 0x4e839605 // sdot v5.4s, v16.16b, v3.16b\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9795d2 // sdot v18.4s, v14.16b, v23.16b\n"
+ "ext v3.16b, v3.16b, v3.16b, #0x1\n"
"add %x[params], %x[params], #0x60\n"
- "mov v19.16b, v30.16b\n"
- ".inst 0x4e8897be // sdot v30.4s, v29.16b, v8.16b\n"
- ".inst 0x4e8297b4 // sdot v20.4s, v29.16b, v2.16b\n"
- ".inst 0x4e9c956f // sdot v15.4s, v11.16b, v28.16b\n"
- ".inst 0x4e82977e // sdot v30.4s, v27.16b, v2.16b\n"
- "ext v2.16b, v2.16b, v2.16b, #0x1\n"
- ".inst 0x4e9c9774 // sdot v20.4s, v27.16b, v28.16b\n"
- "mov v17.16b, v15.16b\n"
- ".inst 0x4e88956f // sdot v15.4s, v11.16b, v8.16b\n"
- ".inst 0x4e9c973e // sdot v30.4s, v25.16b, v28.16b\n"
- "mls v30.4s, v15.4s, v14.4s\n"
- ".inst 0x4e989734 // sdot v20.4s, v25.16b, v24.16b\n"
- ".inst 0x4e989571 // sdot v17.4s, v11.16b, v24.16b\n"
- "mls v20.4s, v17.4s, v14.4s\n"
+ ".inst 0x4e8895d3 // sdot v19.4s, v14.16b, v8.16b\n"
"ext v8.16b, v8.16b, v8.16b, #0x1\n"
+ ".inst 0x4e88941e // sdot v30.4s, v0.16b, v8.16b\n"
+ ".inst 0x4e839414 // sdot v20.4s, v0.16b, v3.16b\n"
+ ".inst 0x4e8395d1 // sdot v17.4s, v14.16b, v3.16b\n"
+ ".inst 0x4e9c9619 // sdot v25.4s, v16.16b, v28.16b\n"
+ ".inst 0x4e9c97e5 // sdot v5.4s, v31.16b, v28.16b\n"
"ext v28.16b, v28.16b, v28.16b, #0x1\n"
- "ext v24.16b, v24.16b, v24.16b, #0x1\n"
- ".inst 0x4e8297b3 // sdot v19.4s, v29.16b, v2.16b\n"
- ".inst 0x4e82956a // sdot v10.4s, v11.16b, v2.16b\n"
- ".inst 0x4e8897b6 // sdot v22.4s, v29.16b, v8.16b\n"
- "sqrdmulh v30.4s, v30.4s, v23.4s\n"
- ".inst 0x4e9c9773 // sdot v19.4s, v27.16b, v28.16b\n"
- ".inst 0x4e9c956a // sdot v10.4s, v11.16b, v28.16b\n"
- ".inst 0x4e829776 // sdot v22.4s, v27.16b, v2.16b\n"
- "and v18.16b, v30.16b, v21.16b\n"
- "sshr v18.4s, v18.4s, #0x1f\n"
- ".inst 0x4e9c9736 // sdot v22.4s, v25.16b, v28.16b\n"
- ".inst 0x4e989733 // sdot v19.4s, v25.16b, v24.16b\n"
- "mov v17.16b, v10.16b\n"
- ".inst 0x4e88956a // sdot v10.4s, v11.16b, v8.16b\n"
- "mls v22.4s, v10.4s, v14.4s\n"
- ".inst 0x4e989571 // sdot v17.4s, v11.16b, v24.16b\n"
- "sqadd v30.4s, v30.4s, v18.4s\n"
- "mls v19.4s, v17.4s, v14.4s\n"
- "srshl v30.4s, v30.4s, v21.4s\n"
- "sqrdmulh v20.4s, v20.4s, v23.4s\n"
- "sqrdmulh v22.4s, v22.4s, v23.4s\n"
- "add v30.4s, v30.4s, v13.4s\n"
- "and v16.16b, v20.16b, v21.16b\n"
+ ".inst 0x4e83961e // sdot v30.4s, v16.16b, v3.16b\n"
+ ".inst 0x4e9c9614 // sdot v20.4s, v16.16b, v28.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ ".inst 0x4e9c95d1 // sdot v17.4s, v14.16b, v28.16b\n"
+ ".inst 0x4e9797f9 // sdot v25.4s, v31.16b, v23.16b\n"
+ "ext v23.16b, v23.16b, v23.16b, #0x1\n"
+ ".inst 0x4e9c97fe // sdot v30.4s, v31.16b, v28.16b\n"
+ ".inst 0x4e9797f4 // sdot v20.4s, v31.16b, v23.16b\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9795d0 // sdot v16.4s, v14.16b, v23.16b\n"
+ ".inst 0x4e8895d1 // sdot v17.4s, v14.16b, v8.16b\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "smax v30.4s, v30.4s, v9.4s\n"
- "and v17.16b, v22.16b, v21.16b\n"
- "sshr v17.4s, v17.4s, #0x1f\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
"smin v30.4s, v30.4s, v12.4s\n"
- "sqrdmulh v19.4s, v19.4s, v23.4s\n"
- "sqadd v20.4s, v20.4s, v16.4s\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "and v16.16b, v19.16b, v21.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sqadd v22.4s, v22.4s, v17.4s\n"
- "srshl v20.4s, v20.4s, v21.4s\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "srshl v22.4s, v22.4s, v21.4s\n"
- "add v20.4s, v20.4s, v13.4s\n"
- "sqadd v19.4s, v19.4s, v16.4s\n"
- "smax v20.4s, v20.4s, v9.4s\n"
- "add v22.4s, v22.4s, v13.4s\n"
- "srshl v19.4s, v19.4s, v21.4s\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
"smin v20.4s, v20.4s, v12.4s\n"
- "smax v22.4s, v22.4s, v9.4s\n"
- "add v19.4s, v19.4s, v13.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "smin v22.4s, v22.4s, v12.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "smax v19.4s, v19.4s, v9.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "smin v19.4s, v19.4s, v12.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "blt 23f\n"
- "str s30, [x25, x23]\n"
- "str s22, [x24, x23]\n"
- "str s20, [x22, x23]\n"
- "str s19, [x21, x23]\n"
+ "blt 24f\n"
+ "str s5, [x24, x27]\n"
+ "str s30, [x23, x27]\n"
+ "str s25, [x22, x27]\n"
+ "str s20, [x21, x27]\n"
+ "b 27f\n"
+ "24:" // Oddments: Unroll 1: Oddment store
+ "add x24, x24, x27\n"
+ "add x23, x23, x27\n"
+ "add x22, x22, x27\n"
+ "add x21, x21, x27\n"
+ "tbz x20, #1, 25f\n"
+ "st1 { v5.h }[0], [x24], #0x2\n"
+ "st1 { v30.h }[0], [x23], #0x2\n"
+ "st1 { v25.h }[0], [x22], #0x2\n"
+ "st1 { v20.h }[0], [x21], #0x2\n"
+ "tbz x20, #0, 26f\n"
+ "st1 { v5.b }[2], [x24], #0x1\n"
+ "st1 { v30.b }[2], [x23], #0x1\n"
+ "st1 { v25.b }[2], [x22], #0x1\n"
+ "st1 { v20.b }[2], [x21], #0x1\n"
"b 26f\n"
- "23:" // Oddments: Unroll 1: Oddment store
- "add x25, x25, x23\n"
- "add x24, x24, x23\n"
- "add x22, x22, x23\n"
- "add x21, x21, x23\n"
- "tbz x19, #1, 24f\n"
- "st1 { v30.h }[0], [x25], #0x2\n"
- "st1 { v22.h }[0], [x24], #0x2\n"
- "st1 { v20.h }[0], [x22], #0x2\n"
- "st1 { v19.h }[0], [x21], #0x2\n"
- "tbz x19, #0, 25f\n"
- "st1 { v30.b }[2], [x25], #0x1\n"
- "st1 { v22.b }[2], [x24], #0x1\n"
- "st1 { v20.b }[2], [x22], #0x1\n"
- "st1 { v19.b }[2], [x21], #0x1\n"
- "b 25f\n"
- "24:" // Oddments: Unroll 1: Oddment store: Bit 1: Unset
- "tbz x19, #0, 25f\n"
- "st1 { v30.b }[0], [x25], #0x1\n"
- "st1 { v22.b }[0], [x24], #0x1\n"
- "st1 { v20.b }[0], [x22], #0x1\n"
- "st1 { v19.b }[0], [x21], #0x1\n"
- "25:" // Oddments: Unroll 1: Oddment store: Bit 1: End
-
- "26:" // Oddments: Unroll 1: After oddment store
- "add x23, x23, #0x4\n"
- "subs x19, x19, #0x4\n"
- "ble 34f\n"
- "movi v15.4s, #0x0\n"
- "ldr q6, [SP, #0x0]\n"
- "movi v10.4s, #0x0\n"
- "ldr q3, [SP, #0x20]\n"
- "cmp x19, #0x4\n"
- ".inst 0x4e83956f // sdot v15.4s, v11.16b, v3.16b\n"
- "ldr q31, [SP, #0x40]\n"
- "ldr q26, [SP, #0x60]\n"
- ".inst 0x4e9f956f // sdot v15.4s, v11.16b, v31.16b\n"
- "ldr q30, [%x[params], #0x0]\n"
- "ldr q29, [%x[params], #0x10]\n"
- "mov v22.16b, v30.16b\n"
- "ldr q27, [%x[params], #0x20]\n"
- "mov v20.16b, v30.16b\n"
- "ldr q25, [%x[params], #0x30]\n"
- "mov v19.16b, v30.16b\n"
- "ldr q23, [%x[params], #0x40]\n"
- ".inst 0x4e8697be // sdot v30.4s, v29.16b, v6.16b\n"
- "ldr q21, [%x[params], #0x50]\n"
+ "25:" // Oddments: Unroll 1: Oddment store: Bit 1: Unset
+ "st1 { v5.b }[0], [x24], #0x1\n"
+ "st1 { v30.b }[0], [x23], #0x1\n"
+ "st1 { v25.b }[0], [x22], #0x1\n"
+ "st1 { v20.b }[0], [x21], #0x1\n"
+ "26:" // Oddments: Unroll 1: Oddment store: Bit 1: End
+ "27:" // Oddments: Unroll 1: After oddment store
+ "subs x20, x20, #0x4\n"
+ "add x27, x27, #0x4\n"
+ "ble 35f\n"
+ "ldr q5, [%x[params], #0x0]\n"
+ "ldr q0, [%x[params], #0x10]\n"
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8295d3 // sdot v19.4s, v14.16b, v2.16b\n"
+ "ldr q16, [%x[params], #0x20]\n"
+ "ldr q31, [%x[params], #0x30]\n"
+ "mov v30.16b, v5.16b\n"
+ "mov v25.16b, v5.16b\n"
+ "ldr q9, [%x[params], #0x40]\n"
+ "ldr q4, [%x[params], #0x50]\n"
+ "mov v20.16b, v5.16b\n"
+ ".inst 0x4e879405 // sdot v5.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e9b95d3 // sdot v19.4s, v14.16b, v27.16b\n"
+ ".inst 0x4e829419 // sdot v25.4s, v0.16b, v2.16b\n"
+ "movi v17.4s, #0x0\n"
+ "cmp x20, #0x4\n"
+ ".inst 0x4e829605 // sdot v5.4s, v16.16b, v2.16b\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9695d2 // sdot v18.4s, v14.16b, v22.16b\n"
+ "ext v2.16b, v2.16b, v2.16b, #0x1\n"
"add %x[params], %x[params], #0x60\n"
- ".inst 0x4e8397b4 // sdot v20.4s, v29.16b, v3.16b\n"
- "mov v17.16b, v15.16b\n"
- ".inst 0x4e86956f // sdot v15.4s, v11.16b, v6.16b\n"
- ".inst 0x4e83977e // sdot v30.4s, v27.16b, v3.16b\n"
- ".inst 0x4e9a9571 // sdot v17.4s, v11.16b, v26.16b\n"
- ".inst 0x4e9f9774 // sdot v20.4s, v27.16b, v31.16b\n"
- "ext v6.16b, v6.16b, v6.16b, #0x1\n"
- ".inst 0x4e9f973e // sdot v30.4s, v25.16b, v31.16b\n"
- "mls v30.4s, v15.4s, v14.4s\n"
- ".inst 0x4e9a9734 // sdot v20.4s, v25.16b, v26.16b\n"
- "ext v3.16b, v3.16b, v3.16b, #0x1\n"
- "mls v20.4s, v17.4s, v14.4s\n"
- "ext v31.16b, v31.16b, v31.16b, #0x1\n"
- "ext v26.16b, v26.16b, v26.16b, #0x1\n"
- ".inst 0x4e8697b6 // sdot v22.4s, v29.16b, v6.16b\n"
- ".inst 0x4e8397b3 // sdot v19.4s, v29.16b, v3.16b\n"
- ".inst 0x4e83956a // sdot v10.4s, v11.16b, v3.16b\n"
- "sqrdmulh v30.4s, v30.4s, v23.4s\n"
- ".inst 0x4e839776 // sdot v22.4s, v27.16b, v3.16b\n"
- ".inst 0x4e9f9773 // sdot v19.4s, v27.16b, v31.16b\n"
- ".inst 0x4e9f956a // sdot v10.4s, v11.16b, v31.16b\n"
- "and v18.16b, v30.16b, v21.16b\n"
- "sshr v18.4s, v18.4s, #0x1f\n"
- ".inst 0x4e9f9736 // sdot v22.4s, v25.16b, v31.16b\n"
- ".inst 0x4e9a9733 // sdot v19.4s, v25.16b, v26.16b\n"
- "mov v17.16b, v10.16b\n"
- ".inst 0x4e86956a // sdot v10.4s, v11.16b, v6.16b\n"
- "mls v22.4s, v10.4s, v14.4s\n"
- ".inst 0x4e9a9571 // sdot v17.4s, v11.16b, v26.16b\n"
- "sqadd v30.4s, v30.4s, v18.4s\n"
- "mls v19.4s, v17.4s, v14.4s\n"
- "srshl v30.4s, v30.4s, v21.4s\n"
- "sqrdmulh v20.4s, v20.4s, v23.4s\n"
- "sqrdmulh v22.4s, v22.4s, v23.4s\n"
- "add v30.4s, v30.4s, v13.4s\n"
- "and v16.16b, v20.16b, v21.16b\n"
+ ".inst 0x4e8795d3 // sdot v19.4s, v14.16b, v7.16b\n"
+ "ext v7.16b, v7.16b, v7.16b, #0x1\n"
+ ".inst 0x4e87941e // sdot v30.4s, v0.16b, v7.16b\n"
+ ".inst 0x4e829414 // sdot v20.4s, v0.16b, v2.16b\n"
+ ".inst 0x4e8295d1 // sdot v17.4s, v14.16b, v2.16b\n"
+ ".inst 0x4e9b9619 // sdot v25.4s, v16.16b, v27.16b\n"
+ ".inst 0x4e9b97e5 // sdot v5.4s, v31.16b, v27.16b\n"
+ "ext v27.16b, v27.16b, v27.16b, #0x1\n"
+ ".inst 0x4e82961e // sdot v30.4s, v16.16b, v2.16b\n"
+ ".inst 0x4e9b9614 // sdot v20.4s, v16.16b, v27.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ ".inst 0x4e9b95d1 // sdot v17.4s, v14.16b, v27.16b\n"
+ ".inst 0x4e9697f9 // sdot v25.4s, v31.16b, v22.16b\n"
+ "ext v22.16b, v22.16b, v22.16b, #0x1\n"
+ ".inst 0x4e9b97fe // sdot v30.4s, v31.16b, v27.16b\n"
+ ".inst 0x4e9697f4 // sdot v20.4s, v31.16b, v22.16b\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9695d0 // sdot v16.4s, v14.16b, v22.16b\n"
+ ".inst 0x4e8795d1 // sdot v17.4s, v14.16b, v7.16b\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "smax v30.4s, v30.4s, v9.4s\n"
- "and v17.16b, v22.16b, v21.16b\n"
- "sshr v17.4s, v17.4s, #0x1f\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
"smin v30.4s, v30.4s, v12.4s\n"
- "sqrdmulh v19.4s, v19.4s, v23.4s\n"
- "sqadd v20.4s, v20.4s, v16.4s\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "and v16.16b, v19.16b, v21.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sqadd v22.4s, v22.4s, v17.4s\n"
- "srshl v20.4s, v20.4s, v21.4s\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "srshl v22.4s, v22.4s, v21.4s\n"
- "add v20.4s, v20.4s, v13.4s\n"
- "sqadd v19.4s, v19.4s, v16.4s\n"
- "smax v20.4s, v20.4s, v9.4s\n"
- "add v22.4s, v22.4s, v13.4s\n"
- "srshl v19.4s, v19.4s, v21.4s\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
"smin v20.4s, v20.4s, v12.4s\n"
- "smax v22.4s, v22.4s, v9.4s\n"
- "add v19.4s, v19.4s, v13.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "smin v22.4s, v22.4s, v12.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "smax v19.4s, v19.4s, v9.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "smin v19.4s, v19.4s, v12.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "blt 27f\n"
- "str s30, [x25, x23]\n"
- "str s22, [x24, x23]\n"
- "str s20, [x22, x23]\n"
- "str s19, [x21, x23]\n"
+ "blt 28f\n"
+ "str s5, [x24, x27]\n"
+ "str s30, [x23, x27]\n"
+ "str s25, [x22, x27]\n"
+ "str s20, [x21, x27]\n"
+ "b 31f\n"
+ "28:" // Oddments: Unroll 2: Oddment store
+ "add x24, x24, x27\n"
+ "add x23, x23, x27\n"
+ "add x22, x22, x27\n"
+ "add x21, x21, x27\n"
+ "tbz x20, #1, 29f\n"
+ "st1 { v5.h }[0], [x24], #0x2\n"
+ "st1 { v30.h }[0], [x23], #0x2\n"
+ "st1 { v25.h }[0], [x22], #0x2\n"
+ "st1 { v20.h }[0], [x21], #0x2\n"
+ "tbz x20, #0, 30f\n"
+ "st1 { v5.b }[2], [x24], #0x1\n"
+ "st1 { v30.b }[2], [x23], #0x1\n"
+ "st1 { v25.b }[2], [x22], #0x1\n"
+ "st1 { v20.b }[2], [x21], #0x1\n"
"b 30f\n"
- "27:" // Oddments: Unroll 2: Oddment store
- "add x25, x25, x23\n"
- "add x24, x24, x23\n"
- "add x22, x22, x23\n"
- "add x21, x21, x23\n"
- "tbz x19, #1, 28f\n"
- "st1 { v30.h }[0], [x25], #0x2\n"
- "st1 { v22.h }[0], [x24], #0x2\n"
- "st1 { v20.h }[0], [x22], #0x2\n"
- "st1 { v19.h }[0], [x21], #0x2\n"
- "tbz x19, #0, 29f\n"
- "st1 { v30.b }[2], [x25], #0x1\n"
- "st1 { v22.b }[2], [x24], #0x1\n"
- "st1 { v20.b }[2], [x22], #0x1\n"
- "st1 { v19.b }[2], [x21], #0x1\n"
- "b 29f\n"
- "28:" // Oddments: Unroll 2: Oddment store: Bit 1: Unset
- "tbz x19, #0, 29f\n"
- "st1 { v30.b }[0], [x25], #0x1\n"
- "st1 { v22.b }[0], [x24], #0x1\n"
- "st1 { v20.b }[0], [x22], #0x1\n"
- "st1 { v19.b }[0], [x21], #0x1\n"
- "29:" // Oddments: Unroll 2: Oddment store: Bit 1: End
-
- "30:" // Oddments: Unroll 2: After oddment store
- "add x23, x23, #0x4\n"
- "subs x19, x19, #0x4\n"
- "ble 34f\n"
- "movi v15.4s, #0x0\n"
- "ldr q8, [SP, #0x10]\n"
- "movi v10.4s, #0x0\n"
- "ldr q2, [SP, #0x30]\n"
- "ldr q28, [SP, #0x50]\n"
- ".inst 0x4e82956f // sdot v15.4s, v11.16b, v2.16b\n"
- "ldr q24, [SP, #0x70]\n"
- "ldr q30, [%x[params], #0x0]\n"
- "mov v22.16b, v30.16b\n"
- "ldr q29, [%x[params], #0x10]\n"
- "mov v20.16b, v30.16b\n"
- "ldr q27, [%x[params], #0x20]\n"
- "mov v19.16b, v30.16b\n"
- "ldr q25, [%x[params], #0x30]\n"
- ".inst 0x4e9c956f // sdot v15.4s, v11.16b, v28.16b\n"
- "ldr q23, [%x[params], #0x40]\n"
- "ldr q21, [%x[params], #0x50]\n"
- ".inst 0x4e8897be // sdot v30.4s, v29.16b, v8.16b\n"
+ "29:" // Oddments: Unroll 2: Oddment store: Bit 1: Unset
+ "st1 { v5.b }[0], [x24], #0x1\n"
+ "st1 { v30.b }[0], [x23], #0x1\n"
+ "st1 { v25.b }[0], [x22], #0x1\n"
+ "st1 { v20.b }[0], [x21], #0x1\n"
+ "30:" // Oddments: Unroll 2: Oddment store: Bit 1: End
+ "31:" // Oddments: Unroll 2: After oddment store
+ "subs x20, x20, #0x4\n"
+ "add x27, x27, #0x4\n"
+ "ble 35f\n"
+ "ldr q5, [%x[params], #0x0]\n"
+ "ldr q0, [%x[params], #0x10]\n"
+ "movi v19.4s, #0x0\n"
+ ".inst 0x4e8195d3 // sdot v19.4s, v14.16b, v1.16b\n"
+ "ldr q16, [%x[params], #0x20]\n"
+ "ldr q31, [%x[params], #0x30]\n"
+ "mov v30.16b, v5.16b\n"
+ "mov v25.16b, v5.16b\n"
+ "ldr q9, [%x[params], #0x40]\n"
+ "ldr q4, [%x[params], #0x50]\n"
+ "mov v20.16b, v5.16b\n"
+ ".inst 0x4e869405 // sdot v5.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e9a95d3 // sdot v19.4s, v14.16b, v26.16b\n"
+ ".inst 0x4e819419 // sdot v25.4s, v0.16b, v1.16b\n"
+ "movi v17.4s, #0x0\n"
"add %x[params], %x[params], #0x60\n"
- ".inst 0x4e8297b4 // sdot v20.4s, v29.16b, v2.16b\n"
- "mov v17.16b, v15.16b\n"
- ".inst 0x4e88956f // sdot v15.4s, v11.16b, v8.16b\n"
- ".inst 0x4e989571 // sdot v17.4s, v11.16b, v24.16b\n"
- ".inst 0x4e82977e // sdot v30.4s, v27.16b, v2.16b\n"
- ".inst 0x4e9c9774 // sdot v20.4s, v27.16b, v28.16b\n"
- "ext v8.16b, v8.16b, v8.16b, #0x1\n"
- "ext v2.16b, v2.16b, v2.16b, #0x1\n"
- ".inst 0x4e9c973e // sdot v30.4s, v25.16b, v28.16b\n"
- "mls v30.4s, v15.4s, v14.4s\n"
- ".inst 0x4e989734 // sdot v20.4s, v25.16b, v24.16b\n"
- "ext v28.16b, v28.16b, v28.16b, #0x1\n"
- "mls v20.4s, v17.4s, v14.4s\n"
- "ext v24.16b, v24.16b, v24.16b, #0x1\n"
- ".inst 0x4e8897b6 // sdot v22.4s, v29.16b, v8.16b\n"
- ".inst 0x4e8297b3 // sdot v19.4s, v29.16b, v2.16b\n"
- ".inst 0x4e82956a // sdot v10.4s, v11.16b, v2.16b\n"
- "sqrdmulh v30.4s, v30.4s, v23.4s\n"
- ".inst 0x4e829776 // sdot v22.4s, v27.16b, v2.16b\n"
- ".inst 0x4e9c9773 // sdot v19.4s, v27.16b, v28.16b\n"
- ".inst 0x4e9c956a // sdot v10.4s, v11.16b, v28.16b\n"
- "and v18.16b, v30.16b, v21.16b\n"
- "sshr v18.4s, v18.4s, #0x1f\n"
- ".inst 0x4e9c9736 // sdot v22.4s, v25.16b, v28.16b\n"
- ".inst 0x4e989733 // sdot v19.4s, v25.16b, v24.16b\n"
- "mov v17.16b, v10.16b\n"
- ".inst 0x4e88956a // sdot v10.4s, v11.16b, v8.16b\n"
- "mls v22.4s, v10.4s, v14.4s\n"
- ".inst 0x4e989571 // sdot v17.4s, v11.16b, v24.16b\n"
- "sqadd v30.4s, v30.4s, v18.4s\n"
- "mls v19.4s, v17.4s, v14.4s\n"
- "srshl v30.4s, v30.4s, v21.4s\n"
- "sqrdmulh v20.4s, v20.4s, v23.4s\n"
- "sqrdmulh v22.4s, v22.4s, v23.4s\n"
- "add v30.4s, v30.4s, v13.4s\n"
- "and v16.16b, v20.16b, v21.16b\n"
+ ".inst 0x4e819605 // sdot v5.4s, v16.16b, v1.16b\n"
+ "mov v18.16b, v19.16b\n .inst 0x4e9595d2 // sdot v18.4s, v14.16b, v21.16b\n"
+ "ext v1.16b, v1.16b, v1.16b, #0x1\n"
+ ".inst 0x4e8695d3 // sdot v19.4s, v14.16b, v6.16b\n"
+ "ext v6.16b, v6.16b, v6.16b, #0x1\n"
+ ".inst 0x4e86941e // sdot v30.4s, v0.16b, v6.16b\n"
+ ".inst 0x4e819414 // sdot v20.4s, v0.16b, v1.16b\n"
+ ".inst 0x4e8195d1 // sdot v17.4s, v14.16b, v1.16b\n"
+ ".inst 0x4e9a9619 // sdot v25.4s, v16.16b, v26.16b\n"
+ ".inst 0x4e9a97e5 // sdot v5.4s, v31.16b, v26.16b\n"
+ "ext v26.16b, v26.16b, v26.16b, #0x1\n"
+ ".inst 0x4e81961e // sdot v30.4s, v16.16b, v1.16b\n"
+ ".inst 0x4e9a9614 // sdot v20.4s, v16.16b, v26.16b\n"
+ "mls v5.4s, v19.4s, v11.4s\n"
+ ".inst 0x4e9a95d1 // sdot v17.4s, v14.16b, v26.16b\n"
+ ".inst 0x4e9597f9 // sdot v25.4s, v31.16b, v21.16b\n"
+ "ext v21.16b, v21.16b, v21.16b, #0x1\n"
+ ".inst 0x4e9a97fe // sdot v30.4s, v31.16b, v26.16b\n"
+ ".inst 0x4e9597f4 // sdot v20.4s, v31.16b, v21.16b\n"
+ "sqrdmulh v5.4s, v5.4s, v9.4s\n"
+ "mov v16.16b, v17.16b\n .inst 0x4e9595d0 // sdot v16.4s, v14.16b, v21.16b\n"
+ ".inst 0x4e8695d1 // sdot v17.4s, v14.16b, v6.16b\n"
+ "mls v30.4s, v17.4s, v11.4s\n"
+ "mls v25.4s, v18.4s, v11.4s\n"
+ "mls v20.4s, v16.4s, v11.4s\n"
+ "and v0.16b, v5.16b, v4.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v30.4s, v30.4s, v9.4s\n"
+ "sqrdmulh v25.4s, v25.4s, v9.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v9.4s\n"
+ "sqadd v5.4s, v5.4s, v0.4s\n"
+ "and v16.16b, v30.16b, v4.16b\n"
+ "and v31.16b, v25.16b, v4.16b\n"
+ "and v0.16b, v20.16b, v4.16b\n"
"sshr v16.4s, v16.4s, #0x1f\n"
- "smax v30.4s, v30.4s, v9.4s\n"
- "and v17.16b, v22.16b, v21.16b\n"
- "sshr v17.4s, v17.4s, #0x1f\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqadd v30.4s, v30.4s, v16.4s\n"
+ "sqadd v25.4s, v25.4s, v31.4s\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "srshl v5.4s, v5.4s, v4.4s\n"
+ "srshl v30.4s, v30.4s, v4.4s\n"
+ "srshl v25.4s, v25.4s, v4.4s\n"
+ "srshl v20.4s, v20.4s, v4.4s\n"
+ "add v5.4s, v5.4s, v10.4s\n"
+ "add v30.4s, v30.4s, v10.4s\n"
+ "add v25.4s, v25.4s, v10.4s\n"
+ "add v20.4s, v20.4s, v10.4s\n"
+ "smax v5.4s, v5.4s, v13.4s\n"
+ "smax v30.4s, v30.4s, v13.4s\n"
+ "smax v25.4s, v25.4s, v13.4s\n"
+ "smax v20.4s, v20.4s, v13.4s\n"
+ "smin v5.4s, v5.4s, v12.4s\n"
"smin v30.4s, v30.4s, v12.4s\n"
- "sqrdmulh v19.4s, v19.4s, v23.4s\n"
- "sqadd v20.4s, v20.4s, v16.4s\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "and v16.16b, v19.16b, v21.16b\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "sqadd v22.4s, v22.4s, v17.4s\n"
- "srshl v20.4s, v20.4s, v21.4s\n"
- "uzp1 v30.16b, v30.16b, v30.16b\n"
- "srshl v22.4s, v22.4s, v21.4s\n"
- "add v20.4s, v20.4s, v13.4s\n"
- "sqadd v19.4s, v19.4s, v16.4s\n"
- "smax v20.4s, v20.4s, v9.4s\n"
- "add v22.4s, v22.4s, v13.4s\n"
- "srshl v19.4s, v19.4s, v21.4s\n"
+ "smin v25.4s, v25.4s, v12.4s\n"
"smin v20.4s, v20.4s, v12.4s\n"
- "smax v22.4s, v22.4s, v9.4s\n"
- "add v19.4s, v19.4s, v13.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "smin v22.4s, v22.4s, v12.4s\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "uzp1 v30.16b, v30.16b, v30.16b\n"
+ "uzp1 v25.16b, v25.16b, v25.16b\n"
"uzp1 v20.16b, v20.16b, v20.16b\n"
- "smax v19.4s, v19.4s, v9.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "smin v19.4s, v19.4s, v12.4s\n"
- "uzp1 v22.16b, v22.16b, v22.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "uzp1 v19.16b, v19.16b, v19.16b\n"
- "31:" // Oddments: Unroll 3: Oddment store
- "add x25, x25, x23\n"
- "add x24, x24, x23\n"
- "add x22, x22, x23\n"
- "add x21, x21, x23\n"
- "tbz x19, #1, 32f\n"
- "st1 { v30.h }[0], [x25], #0x2\n"
- "st1 { v22.h }[0], [x24], #0x2\n"
- "st1 { v20.h }[0], [x22], #0x2\n"
- "st1 { v19.h }[0], [x21], #0x2\n"
- "tbz x19, #0, 33f\n"
- "st1 { v30.b }[2], [x25], #0x1\n"
- "st1 { v22.b }[2], [x24], #0x1\n"
- "st1 { v20.b }[2], [x22], #0x1\n"
- "st1 { v19.b }[2], [x21], #0x1\n"
- "b 33f\n"
- "32:" // Oddments: Unroll 3: Oddment store: Bit 1: Unset
- "tbz x19, #0, 33f\n"
- "st1 { v30.b }[0], [x25], #0x1\n"
- "st1 { v22.b }[0], [x24], #0x1\n"
- "st1 { v20.b }[0], [x22], #0x1\n"
- "st1 { v19.b }[0], [x21], #0x1\n"
- "33:" // Oddments: Unroll 3: Oddment store: Bit 1: End
-
- "34:" // End
- "add SP, SP, #0x80\n"
+ "32:" // Oddments: Unroll 3: Oddment store
+ "add x24, x24, x27\n"
+ "add x23, x23, x27\n"
+ "add x22, x22, x27\n"
+ "add x21, x21, x27\n"
+ "tbz x20, #1, 33f\n"
+ "st1 { v5.h }[0], [x24], #0x2\n"
+ "st1 { v30.h }[0], [x23], #0x2\n"
+ "st1 { v25.h }[0], [x22], #0x2\n"
+ "st1 { v20.h }[0], [x21], #0x2\n"
+ "tbz x20, #0, 34f\n"
+ "st1 { v5.b }[2], [x24], #0x1\n"
+ "st1 { v30.b }[2], [x23], #0x1\n"
+ "st1 { v25.b }[2], [x22], #0x1\n"
+ "st1 { v20.b }[2], [x21], #0x1\n"
+ "b 34f\n"
+ "33:" // Oddments: Unroll 3: Oddment store: Bit 1: Unset
+ "st1 { v5.b }[0], [x24], #0x1\n"
+ "st1 { v30.b }[0], [x23], #0x1\n"
+ "st1 { v25.b }[0], [x22], #0x1\n"
+ "st1 { v20.b }[0], [x21], #0x1\n"
+ "34:" // Oddments: Unroll 3: Oddment store: Bit 1: End
+ "35:" // End
: [params] "+&r" (params)
- : [inptrs] "r" (inptrs), [n_channels] "r" ((long unsigned int) n_channels), [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", "x12", "x13", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+ : [inptrs] "r" (inptrs), [n_channels] "r" (n_channels), [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", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
);
}