aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp
diff options
context:
space:
mode:
authorramelg01 <ramy.elgammal@arm.com>2022-04-07 02:42:52 +0100
committerRamy Elgammal <ramy.elgammal@arm.com>2022-04-26 15:51:22 +0000
commit8a164884dddf769643cf3b9f7f94e43cb4f3c20b (patch)
tree35958dd48b6df1a851c880dad2b2ce285671b611 /src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp
parentc827e99fc46521f43719b0c2d1b6f05d66abf68c (diff)
downloadComputeLibrary-8a164884dddf769643cf3b9f7f94e43cb4f3c20b.tar.gz
Update Neon™ depthwise kernel
- Reduce duplication and simplify overall structure. - Improve multi-threaded performance by sharing more data in lower-level caches. Partially Resolves: COMPMID-5054 Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com> Change-Id: Iac747f39b21c540122fa75218762631c4d787911 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7449 Tested-by: Arm Jenkins <bsgcomp@arm.com> Reviewed-by: Andrew Mundy Reviewed-by: Sheri Zhang <sheri.zhang@arm.com> Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp1412
1 files changed, 693 insertions, 719 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp
index 8cbbfae00d..22f95746fc 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -46,7 +46,7 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
struct Params
{
long unsigned int n_channels;
- const int8_t *weights;
+ const void *weights;
const int32_t *bias;
const arm_gemm::Requantize32 *requant;
const int32_t *const requant_muls;
@@ -57,7 +57,7 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
Params(
long unsigned int n_channels,
const uint8_t *const *inptrs_raw,
- const int8_t *const weights,
+ const void *const weights,
const int32_t *const bias,
const arm_gemm::Requantize32 &qp,
const int32_t *const requant_muls,
@@ -91,505 +91,489 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
requant_muls, requant_shifts, outptrs);
__asm__ __volatile__(
+ "ldr x19, [%x[params], %[offsetof_Params_requant]]\n"
"ldr x8, [%x[params], %[offsetof_Params_n_channels]]\n"
- "mov x17, #0x0\n"
- "ldr x16, [%x[params], %[offsetof_Params_weights]]\n"
+ "add x24, x19, %[offsetof_Requantize32_a_offset]\n"
+ "add x23, x19, %[offsetof_Requantize32_b_offset]\n"
+ "ldr x22, [%x[params], %[offsetof_Params_outptrs]]\n"
+ "add x21, x19, %[offsetof_Requantize32_c_offset]\n"
+ "add x20, x19, %[offsetof_Requantize32_minval]\n"
+ "ldr x17, [%x[params], %[offsetof_Params_weights]]\n"
+ "add x19, x19, %[offsetof_Requantize32_maxval]\n"
+ "ld1r { v22.16b }, [x24]\n"
+ "ld1r { v12.16b }, [x23]\n"
+ "lsr x16, x8, #0x3\n"
+ "ld1r { v14.8h }, [x21]\n"
+ "ld1r { v17.8h }, [x20]\n"
"mov x15, #0x0\n"
- "ldr x22, [%x[params], %[offsetof_Params_requant]]\n"
- "add x14, %x[params], %[offsetof_Params_inptrs]\n"
+ "mov x14, #0x0\n"
+ "ld1r { v15.8h }, [x19]\n"
"ldr x13, [%x[params], %[offsetof_Params_requant_muls]]\n"
- "lsr x12, x8, #0x3\n"
+ "add x12, %x[params], %[offsetof_Params_inptrs]\n"
"ldr x11, [%x[params], %[offsetof_Params_requant_shifts]]\n"
- "add x19, x22, %[offsetof_Requantize32_a_offset]\n"
- "ldr x21, [%x[params], %[offsetof_Params_outptrs]]\n"
- "add x20, x22, %[offsetof_Requantize32_b_offset]\n"
- "ld1r { v21.16b }, [x19]\n"
- "add x19, x22, %[offsetof_Requantize32_c_offset]\n"
- "ld1r { v17.16b }, [x20]\n"
- "add x20, x22, %[offsetof_Requantize32_minval]\n"
- "ld1r { v13.4s }, [x19]\n"
- "add x19, x22, %[offsetof_Requantize32_maxval]\n"
- "ld1r { v15.4s }, [x20]\n"
- "ld1r { v14.4s }, [x19]\n"
- "ldp x10, x9, [x21, #0x0]\n"
- "ldp x28, x27, [x21, #0x10]\n"
- "cbz x12, 3f\n"
- "subs x12, x12, #0x1\n"
+ "ldp x10, x9, [x22, #0x0]\n"
+ "ldp x28, x27, [x22, #0x10]\n"
+ "cbz x16, 3f\n"
"ldr x19, [%x[params], %[offsetof_Params_bias]]\n"
- "ldr q11, [x19, #0x0]\n"
- "mov v23.16b, v11.16b\n"
+ "ldr q13, [x19, #0x0]\n"
+ "subs x16, x16, #0x1\n"
+ "mov v19.16b, v13.16b\n"
"ldr q26, [x19, #0x10]\n"
"add x19, x19, #0x20\n"
- "mov v12.16b, v11.16b\n"
"str x19, [%x[params], %[offsetof_Params_bias]]\n"
- "mov v24.16b, v11.16b\n"
- "ldr d0, [x16, #0x0]\n"
- "ldr d1, [x16, #0x8]\n"
- "mov v9.16b, v26.16b\n"
- "ldr d2, [x16, #0x10]\n"
- "mov v22.16b, v26.16b\n"
- "ldr d3, [x16, #0x18]\n"
- "mov v10.16b, v26.16b\n"
- "ldr d4, [x16, #0x20]\n"
- "ssubl v0.8h, v0.8b, v17.8b\n"
- "ldr d5, [x16, #0x28]\n"
- "ssubl v1.8h, v1.8b, v17.8b\n"
- "ldr d6, [x16, #0x30]\n"
- "ssubl v2.8h, v2.8b, v17.8b\n"
- "ldr d7, [x16, #0x38]\n"
- "ssubl v3.8h, v3.8b, v17.8b\n"
- "ldr d8, [x16, #0x40]\n"
- "ssubl v4.8h, v4.8b, v17.8b\n"
- "ldp x23, x22, [x14, #0x0]\n"
- "ssubl v5.8h, v5.8b, v17.8b\n"
- "ldp x21, x20, [x14, #0x10]\n"
- "ssubl v6.8h, v6.8b, v17.8b\n"
- "ssubl v7.8h, v7.8b, v17.8b\n"
- "ldr x19, [x14, #0x20]\n"
- "ssubl v8.8h, v8.8b, v17.8b\n"
- "ldr d31, [x23, x17]\n"
- "usubl v31.8h, v31.8b, v21.8b\n"
- "ldr d30, [x22, x17]\n"
- "ldr d29, [x21, x17]\n"
- "usubl v30.8h, v30.8b, v21.8b\n"
- "ldr d28, [x20, x17]\n"
- "usubl v29.8h, v29.8b, v21.8b\n"
- "ldr d27, [x19, x17]\n"
- "usubl v28.8h, v28.8b, v21.8b\n"
- "usubl v27.8h, v27.8b, v21.8b\n"
+ "ldr d0, [x17, #0x0]\n"
+ "ldr d1, [x17, #0x8]\n"
+ "ldr d2, [x17, #0x10]\n"
+ "mov v11.16b, v26.16b\n"
+ "mov v18.16b, v13.16b\n"
+ "ldr d3, [x17, #0x18]\n"
+ "ldr d4, [x17, #0x20]\n"
+ "mov v24.16b, v26.16b\n"
+ "mov v9.16b, v13.16b\n"
+ "ldr d5, [x17, #0x28]\n"
+ "ldr d6, [x17, #0x30]\n"
+ "mov v23.16b, v26.16b\n"
+ "ssubl v0.8h, v0.8b, v12.8b\n"
+ "ldr d7, [x17, #0x38]\n"
+ "ldr d8, [x17, #0x40]\n"
+ "ssubl v1.8h, v1.8b, v12.8b\n"
+ "ssubl v2.8h, v2.8b, v12.8b\n"
+ "ldp x23, x22, [x12, #0x0]\n"
+ "ldp x21, x20, [x12, #0x10]\n"
+ "ssubl v3.8h, v3.8b, v12.8b\n"
+ "ssubl v4.8h, v4.8b, v12.8b\n"
+ "ldr x19, [x12, #0x20]\n"
+ "ldr d31, [x23, x15]\n"
+ "ssubl v5.8h, v5.8b, v12.8b\n"
+ "ssubl v6.8h, v6.8b, v12.8b\n"
+ "ldr d30, [x22, x15]\n"
+ "ldr d29, [x21, x15]\n"
+ "ssubl v7.8h, v7.8b, v12.8b\n"
+ "ssubl v8.8h, v8.8b, v12.8b\n"
+ "ldr d28, [x20, x15]\n"
+ "ldr d27, [x19, x15]\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "usubl v30.8h, v30.8b, v22.8b\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "usubl v27.8h, v27.8b, v22.8b\n"
"beq 2f\n"
"1:" // Loop
- "smlal v11.4s, v31.4h, v4.4h\n"
- "ldr x21, [x14, #0x28]\n"
- "add x16, x16, #0x48\n"
+ "smlal v13.4s, v31.4h, v4.4h\n"
"smlal2 v26.4s, v31.8h, v4.8h\n"
- "ldr x20, [x14, #0x30]\n"
- "subs x12, x12, #0x1\n"
- "smlal v23.4s, v31.4h, v3.4h\n"
- "ldr x26, [x14, #0x38]\n"
- "smlal2 v9.4s, v31.8h, v3.8h\n"
- "ldr x25, [x14, #0x40]\n"
- "smlal v12.4s, v31.4h, v1.4h\n"
- "ldr x19, [x14, #0x48]\n"
- "smlal2 v22.4s, v31.8h, v1.8h\n"
- "ldr x24, [x14, #0x50]\n"
- "smlal v24.4s, v31.4h, v0.4h\n"
- "ldr x23, [x14, #0x58]\n"
- "smlal2 v10.4s, v31.8h, v0.8h\n"
- "ldr d31, [x21, x17]\n"
- "smlal v11.4s, v30.4h, v0.4h\n"
- "ldr x22, [x14, #0x60]\n"
+ "ldr x21, [x12, #0x28]\n"
+ "ldr x26, [x12, #0x38]\n"
+ "smlal v19.4s, v31.4h, v3.4h\n"
+ "smlal2 v11.4s, v31.8h, v3.8h\n"
+ "ldr x20, [x12, #0x30]\n"
+ "ldr x25, [x12, #0x40]\n"
+ "smlal v13.4s, v30.4h, v0.4h\n"
"smlal2 v26.4s, v30.8h, v0.8h\n"
- "ldr d30, [x19, x17]\n"
- "smlal v23.4s, v29.4h, v2.4h\n"
- "ldr x21, [x14, #0x68]\n"
- "smlal2 v9.4s, v29.8h, v2.8h\n"
- "ldr d29, [x20, x17]\n"
- "smlal v11.4s, v28.4h, v5.4h\n"
- "ldr x20, [x14, #0x70]\n"
+ "ldr x19, [x12, #0x48]\n"
+ "ldr d30, [x19, x15]\n"
+ "smlal v19.4s, v29.4h, v2.4h\n"
+ "smlal2 v11.4s, v29.8h, v2.8h\n"
+ "ldr d29, [x20, x15]\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "smlal v18.4s, v31.4h, v1.4h\n"
+ "smlal2 v24.4s, v31.8h, v1.8h\n"
+ "ldr x24, [x12, #0x50]\n"
+ "ldr x23, [x12, #0x58]\n"
+ "smlal v9.4s, v31.4h, v0.4h\n"
+ "smlal2 v23.4s, v31.8h, v0.8h\n"
+ "ldr d31, [x21, x15]\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "smlal v13.4s, v28.4h, v5.4h\n"
"smlal2 v26.4s, v28.8h, v5.8h\n"
- "ldr x19, [x14, #0x78]\n"
- "smlal v23.4s, v28.4h, v4.4h\n"
- "ldr q25, [x13, #0x0]\n"
- "smlal2 v9.4s, v28.8h, v4.8h\n"
- "ldr q18, [x11, #0x0]\n"
- "smlal v12.4s, v28.4h, v2.4h\n"
- "ldr q16, [x13, #0x10]\n"
- "add x13, x13, #0x20\n"
- "smlal2 v22.4s, v28.8h, v2.8h\n"
- "ldr q20, [x11, #0x10]\n"
- "add x11, x11, #0x20\n"
- "smlal v24.4s, v28.4h, v1.4h\n"
- "smlal2 v10.4s, v28.8h, v1.8h\n"
- "ldr d28, [x26, x17]\n"
- "usubl v31.8h, v31.8b, v21.8b\n"
- "smlal v11.4s, v27.4h, v7.4h\n"
+ "usubl v30.8h, v30.8b, v22.8b\n"
+ "ldr x22, [x12, #0x60]\n"
+ "smlal v19.4s, v28.4h, v4.4h\n"
+ "smlal2 v11.4s, v28.8h, v4.8h\n"
+ "ldr x21, [x12, #0x68]\n"
+ "ldr x20, [x12, #0x70]\n"
+ "smlal v18.4s, v28.4h, v2.4h\n"
+ "smlal2 v24.4s, v28.8h, v2.8h\n"
+ "ldr x19, [x12, #0x78]\n"
+ "ldr q21, [x13, #0x0]\n"
+ "smlal v9.4s, v28.4h, v1.4h\n"
+ "smlal2 v23.4s, v28.8h, v1.8h\n"
+ "ldr d28, [x26, x15]\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "smlal v13.4s, v27.4h, v7.4h\n"
"smlal2 v26.4s, v27.8h, v7.8h\n"
- "smlal v12.4s, v31.4h, v6.4h\n"
- "smlal2 v22.4s, v31.8h, v6.8h\n"
- "ldr d31, [x25, x17]\n"
- "smlal v23.4s, v27.4h, v6.4h\n"
- "smlal2 v9.4s, v27.8h, v6.8h\n"
- "smlal v12.4s, v27.4h, v4.4h\n"
- "smlal2 v22.4s, v27.8h, v4.8h\n"
- "smlal v24.4s, v27.4h, v3.4h\n"
- "smlal2 v10.4s, v27.8h, v3.8h\n"
- "usubl v29.8h, v29.8b, v21.8b\n"
- "usubl v28.8h, v28.8b, v21.8b\n"
- "usubl v31.8h, v31.8b, v21.8b\n"
- "smlal v24.4s, v29.4h, v8.4h\n"
- "smlal2 v10.4s, v29.8h, v8.8h\n"
- "ldr d29, [x24, x17]\n"
- "smlal v11.4s, v28.4h, v1.4h\n"
+ "ldr q25, [x11, #0x0]\n"
+ "ldr q10, [x13, #0x10]\n"
+ "smlal v19.4s, v27.4h, v6.4h\n"
+ "smlal2 v11.4s, v27.8h, v6.8h\n"
+ "ldr q16, [x11, #0x10]\n"
+ "add x17, x17, #0x48\n"
+ "smlal v18.4s, v31.4h, v6.4h\n"
+ "smlal2 v24.4s, v31.8h, v6.8h\n"
+ "ldr d31, [x25, x15]\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "smlal v9.4s, v27.4h, v3.4h\n"
+ "smlal2 v23.4s, v27.8h, v3.8h\n"
+ "subs x16, x16, #0x1\n"
+ "add x13, x13, #0x20\n"
+ "smlal v13.4s, v28.4h, v1.4h\n"
"smlal2 v26.4s, v28.8h, v1.8h\n"
- "smlal v23.4s, v28.4h, v0.4h\n"
- "smlal2 v9.4s, v28.8h, v0.8h\n"
- "ldr d28, [x23, x17]\n"
- "smlal v11.4s, v31.4h, v2.4h\n"
+ "add x11, x11, #0x20\n"
+ "smlal v19.4s, v28.4h, v0.4h\n"
+ "smlal2 v11.4s, v28.8h, v0.8h\n"
+ "ldr d28, [x23, x15]\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "smlal v18.4s, v27.4h, v4.4h\n"
+ "smlal v9.4s, v29.4h, v8.4h\n"
+ "smlal2 v24.4s, v27.8h, v4.8h\n"
+ "smlal2 v23.4s, v29.8h, v8.8h\n"
+ "ldr d29, [x24, x15]\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "smlal v13.4s, v31.4h, v2.4h\n"
"smlal2 v26.4s, v31.8h, v2.8h\n"
- "smlal v23.4s, v31.4h, v1.4h\n"
- "smlal2 v9.4s, v31.8h, v1.8h\n"
- "ldr d31, [x22, x17]\n"
- "usubl v30.8h, v30.8b, v21.8b\n"
- "usubl v29.8h, v29.8b, v21.8b\n"
- "usubl v28.8h, v28.8b, v21.8b\n"
- "smlal v11.4s, v30.4h, v8.4h\n"
+ "smlal v19.4s, v31.4h, v1.4h\n"
+ "smlal2 v11.4s, v31.8h, v1.8h\n"
+ "ldr d31, [x22, x15]\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "smlal v18.4s, v30.4h, v5.4h\n"
+ "smlal v9.4s, v30.4h, v4.4h\n"
+ "smlal v13.4s, v30.4h, v8.4h\n"
"smlal2 v26.4s, v30.8h, v8.8h\n"
- "smlal v23.4s, v30.4h, v7.4h\n"
- "smlal2 v9.4s, v30.8h, v7.8h\n"
- "smlal v12.4s, v30.4h, v5.4h\n"
- "smlal2 v22.4s, v30.8h, v5.8h\n"
- "smlal v24.4s, v30.4h, v4.4h\n"
- "smlal2 v10.4s, v30.8h, v4.8h\n"
- "ldr d30, [x21, x17]\n"
- "smlal v11.4s, v29.4h, v3.4h\n"
+ "smlal v19.4s, v30.4h, v7.4h\n"
+ "smlal2 v11.4s, v30.8h, v7.8h\n"
+ "smlal2 v24.4s, v30.8h, v5.8h\n"
+ "smlal2 v23.4s, v30.8h, v4.8h\n"
+ "ldr d30, [x21, x15]\n"
+ "usubl v30.8h, v30.8b, v22.8b\n"
+ "smlal v18.4s, v29.4h, v0.4h\n"
+ "smlal v9.4s, v28.4h, v2.4h\n"
+ "smlal v13.4s, v29.4h, v3.4h\n"
"smlal2 v26.4s, v29.8h, v3.8h\n"
- "smlal v12.4s, v29.4h, v0.4h\n"
- "smlal2 v22.4s, v29.8h, v0.8h\n"
- "ldr d29, [x20, x17]\n"
- "smlal v23.4s, v28.4h, v5.4h\n"
- "smlal2 v9.4s, v28.8h, v5.8h\n"
- "smlal v24.4s, v28.4h, v2.4h\n"
- "smlal2 v10.4s, v28.8h, v2.8h\n"
- "ldr d28, [x19, x17]\n"
- "add x17, x17, #0x8\n"
- "usubl v31.8h, v31.8b, v21.8b\n"
- "usubl v30.8h, v30.8b, v21.8b\n"
- "usubl v29.8h, v29.8b, v21.8b\n"
- "smlal v11.4s, v31.4h, v6.4h\n"
- "smlal2 v26.4s, v31.8h, v6.8h\n"
- "smlal v12.4s, v31.4h, v3.4h\n"
- "smlal2 v22.4s, v31.8h, v3.8h\n"
- "smlal v23.4s, v30.4h, v8.4h\n"
- "smlal2 v9.4s, v30.8h, v8.8h\n"
- "smlal v24.4s, v30.4h, v5.4h\n"
- "smlal2 v10.4s, v30.8h, v5.8h\n"
- "smlal v12.4s, v29.4h, v7.4h\n"
- "smlal2 v22.4s, v29.8h, v7.8h\n"
- "smlal v24.4s, v29.4h, v6.4h\n"
- "smlal2 v10.4s, v29.8h, v6.8h\n"
- "usubl v28.8h, v28.8b, v21.8b\n"
- "sqrdmulh v11.4s, v11.4s, v25.4s\n"
- "sqrdmulh v26.4s, v26.4s, v16.4s\n"
- "smlal v12.4s, v28.4h, v8.4h\n"
- "smlal2 v22.4s, v28.8h, v8.8h\n"
- "smlal v24.4s, v28.4h, v7.4h\n"
- "smlal2 v10.4s, v28.8h, v7.8h\n"
- "and v19.16b, v11.16b, v18.16b\n"
- "and v5.16b, v26.16b, v20.16b\n"
- "sqrdmulh v23.4s, v23.4s, v25.4s\n"
- "sshr v19.4s, v19.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sqrdmulh v9.4s, v9.4s, v16.4s\n"
- "sqadd v11.4s, v11.4s, v19.4s\n"
- "sqadd v26.4s, v26.4s, v5.4s\n"
- "and v28.16b, v23.16b, v18.16b\n"
- "and v8.16b, v9.16b, v20.16b\n"
- "srshl v11.4s, v11.4s, v18.4s\n"
- "srshl v26.4s, v26.4s, v20.4s\n"
- "sshr v28.4s, v28.4s, #0x1f\n"
- "sshr v8.4s, v8.4s, #0x1f\n"
- "add v11.4s, v11.4s, v13.4s\n"
- "add v26.4s, v26.4s, v13.4s\n"
- "sqadd v23.4s, v23.4s, v28.4s\n"
- "smin v11.4s, v11.4s, v14.4s\n"
- "smin v26.4s, v26.4s, v14.4s\n"
- "sqadd v9.4s, v9.4s, v8.4s\n"
- "smax v11.4s, v11.4s, v15.4s\n"
- "smax v26.4s, v26.4s, v15.4s\n"
- "srshl v23.4s, v23.4s, v18.4s\n"
- "srshl v9.4s, v9.4s, v20.4s\n"
- "uzp1 v11.16b, v11.16b, v26.16b\n"
- "sqrdmulh v12.4s, v12.4s, v25.4s\n"
- "uzp1 v11.16b, v11.16b, v11.16b\n"
- "str d11, [x10, x15]\n"
- "add v23.4s, v23.4s, v13.4s\n"
- "add v9.4s, v9.4s, v13.4s\n"
- "and v1.16b, v12.16b, v18.16b\n"
- "sqrdmulh v22.4s, v22.4s, v16.4s\n"
- "smin v23.4s, v23.4s, v14.4s\n"
- "smin v9.4s, v9.4s, v14.4s\n"
- "sshr v1.4s, v1.4s, #0x1f\n"
- "smax v23.4s, v23.4s, v15.4s\n"
- "smax v9.4s, v9.4s, v15.4s\n"
- "sqadd v12.4s, v12.4s, v1.4s\n"
- "and v0.16b, v22.16b, v20.16b\n"
- "uzp1 v23.16b, v23.16b, v9.16b\n"
- "sqrdmulh v24.4s, v24.4s, v25.4s\n"
- "uzp1 v23.16b, v23.16b, v23.16b\n"
- "str d23, [x9, x15]\n"
- "srshl v12.4s, v12.4s, v18.4s\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "and v26.16b, v24.16b, v18.16b\n"
- "sqrdmulh v10.4s, v10.4s, v16.4s\n"
- "sqadd v22.4s, v22.4s, v0.4s\n"
- "add v12.4s, v12.4s, v13.4s\n"
- "sshr v26.4s, v26.4s, #0x1f\n"
- "and v16.16b, v10.16b, v20.16b\n"
- "smin v12.4s, v12.4s, v14.4s\n"
- "srshl v22.4s, v22.4s, v20.4s\n"
- "sqadd v24.4s, v24.4s, v26.4s\n"
- "smax v12.4s, v12.4s, v15.4s\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "add v22.4s, v22.4s, v13.4s\n"
- "srshl v24.4s, v24.4s, v18.4s\n"
- "sqadd v10.4s, v10.4s, v16.4s\n"
- "smin v22.4s, v22.4s, v14.4s\n"
- "add v24.4s, v24.4s, v13.4s\n"
- "smax v22.4s, v22.4s, v15.4s\n"
- "srshl v10.4s, v10.4s, v20.4s\n"
- "smin v24.4s, v24.4s, v14.4s\n"
- "uzp1 v12.16b, v12.16b, v22.16b\n"
- "add v10.4s, v10.4s, v13.4s\n"
- "uzp1 v12.16b, v12.16b, v12.16b\n"
- "str d12, [x28, x15]\n"
- "smax v24.4s, v24.4s, v15.4s\n"
- "smin v10.4s, v10.4s, v14.4s\n"
- "smax v10.4s, v10.4s, v15.4s\n"
- "uzp1 v24.16b, v24.16b, v10.16b\n"
- "uzp1 v24.16b, v24.16b, v24.16b\n"
- "str d24, [x27, x15]\n"
+ "smlal2 v24.4s, v29.8h, v0.8h\n"
+ "ldr d29, [x20, x15]\n"
+ "smlal2 v23.4s, v28.8h, v2.8h\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "smlal v18.4s, v31.4h, v3.4h\n"
+ "smlal v9.4s, v30.4h, v5.4h\n"
+ "smlal v19.4s, v28.4h, v5.4h\n"
+ "smlal2 v11.4s, v28.8h, v5.8h\n"
+ "ldr d28, [x19, x15]\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "smlal2 v24.4s, v31.8h, v3.8h\n"
+ "smlal2 v23.4s, v30.8h, v5.8h\n"
"add x15, x15, #0x8\n"
+ "smlal v18.4s, v29.4h, v7.4h\n"
+ "smlal v9.4s, v29.4h, v6.4h\n"
+ "smlal2 v24.4s, v29.8h, v7.8h\n"
+ "smlal2 v23.4s, v29.8h, v6.8h\n"
+ "smlal v13.4s, v31.4h, v6.4h\n"
+ "smlal v19.4s, v30.4h, v8.4h\n"
+ "sqdmulh v13.4s, v13.4s, v21.4s\n"
+ "smlal v18.4s, v28.4h, v8.4h\n"
+ "smlal v9.4s, v28.4h, v7.4h\n"
+ "sqdmulh v19.4s, v19.4s, v21.4s\n"
+ "smlal2 v26.4s, v31.8h, v6.8h\n"
+ "smlal2 v11.4s, v30.8h, v8.8h\n"
+ "sqdmulh v18.4s, v18.4s, v21.4s\n"
+ "smlal2 v24.4s, v28.8h, v8.8h\n"
+ "smlal2 v23.4s, v28.8h, v7.8h\n"
+ "sqdmulh v9.4s, v9.4s, v21.4s\n"
+ "and v7.16b, v13.16b, v25.16b\n"
+ "sqdmulh v26.4s, v26.4s, v10.4s\n"
+ "and v4.16b, v19.16b, v25.16b\n"
+ "sqdmulh v11.4s, v11.4s, v10.4s\n"
+ "and v21.16b, v18.16b, v25.16b\n"
+ "sqdmulh v24.4s, v24.4s, v10.4s\n"
+ "and v20.16b, v9.16b, v25.16b\n"
+ "sqdmulh v23.4s, v23.4s, v10.4s\n"
+ "sshr v7.4s, v7.4s, #0x1f\n"
+ "and v29.16b, v26.16b, v16.16b\n"
+ "sshr v4.4s, v4.4s, #0x1f\n"
+ "and v10.16b, v11.16b, v16.16b\n"
+ "sshr v21.4s, v21.4s, #0x1f\n"
+ "and v31.16b, v24.16b, v16.16b\n"
+ "sshr v20.4s, v20.4s, #0x1f\n"
+ "and v30.16b, v23.16b, v16.16b\n"
+ "sqadd v13.4s, v13.4s, v7.4s\n"
+ "sshr v29.4s, v29.4s, #0x1f\n"
+ "sqadd v19.4s, v19.4s, v4.4s\n"
+ "sshr v10.4s, v10.4s, #0x1f\n"
+ "sqadd v18.4s, v18.4s, v21.4s\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sqadd v9.4s, v9.4s, v20.4s\n"
+ "sshr v30.4s, v30.4s, #0x1f\n"
+ "srshl v13.4s, v13.4s, v25.4s\n"
+ "sqadd v26.4s, v26.4s, v29.4s\n"
+ "srshl v19.4s, v19.4s, v25.4s\n"
+ "sqadd v11.4s, v11.4s, v10.4s\n"
+ "srshl v18.4s, v18.4s, v25.4s\n"
+ "sqadd v24.4s, v24.4s, v31.4s\n"
+ "srshl v9.4s, v9.4s, v25.4s\n"
+ "sqadd v23.4s, v23.4s, v30.4s\n"
+ "srshl v26.4s, v26.4s, v16.4s\n"
+ "sqxtn v13.4h, v13.4s\n"
+ "srshl v11.4s, v11.4s, v16.4s\n"
+ "sqxtn v19.4h, v19.4s\n"
+ "srshl v24.4s, v24.4s, v16.4s\n"
+ "sqxtn v18.4h, v18.4s\n"
+ "srshl v23.4s, v23.4s, v16.4s\n"
+ "sqxtn v9.4h, v9.4s\n"
+ "sqxtn2 v13.8h, v26.4s\n"
+ "sqxtn2 v19.8h, v11.4s\n"
+ "sqxtn2 v18.8h, v24.4s\n"
+ "sqxtn2 v9.8h, v23.4s\n"
+ "sqadd v13.8h, v13.8h, v14.8h\n"
+ "sqadd v19.8h, v19.8h, v14.8h\n"
+ "sqadd v18.8h, v18.8h, v14.8h\n"
+ "sqadd v9.8h, v9.8h, v14.8h\n"
+ "smax v13.8h, v13.8h, v17.8h\n"
+ "smax v19.8h, v19.8h, v17.8h\n"
+ "smax v18.8h, v18.8h, v17.8h\n"
+ "smax v9.8h, v9.8h, v17.8h\n"
+ "smin v13.8h, v13.8h, v15.8h\n"
+ "smin v19.8h, v19.8h, v15.8h\n"
+ "smin v18.8h, v18.8h, v15.8h\n"
+ "smin v9.8h, v9.8h, v15.8h\n"
+ "uzp1 v13.16b, v13.16b, v13.16b\n"
+ "uzp1 v19.16b, v19.16b, v19.16b\n"
+ "str d13, [x10, x14]\n"
+ "uzp1 v18.16b, v18.16b, v18.16b\n"
+ "uzp1 v9.16b, v9.16b, v9.16b\n"
+ "str d19, [x9, x14]\n"
+ "str d18, [x28, x14]\n"
+ "str d9, [x27, x14]\n"
"ldr x19, [%x[params], %[offsetof_Params_bias]]\n"
- "ldr q11, [x19, #0x0]\n"
- "mov v23.16b, v11.16b\n"
+ "ldr q13, [x19, #0x0]\n"
+ "add x14, x14, #0x8\n"
"ldr q26, [x19, #0x10]\n"
"add x19, x19, #0x20\n"
- "mov v12.16b, v11.16b\n"
"str x19, [%x[params], %[offsetof_Params_bias]]\n"
- "mov v24.16b, v11.16b\n"
- "ldr d0, [x16, #0x0]\n"
- "ldr d1, [x16, #0x8]\n"
- "mov v9.16b, v26.16b\n"
- "ldr d2, [x16, #0x10]\n"
- "mov v22.16b, v26.16b\n"
- "ldr d3, [x16, #0x18]\n"
- "mov v10.16b, v26.16b\n"
- "ldr d4, [x16, #0x20]\n"
- "ssubl v0.8h, v0.8b, v17.8b\n"
- "ldr d5, [x16, #0x28]\n"
- "ssubl v1.8h, v1.8b, v17.8b\n"
- "ldr d6, [x16, #0x30]\n"
- "ssubl v2.8h, v2.8b, v17.8b\n"
- "ldr d7, [x16, #0x38]\n"
- "ssubl v3.8h, v3.8b, v17.8b\n"
- "ldr d8, [x16, #0x40]\n"
- "ssubl v4.8h, v4.8b, v17.8b\n"
- "ldp x23, x22, [x14, #0x0]\n"
- "ssubl v5.8h, v5.8b, v17.8b\n"
- "ldp x21, x20, [x14, #0x10]\n"
- "ssubl v6.8h, v6.8b, v17.8b\n"
- "ssubl v7.8h, v7.8b, v17.8b\n"
- "ldr x19, [x14, #0x20]\n"
- "ssubl v8.8h, v8.8b, v17.8b\n"
- "ldr d31, [x23, x17]\n"
- "usubl v31.8h, v31.8b, v21.8b\n"
- "ldr d30, [x22, x17]\n"
- "ldr d29, [x21, x17]\n"
- "usubl v30.8h, v30.8b, v21.8b\n"
- "ldr d28, [x20, x17]\n"
- "usubl v29.8h, v29.8b, v21.8b\n"
- "ldr d27, [x19, x17]\n"
- "usubl v28.8h, v28.8b, v21.8b\n"
- "usubl v27.8h, v27.8b, v21.8b\n"
+ "ldr d0, [x17, #0x0]\n"
+ "ldr d1, [x17, #0x8]\n"
+ "ldr d2, [x17, #0x10]\n"
+ "mov v19.16b, v13.16b\n"
+ "mov v11.16b, v26.16b\n"
+ "ldr d3, [x17, #0x18]\n"
+ "ldr d4, [x17, #0x20]\n"
+ "mov v18.16b, v13.16b\n"
+ "mov v24.16b, v26.16b\n"
+ "ldr d5, [x17, #0x28]\n"
+ "ldr d6, [x17, #0x30]\n"
+ "mov v9.16b, v13.16b\n"
+ "mov v23.16b, v26.16b\n"
+ "ldr d7, [x17, #0x38]\n"
+ "ldr d8, [x17, #0x40]\n"
+ "ssubl v0.8h, v0.8b, v12.8b\n"
+ "ssubl v1.8h, v1.8b, v12.8b\n"
+ "ldp x23, x22, [x12, #0x0]\n"
+ "ldp x21, x20, [x12, #0x10]\n"
+ "ssubl v2.8h, v2.8b, v12.8b\n"
+ "ssubl v3.8h, v3.8b, v12.8b\n"
+ "ldr x19, [x12, #0x20]\n"
+ "ldr d31, [x23, x15]\n"
+ "ssubl v4.8h, v4.8b, v12.8b\n"
+ "ssubl v5.8h, v5.8b, v12.8b\n"
+ "ldr d30, [x22, x15]\n"
+ "ldr d29, [x21, x15]\n"
+ "ssubl v6.8h, v6.8b, v12.8b\n"
+ "ssubl v7.8h, v7.8b, v12.8b\n"
+ "ldr d28, [x20, x15]\n"
+ "ldr d27, [x19, x15]\n"
+ "ssubl v8.8h, v8.8b, v12.8b\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "usubl v30.8h, v30.8b, v22.8b\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "usubl v27.8h, v27.8b, v22.8b\n"
"bgt 1b\n"
"2:" // Tail
- "smlal v11.4s, v31.4h, v4.4h\n"
- "ldr x21, [x14, #0x28]\n"
- "tst x8, #0x7\n"
+ "smlal v13.4s, v31.4h, v4.4h\n"
"smlal2 v26.4s, v31.8h, v4.8h\n"
- "ldr x20, [x14, #0x30]\n"
- "smlal v23.4s, v31.4h, v3.4h\n"
- "ldr x26, [x14, #0x38]\n"
- "smlal2 v9.4s, v31.8h, v3.8h\n"
- "ldr x25, [x14, #0x40]\n"
- "smlal v12.4s, v31.4h, v1.4h\n"
- "ldr x19, [x14, #0x48]\n"
- "smlal2 v22.4s, v31.8h, v1.8h\n"
- "ldr x24, [x14, #0x50]\n"
- "smlal v24.4s, v31.4h, v0.4h\n"
- "ldr x23, [x14, #0x58]\n"
- "smlal2 v10.4s, v31.8h, v0.8h\n"
- "ldr d31, [x21, x17]\n"
- "smlal v11.4s, v30.4h, v0.4h\n"
- "ldr x22, [x14, #0x60]\n"
+ "ldr x21, [x12, #0x28]\n"
+ "ldr x26, [x12, #0x38]\n"
+ "smlal v19.4s, v31.4h, v3.4h\n"
+ "smlal2 v11.4s, v31.8h, v3.8h\n"
+ "ldr x20, [x12, #0x30]\n"
+ "ldr x25, [x12, #0x40]\n"
+ "smlal v13.4s, v30.4h, v0.4h\n"
"smlal2 v26.4s, v30.8h, v0.8h\n"
- "ldr d30, [x19, x17]\n"
- "smlal v23.4s, v29.4h, v2.4h\n"
- "ldr x21, [x14, #0x68]\n"
- "smlal2 v9.4s, v29.8h, v2.8h\n"
- "ldr d29, [x20, x17]\n"
- "smlal v11.4s, v28.4h, v5.4h\n"
- "ldr x20, [x14, #0x70]\n"
+ "ldr x19, [x12, #0x48]\n"
+ "ldr d30, [x19, x15]\n"
+ "smlal v19.4s, v29.4h, v2.4h\n"
+ "smlal2 v11.4s, v29.8h, v2.8h\n"
+ "ldr d29, [x20, x15]\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "smlal v18.4s, v31.4h, v1.4h\n"
+ "smlal2 v24.4s, v31.8h, v1.8h\n"
+ "ldr x24, [x12, #0x50]\n"
+ "ldr x23, [x12, #0x58]\n"
+ "smlal v9.4s, v31.4h, v0.4h\n"
+ "smlal2 v23.4s, v31.8h, v0.8h\n"
+ "ldr d31, [x21, x15]\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "smlal v13.4s, v28.4h, v5.4h\n"
"smlal2 v26.4s, v28.8h, v5.8h\n"
- "ldr x19, [x14, #0x78]\n"
- "smlal v23.4s, v28.4h, v4.4h\n"
- "ldr q25, [x13, #0x0]\n"
- "smlal2 v9.4s, v28.8h, v4.8h\n"
- "ldr q18, [x11, #0x0]\n"
- "smlal v12.4s, v28.4h, v2.4h\n"
- "ldr q16, [x13, #0x10]\n"
+ "usubl v30.8h, v30.8b, v22.8b\n"
+ "ldr x22, [x12, #0x60]\n"
+ "smlal v19.4s, v28.4h, v4.4h\n"
+ "smlal2 v11.4s, v28.8h, v4.8h\n"
+ "ldr x21, [x12, #0x68]\n"
+ "ldr x20, [x12, #0x70]\n"
+ "smlal v18.4s, v28.4h, v2.4h\n"
+ "smlal2 v24.4s, v28.8h, v2.8h\n"
+ "ldr x19, [x12, #0x78]\n"
+ "ldr q21, [x13, #0x0]\n"
+ "smlal v9.4s, v28.4h, v1.4h\n"
+ "smlal2 v23.4s, v28.8h, v1.8h\n"
+ "ldr d28, [x26, x15]\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "smlal v13.4s, v27.4h, v7.4h\n"
+ "smlal2 v26.4s, v27.8h, v7.8h\n"
+ "ldr q25, [x11, #0x0]\n"
+ "ldr q10, [x13, #0x10]\n"
+ "smlal v19.4s, v27.4h, v6.4h\n"
+ "smlal2 v11.4s, v27.8h, v6.8h\n"
+ "ldr q16, [x11, #0x10]\n"
+ "tst x8, #0x7\n"
+ "smlal v18.4s, v31.4h, v6.4h\n"
+ "smlal2 v24.4s, v31.8h, v6.8h\n"
+ "ldr d31, [x25, x15]\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "smlal v9.4s, v27.4h, v3.4h\n"
+ "smlal2 v23.4s, v27.8h, v3.8h\n"
"add x13, x13, #0x20\n"
- "smlal2 v22.4s, v28.8h, v2.8h\n"
- "ldr q20, [x11, #0x10]\n"
"add x11, x11, #0x20\n"
- "smlal v24.4s, v28.4h, v1.4h\n"
- "smlal2 v10.4s, v28.8h, v1.8h\n"
- "ldr d28, [x26, x17]\n"
- "usubl v31.8h, v31.8b, v21.8b\n"
- "smlal v11.4s, v27.4h, v7.4h\n"
- "smlal2 v26.4s, v27.8h, v7.8h\n"
- "smlal v12.4s, v31.4h, v6.4h\n"
- "smlal2 v22.4s, v31.8h, v6.8h\n"
- "ldr d31, [x25, x17]\n"
- "smlal v23.4s, v27.4h, v6.4h\n"
- "smlal2 v9.4s, v27.8h, v6.8h\n"
- "smlal v12.4s, v27.4h, v4.4h\n"
- "smlal2 v22.4s, v27.8h, v4.8h\n"
- "smlal v24.4s, v27.4h, v3.4h\n"
- "smlal2 v10.4s, v27.8h, v3.8h\n"
- "usubl v29.8h, v29.8b, v21.8b\n"
- "usubl v28.8h, v28.8b, v21.8b\n"
- "usubl v31.8h, v31.8b, v21.8b\n"
- "smlal v24.4s, v29.4h, v8.4h\n"
- "smlal2 v10.4s, v29.8h, v8.8h\n"
- "ldr d29, [x24, x17]\n"
- "smlal v11.4s, v28.4h, v1.4h\n"
+ "smlal v13.4s, v28.4h, v1.4h\n"
"smlal2 v26.4s, v28.8h, v1.8h\n"
- "smlal v23.4s, v28.4h, v0.4h\n"
- "smlal2 v9.4s, v28.8h, v0.8h\n"
- "ldr d28, [x23, x17]\n"
- "smlal v11.4s, v31.4h, v2.4h\n"
+ "smlal v19.4s, v28.4h, v0.4h\n"
+ "smlal2 v11.4s, v28.8h, v0.8h\n"
+ "ldr d28, [x23, x15]\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "smlal v18.4s, v27.4h, v4.4h\n"
+ "smlal v9.4s, v29.4h, v8.4h\n"
+ "smlal2 v24.4s, v27.8h, v4.8h\n"
+ "smlal2 v23.4s, v29.8h, v8.8h\n"
+ "ldr d29, [x24, x15]\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "smlal v13.4s, v31.4h, v2.4h\n"
"smlal2 v26.4s, v31.8h, v2.8h\n"
- "smlal v23.4s, v31.4h, v1.4h\n"
- "smlal2 v9.4s, v31.8h, v1.8h\n"
- "ldr d31, [x22, x17]\n"
- "usubl v30.8h, v30.8b, v21.8b\n"
- "usubl v29.8h, v29.8b, v21.8b\n"
- "usubl v28.8h, v28.8b, v21.8b\n"
- "smlal v11.4s, v30.4h, v8.4h\n"
+ "smlal v19.4s, v31.4h, v1.4h\n"
+ "smlal2 v11.4s, v31.8h, v1.8h\n"
+ "ldr d31, [x22, x15]\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "smlal v18.4s, v30.4h, v5.4h\n"
+ "smlal v9.4s, v30.4h, v4.4h\n"
+ "smlal v13.4s, v30.4h, v8.4h\n"
"smlal2 v26.4s, v30.8h, v8.8h\n"
- "smlal v23.4s, v30.4h, v7.4h\n"
- "smlal2 v9.4s, v30.8h, v7.8h\n"
- "smlal v12.4s, v30.4h, v5.4h\n"
- "smlal2 v22.4s, v30.8h, v5.8h\n"
- "smlal v24.4s, v30.4h, v4.4h\n"
- "smlal2 v10.4s, v30.8h, v4.8h\n"
- "ldr d30, [x21, x17]\n"
- "smlal v11.4s, v29.4h, v3.4h\n"
+ "smlal v19.4s, v30.4h, v7.4h\n"
+ "smlal2 v11.4s, v30.8h, v7.8h\n"
+ "smlal2 v24.4s, v30.8h, v5.8h\n"
+ "smlal2 v23.4s, v30.8h, v4.8h\n"
+ "ldr d30, [x21, x15]\n"
+ "usubl v30.8h, v30.8b, v22.8b\n"
+ "smlal v18.4s, v29.4h, v0.4h\n"
+ "smlal v9.4s, v28.4h, v2.4h\n"
+ "smlal v13.4s, v29.4h, v3.4h\n"
"smlal2 v26.4s, v29.8h, v3.8h\n"
- "smlal v12.4s, v29.4h, v0.4h\n"
- "smlal2 v22.4s, v29.8h, v0.8h\n"
- "ldr d29, [x20, x17]\n"
- "smlal v23.4s, v28.4h, v5.4h\n"
- "smlal2 v9.4s, v28.8h, v5.8h\n"
- "smlal v24.4s, v28.4h, v2.4h\n"
- "smlal2 v10.4s, v28.8h, v2.8h\n"
- "ldr d28, [x19, x17]\n"
- "add x17, x17, #0x8\n"
- "usubl v31.8h, v31.8b, v21.8b\n"
- "usubl v30.8h, v30.8b, v21.8b\n"
- "usubl v29.8h, v29.8b, v21.8b\n"
- "smlal v11.4s, v31.4h, v6.4h\n"
- "smlal2 v26.4s, v31.8h, v6.8h\n"
- "smlal v12.4s, v31.4h, v3.4h\n"
- "smlal2 v22.4s, v31.8h, v3.8h\n"
- "smlal v23.4s, v30.4h, v8.4h\n"
- "smlal2 v9.4s, v30.8h, v8.8h\n"
- "smlal v24.4s, v30.4h, v5.4h\n"
- "smlal2 v10.4s, v30.8h, v5.8h\n"
- "smlal v12.4s, v29.4h, v7.4h\n"
- "smlal2 v22.4s, v29.8h, v7.8h\n"
- "smlal v24.4s, v29.4h, v6.4h\n"
- "smlal2 v10.4s, v29.8h, v6.8h\n"
- "usubl v28.8h, v28.8b, v21.8b\n"
- "sqrdmulh v11.4s, v11.4s, v25.4s\n"
- "sqrdmulh v26.4s, v26.4s, v16.4s\n"
- "smlal v12.4s, v28.4h, v8.4h\n"
- "smlal2 v22.4s, v28.8h, v8.8h\n"
- "smlal v24.4s, v28.4h, v7.4h\n"
- "smlal2 v10.4s, v28.8h, v7.8h\n"
- "and v19.16b, v11.16b, v18.16b\n"
- "and v5.16b, v26.16b, v20.16b\n"
- "sqrdmulh v23.4s, v23.4s, v25.4s\n"
- "sshr v19.4s, v19.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sqrdmulh v9.4s, v9.4s, v16.4s\n"
- "sqadd v11.4s, v11.4s, v19.4s\n"
- "sqadd v26.4s, v26.4s, v5.4s\n"
- "and v28.16b, v23.16b, v18.16b\n"
- "and v8.16b, v9.16b, v20.16b\n"
- "srshl v11.4s, v11.4s, v18.4s\n"
- "srshl v26.4s, v26.4s, v20.4s\n"
- "sshr v28.4s, v28.4s, #0x1f\n"
- "sshr v8.4s, v8.4s, #0x1f\n"
- "add v11.4s, v11.4s, v13.4s\n"
- "add v26.4s, v26.4s, v13.4s\n"
- "sqadd v23.4s, v23.4s, v28.4s\n"
- "smin v11.4s, v11.4s, v14.4s\n"
- "smin v26.4s, v26.4s, v14.4s\n"
- "sqadd v9.4s, v9.4s, v8.4s\n"
- "smax v11.4s, v11.4s, v15.4s\n"
- "smax v26.4s, v26.4s, v15.4s\n"
- "srshl v23.4s, v23.4s, v18.4s\n"
- "srshl v9.4s, v9.4s, v20.4s\n"
- "uzp1 v11.16b, v11.16b, v26.16b\n"
- "sqrdmulh v12.4s, v12.4s, v25.4s\n"
- "uzp1 v11.16b, v11.16b, v11.16b\n"
- "str d11, [x10, x15]\n"
- "add v23.4s, v23.4s, v13.4s\n"
- "add v9.4s, v9.4s, v13.4s\n"
- "and v1.16b, v12.16b, v18.16b\n"
- "sqrdmulh v22.4s, v22.4s, v16.4s\n"
- "smin v23.4s, v23.4s, v14.4s\n"
- "smin v9.4s, v9.4s, v14.4s\n"
- "sshr v1.4s, v1.4s, #0x1f\n"
- "smax v23.4s, v23.4s, v15.4s\n"
- "smax v9.4s, v9.4s, v15.4s\n"
- "sqadd v12.4s, v12.4s, v1.4s\n"
- "and v0.16b, v22.16b, v20.16b\n"
- "uzp1 v23.16b, v23.16b, v9.16b\n"
- "sqrdmulh v24.4s, v24.4s, v25.4s\n"
- "uzp1 v23.16b, v23.16b, v23.16b\n"
- "str d23, [x9, x15]\n"
- "srshl v12.4s, v12.4s, v18.4s\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "and v26.16b, v24.16b, v18.16b\n"
- "sqrdmulh v10.4s, v10.4s, v16.4s\n"
- "sqadd v22.4s, v22.4s, v0.4s\n"
- "add v12.4s, v12.4s, v13.4s\n"
- "sshr v26.4s, v26.4s, #0x1f\n"
- "and v16.16b, v10.16b, v20.16b\n"
- "smin v12.4s, v12.4s, v14.4s\n"
- "srshl v22.4s, v22.4s, v20.4s\n"
- "sqadd v24.4s, v24.4s, v26.4s\n"
- "smax v12.4s, v12.4s, v15.4s\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "add v22.4s, v22.4s, v13.4s\n"
- "srshl v24.4s, v24.4s, v18.4s\n"
- "sqadd v10.4s, v10.4s, v16.4s\n"
- "smin v22.4s, v22.4s, v14.4s\n"
- "add v24.4s, v24.4s, v13.4s\n"
- "smax v22.4s, v22.4s, v15.4s\n"
- "srshl v10.4s, v10.4s, v20.4s\n"
- "smin v24.4s, v24.4s, v14.4s\n"
- "uzp1 v12.16b, v12.16b, v22.16b\n"
- "add v10.4s, v10.4s, v13.4s\n"
- "uzp1 v12.16b, v12.16b, v12.16b\n"
- "str d12, [x28, x15]\n"
- "smax v24.4s, v24.4s, v15.4s\n"
- "smin v10.4s, v10.4s, v14.4s\n"
- "smax v10.4s, v10.4s, v15.4s\n"
- "uzp1 v24.16b, v24.16b, v10.16b\n"
- "uzp1 v24.16b, v24.16b, v24.16b\n"
- "str d24, [x27, x15]\n"
+ "smlal2 v24.4s, v29.8h, v0.8h\n"
+ "ldr d29, [x20, x15]\n"
+ "smlal2 v23.4s, v28.8h, v2.8h\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "smlal v18.4s, v31.4h, v3.4h\n"
+ "smlal v9.4s, v30.4h, v5.4h\n"
+ "smlal v19.4s, v28.4h, v5.4h\n"
+ "smlal2 v11.4s, v28.8h, v5.8h\n"
+ "ldr d28, [x19, x15]\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "smlal2 v24.4s, v31.8h, v3.8h\n"
+ "smlal2 v23.4s, v30.8h, v5.8h\n"
"add x15, x15, #0x8\n"
+ "smlal v18.4s, v29.4h, v7.4h\n"
+ "smlal v9.4s, v29.4h, v6.4h\n"
+ "smlal2 v24.4s, v29.8h, v7.8h\n"
+ "smlal2 v23.4s, v29.8h, v6.8h\n"
+ "smlal v13.4s, v31.4h, v6.4h\n"
+ "smlal v19.4s, v30.4h, v8.4h\n"
+ "sqdmulh v13.4s, v13.4s, v21.4s\n"
+ "smlal v18.4s, v28.4h, v8.4h\n"
+ "smlal v9.4s, v28.4h, v7.4h\n"
+ "sqdmulh v19.4s, v19.4s, v21.4s\n"
+ "smlal2 v26.4s, v31.8h, v6.8h\n"
+ "smlal2 v11.4s, v30.8h, v8.8h\n"
+ "sqdmulh v18.4s, v18.4s, v21.4s\n"
+ "smlal2 v24.4s, v28.8h, v8.8h\n"
+ "smlal2 v23.4s, v28.8h, v7.8h\n"
+ "sqdmulh v9.4s, v9.4s, v21.4s\n"
+ "and v7.16b, v13.16b, v25.16b\n"
+ "sqdmulh v26.4s, v26.4s, v10.4s\n"
+ "and v4.16b, v19.16b, v25.16b\n"
+ "sqdmulh v11.4s, v11.4s, v10.4s\n"
+ "and v21.16b, v18.16b, v25.16b\n"
+ "sqdmulh v24.4s, v24.4s, v10.4s\n"
+ "and v20.16b, v9.16b, v25.16b\n"
+ "sqdmulh v23.4s, v23.4s, v10.4s\n"
+ "sshr v7.4s, v7.4s, #0x1f\n"
+ "and v29.16b, v26.16b, v16.16b\n"
+ "sshr v4.4s, v4.4s, #0x1f\n"
+ "and v10.16b, v11.16b, v16.16b\n"
+ "sshr v21.4s, v21.4s, #0x1f\n"
+ "and v31.16b, v24.16b, v16.16b\n"
+ "sshr v20.4s, v20.4s, #0x1f\n"
+ "and v30.16b, v23.16b, v16.16b\n"
+ "sqadd v13.4s, v13.4s, v7.4s\n"
+ "sshr v29.4s, v29.4s, #0x1f\n"
+ "sqadd v19.4s, v19.4s, v4.4s\n"
+ "sshr v10.4s, v10.4s, #0x1f\n"
+ "sqadd v18.4s, v18.4s, v21.4s\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sqadd v9.4s, v9.4s, v20.4s\n"
+ "sshr v30.4s, v30.4s, #0x1f\n"
+ "srshl v13.4s, v13.4s, v25.4s\n"
+ "sqadd v26.4s, v26.4s, v29.4s\n"
+ "srshl v19.4s, v19.4s, v25.4s\n"
+ "sqadd v11.4s, v11.4s, v10.4s\n"
+ "srshl v18.4s, v18.4s, v25.4s\n"
+ "sqadd v24.4s, v24.4s, v31.4s\n"
+ "srshl v9.4s, v9.4s, v25.4s\n"
+ "sqadd v23.4s, v23.4s, v30.4s\n"
+ "srshl v26.4s, v26.4s, v16.4s\n"
+ "sqxtn v13.4h, v13.4s\n"
+ "srshl v11.4s, v11.4s, v16.4s\n"
+ "sqxtn v19.4h, v19.4s\n"
+ "srshl v24.4s, v24.4s, v16.4s\n"
+ "sqxtn v18.4h, v18.4s\n"
+ "srshl v23.4s, v23.4s, v16.4s\n"
+ "sqxtn v9.4h, v9.4s\n"
+ "sqxtn2 v13.8h, v26.4s\n"
+ "sqxtn2 v19.8h, v11.4s\n"
+ "sqxtn2 v18.8h, v24.4s\n"
+ "sqxtn2 v9.8h, v23.4s\n"
+ "sqadd v13.8h, v13.8h, v14.8h\n"
+ "sqadd v19.8h, v19.8h, v14.8h\n"
+ "sqadd v18.8h, v18.8h, v14.8h\n"
+ "sqadd v9.8h, v9.8h, v14.8h\n"
+ "smax v13.8h, v13.8h, v17.8h\n"
+ "smax v19.8h, v19.8h, v17.8h\n"
+ "smax v18.8h, v18.8h, v17.8h\n"
+ "smax v9.8h, v9.8h, v17.8h\n"
+ "smin v13.8h, v13.8h, v15.8h\n"
+ "smin v19.8h, v19.8h, v15.8h\n"
+ "smin v18.8h, v18.8h, v15.8h\n"
+ "smin v9.8h, v9.8h, v15.8h\n"
+ "uzp1 v13.16b, v13.16b, v13.16b\n"
+ "uzp1 v19.16b, v19.16b, v19.16b\n"
+ "str d13, [x10, x14]\n"
+ "uzp1 v18.16b, v18.16b, v18.16b\n"
+ "uzp1 v9.16b, v9.16b, v9.16b\n"
+ "str d19, [x9, x14]\n"
+ "str d18, [x28, x14]\n"
+ "str d9, [x27, x14]\n"
+ "add x14, x14, #0x8\n"
"beq 64f\n"
- "add x16, x16, #0x48\n"
+ "add x17, x17, #0x48\n"
"3:" // Oddments
"ldr x19, [%x[params], %[offsetof_Params_bias]]\n"
"tbz x8, #2, 5f\n"
- "ld1 { v11.4s }, [x19], #0x10\n"
+ "ld1 { v13.4s }, [x19], #0x10\n"
"tbz x8, #1, 4f\n"
"ld1 { v26.d }[0], [x19], #0x8\n"
"tbz x8, #0, 7f\n"
@@ -601,46 +585,46 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"b 7f\n"
"5:" // Oddments: Load bias: Bit 2: Unset
"tbz x8, #1, 6f\n"
- "ld1 { v11.d }[0], [x19], #0x8\n"
+ "ld1 { v13.d }[0], [x19], #0x8\n"
"tbz x8, #0, 7f\n"
- "ld1 { v11.s }[2], [x19]\n"
+ "ld1 { v13.s }[2], [x19]\n"
"b 7f\n"
"6:" // Oddments: Load bias: Bit 2: Unset: Bit 1: Unset
"tbz x8, #0, 7f\n"
- "ld1 { v11.s }[0], [x19]\n"
+ "ld1 { v13.s }[0], [x19]\n"
"7:" // Oddments: Load bias: Bit 2: End
- "mov v23.16b, v11.16b\n"
- "ldr d0, [x16, #0x0]\n"
- "mov v9.16b, v26.16b\n"
- "ldr d1, [x16, #0x8]\n"
- "mov v12.16b, v11.16b\n"
- "ldr d2, [x16, #0x10]\n"
- "mov v22.16b, v26.16b\n"
- "ldr d3, [x16, #0x18]\n"
- "mov v24.16b, v11.16b\n"
- "ldr d4, [x16, #0x20]\n"
- "mov v10.16b, v26.16b\n"
- "ldr d5, [x16, #0x28]\n"
- "ssubl v0.8h, v0.8b, v17.8b\n"
- "ldr d6, [x16, #0x30]\n"
- "ssubl v1.8h, v1.8b, v17.8b\n"
- "ldr d7, [x16, #0x38]\n"
- "ssubl v2.8h, v2.8b, v17.8b\n"
- "ldr d8, [x16, #0x40]\n"
- "ssubl v3.8h, v3.8b, v17.8b\n"
- "ldp x23, x22, [x14, #0x0]\n"
- "add x23, x23, x17\n"
- "ssubl v4.8h, v4.8b, v17.8b\n"
- "ldp x21, x20, [x14, #0x10]\n"
- "ssubl v5.8h, v5.8b, v17.8b\n"
- "ldr x19, [x14, #0x20]\n"
- "ssubl v6.8h, v6.8b, v17.8b\n"
- "add x22, x22, x17\n"
- "ssubl v7.8h, v7.8b, v17.8b\n"
- "add x21, x21, x17\n"
- "ssubl v8.8h, v8.8b, v17.8b\n"
- "add x20, x20, x17\n"
- "add x19, x19, x17\n"
+ "ldr d0, [x17, #0x0]\n"
+ "ldr d1, [x17, #0x8]\n"
+ "mov v19.16b, v13.16b\n"
+ "mov v11.16b, v26.16b\n"
+ "ldr d2, [x17, #0x10]\n"
+ "ldr d3, [x17, #0x18]\n"
+ "mov v18.16b, v13.16b\n"
+ "mov v24.16b, v26.16b\n"
+ "ldr d4, [x17, #0x20]\n"
+ "ldr d5, [x17, #0x28]\n"
+ "mov v9.16b, v13.16b\n"
+ "mov v23.16b, v26.16b\n"
+ "ldr d6, [x17, #0x30]\n"
+ "ldr d7, [x17, #0x38]\n"
+ "ssubl v0.8h, v0.8b, v12.8b\n"
+ "ssubl v1.8h, v1.8b, v12.8b\n"
+ "ldr d8, [x17, #0x40]\n"
+ "ldp x23, x22, [x12, #0x0]\n"
+ "ssubl v2.8h, v2.8b, v12.8b\n"
+ "ssubl v3.8h, v3.8b, v12.8b\n"
+ "ldp x21, x20, [x12, #0x10]\n"
+ "ldr x19, [x12, #0x20]\n"
+ "ssubl v4.8h, v4.8b, v12.8b\n"
+ "ssubl v5.8h, v5.8b, v12.8b\n"
+ "ssubl v6.8h, v6.8b, v12.8b\n"
+ "ssubl v7.8h, v7.8b, v12.8b\n"
+ "ssubl v8.8h, v8.8b, v12.8b\n"
+ "add x23, x23, x15\n"
+ "add x22, x22, x15\n"
+ "add x21, x21, x15\n"
+ "add x20, x20, x15\n"
+ "add x19, x19, x15\n"
"tbz x8, #2, 9f\n"
"ld1 { v31.s }[0], [x23], #0x4\n"
"ld1 { v30.s }[0], [x22], #0x4\n"
@@ -690,33 +674,33 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"ld1 { v28.b }[0], [x20]\n"
"ld1 { v27.b }[0], [x19]\n"
"11:" // Oddments: Initial loads: Bit 2: End
- "usubl v31.8h, v31.8b, v21.8b\n"
- "ldr x21, [x14, #0x28]\n"
- "add x21, x21, x17\n"
- "usubl v30.8h, v30.8b, v21.8b\n"
- "usubl v29.8h, v29.8b, v21.8b\n"
- "usubl v28.8h, v28.8b, v21.8b\n"
- "usubl v27.8h, v27.8b, v21.8b\n"
- "smlal v11.4s, v31.4h, v4.4h\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "smlal v13.4s, v31.4h, v4.4h\n"
"smlal2 v26.4s, v31.8h, v4.8h\n"
- "smlal v23.4s, v31.4h, v3.4h\n"
- "smlal2 v9.4s, v31.8h, v3.8h\n"
- "smlal v12.4s, v31.4h, v1.4h\n"
- "smlal2 v22.4s, v31.8h, v1.8h\n"
- "smlal v24.4s, v31.4h, v0.4h\n"
- "smlal2 v10.4s, v31.8h, v0.8h\n"
- "smlal v11.4s, v30.4h, v0.4h\n"
+ "ldr x21, [x12, #0x28]\n"
+ "smlal v19.4s, v31.4h, v3.4h\n"
+ "smlal2 v11.4s, v31.8h, v3.8h\n"
+ "usubl v30.8h, v30.8b, v22.8b\n"
+ "add x21, x21, x15\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "smlal v18.4s, v31.4h, v1.4h\n"
+ "smlal2 v24.4s, v31.8h, v1.8h\n"
+ "smlal v9.4s, v31.4h, v0.4h\n"
+ "smlal2 v23.4s, v31.8h, v0.8h\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "smlal v13.4s, v30.4h, v0.4h\n"
"smlal2 v26.4s, v30.8h, v0.8h\n"
- "smlal v23.4s, v29.4h, v2.4h\n"
- "smlal2 v9.4s, v29.8h, v2.8h\n"
- "smlal v11.4s, v28.4h, v5.4h\n"
+ "usubl v27.8h, v27.8b, v22.8b\n"
+ "smlal v19.4s, v29.4h, v2.4h\n"
+ "smlal2 v11.4s, v29.8h, v2.8h\n"
+ "smlal v13.4s, v28.4h, v5.4h\n"
"smlal2 v26.4s, v28.8h, v5.8h\n"
- "smlal v23.4s, v28.4h, v4.4h\n"
- "smlal2 v9.4s, v28.8h, v4.8h\n"
- "smlal v12.4s, v28.4h, v2.4h\n"
- "smlal2 v22.4s, v28.8h, v2.8h\n"
- "smlal v24.4s, v28.4h, v1.4h\n"
- "smlal2 v10.4s, v28.8h, v1.8h\n"
+ "smlal v19.4s, v28.4h, v4.4h\n"
+ "smlal2 v11.4s, v28.8h, v4.8h\n"
+ "smlal v18.4s, v28.4h, v2.4h\n"
+ "smlal2 v24.4s, v28.8h, v2.8h\n"
+ "smlal v9.4s, v28.4h, v1.4h\n"
+ "smlal2 v23.4s, v28.8h, v1.8h\n"
"tbz x8, #2, 13f\n"
"ld1 { v31.s }[0], [x21], #0x4\n"
"tbz x8, #1, 12f\n"
@@ -738,19 +722,19 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"tbz x8, #0, 15f\n"
"ld1 { v31.b }[0], [x21]\n"
"15:" // Oddments: Load (3, 0): Bit 2: End
- "usubl v31.8h, v31.8b, v21.8b\n"
- "ldr x20, [x14, #0x30]\n"
- "smlal v11.4s, v27.4h, v7.4h\n"
- "add x20, x20, x17\n"
- "smlal v12.4s, v31.4h, v6.4h\n"
- "smlal2 v22.4s, v31.8h, v6.8h\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "smlal v18.4s, v31.4h, v6.4h\n"
+ "smlal2 v24.4s, v31.8h, v6.8h\n"
+ "ldr x20, [x12, #0x30]\n"
+ "smlal v13.4s, v27.4h, v7.4h\n"
"smlal2 v26.4s, v27.8h, v7.8h\n"
- "smlal v23.4s, v27.4h, v6.4h\n"
- "smlal2 v9.4s, v27.8h, v6.8h\n"
- "smlal v12.4s, v27.4h, v4.4h\n"
- "smlal2 v22.4s, v27.8h, v4.8h\n"
- "smlal v24.4s, v27.4h, v3.4h\n"
- "smlal2 v10.4s, v27.8h, v3.8h\n"
+ "add x20, x20, x15\n"
+ "smlal v19.4s, v27.4h, v6.4h\n"
+ "smlal2 v11.4s, v27.8h, v6.8h\n"
+ "smlal v18.4s, v27.4h, v4.4h\n"
+ "smlal2 v24.4s, v27.8h, v4.8h\n"
+ "smlal v9.4s, v27.4h, v3.4h\n"
+ "smlal2 v23.4s, v27.8h, v3.8h\n"
"tbz x8, #2, 17f\n"
"ld1 { v29.s }[0], [x20], #0x4\n"
"tbz x8, #1, 16f\n"
@@ -772,11 +756,11 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"tbz x8, #0, 19f\n"
"ld1 { v29.b }[0], [x20]\n"
"19:" // Oddments: Load (3, 3): Bit 2: End
- "usubl v29.8h, v29.8b, v21.8b\n"
- "ldr x26, [x14, #0x38]\n"
- "smlal v24.4s, v29.4h, v8.4h\n"
- "add x26, x26, x17\n"
- "smlal2 v10.4s, v29.8h, v8.8h\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "ldr x26, [x12, #0x38]\n"
+ "smlal v9.4s, v29.4h, v8.4h\n"
+ "smlal2 v23.4s, v29.8h, v8.8h\n"
+ "add x26, x26, x15\n"
"tbz x8, #2, 21f\n"
"ld1 { v28.s }[0], [x26], #0x4\n"
"tbz x8, #1, 20f\n"
@@ -798,13 +782,13 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"tbz x8, #0, 23f\n"
"ld1 { v28.b }[0], [x26]\n"
"23:" // Oddments: Load (0, 1): Bit 2: End
- "usubl v28.8h, v28.8b, v21.8b\n"
- "ldr x25, [x14, #0x40]\n"
- "smlal v11.4s, v28.4h, v1.4h\n"
- "add x25, x25, x17\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "ldr x25, [x12, #0x40]\n"
+ "smlal v13.4s, v28.4h, v1.4h\n"
"smlal2 v26.4s, v28.8h, v1.8h\n"
- "smlal v23.4s, v28.4h, v0.4h\n"
- "smlal2 v9.4s, v28.8h, v0.8h\n"
+ "smlal v19.4s, v28.4h, v0.4h\n"
+ "smlal2 v11.4s, v28.8h, v0.8h\n"
+ "add x25, x25, x15\n"
"tbz x8, #2, 25f\n"
"ld1 { v31.s }[0], [x25], #0x4\n"
"tbz x8, #1, 24f\n"
@@ -826,13 +810,13 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"tbz x8, #0, 27f\n"
"ld1 { v31.b }[0], [x25]\n"
"27:" // Oddments: Load (0, 2): Bit 2: End
- "usubl v31.8h, v31.8b, v21.8b\n"
- "ldr x19, [x14, #0x48]\n"
- "smlal v11.4s, v31.4h, v2.4h\n"
- "add x19, x19, x17\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "ldr x19, [x12, #0x48]\n"
+ "smlal v13.4s, v31.4h, v2.4h\n"
"smlal2 v26.4s, v31.8h, v2.8h\n"
- "smlal v23.4s, v31.4h, v1.4h\n"
- "smlal2 v9.4s, v31.8h, v1.8h\n"
+ "smlal v19.4s, v31.4h, v1.4h\n"
+ "smlal2 v11.4s, v31.8h, v1.8h\n"
+ "add x19, x19, x15\n"
"tbz x8, #2, 29f\n"
"ld1 { v30.s }[0], [x19], #0x4\n"
"tbz x8, #1, 28f\n"
@@ -854,17 +838,17 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"tbz x8, #0, 31f\n"
"ld1 { v30.b }[0], [x19]\n"
"31:" // Oddments: Load (2, 2): Bit 2: End
- "usubl v30.8h, v30.8b, v21.8b\n"
- "ldr x24, [x14, #0x50]\n"
- "smlal v11.4s, v30.4h, v8.4h\n"
- "add x24, x24, x17\n"
+ "usubl v30.8h, v30.8b, v22.8b\n"
+ "ldr x24, [x12, #0x50]\n"
+ "smlal v13.4s, v30.4h, v8.4h\n"
"smlal2 v26.4s, v30.8h, v8.8h\n"
- "smlal v23.4s, v30.4h, v7.4h\n"
- "smlal2 v9.4s, v30.8h, v7.8h\n"
- "smlal v12.4s, v30.4h, v5.4h\n"
- "smlal2 v22.4s, v30.8h, v5.8h\n"
- "smlal v24.4s, v30.4h, v4.4h\n"
- "smlal2 v10.4s, v30.8h, v4.8h\n"
+ "smlal v19.4s, v30.4h, v7.4h\n"
+ "smlal2 v11.4s, v30.8h, v7.8h\n"
+ "add x24, x24, x15\n"
+ "smlal v18.4s, v30.4h, v5.4h\n"
+ "smlal2 v24.4s, v30.8h, v5.8h\n"
+ "smlal v9.4s, v30.4h, v4.4h\n"
+ "smlal2 v23.4s, v30.8h, v4.8h\n"
"tbz x8, #2, 33f\n"
"ld1 { v29.s }[0], [x24], #0x4\n"
"tbz x8, #1, 32f\n"
@@ -886,13 +870,13 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"tbz x8, #0, 35f\n"
"ld1 { v29.b }[0], [x24]\n"
"35:" // Oddments: Load (1, 0): Bit 2: End
- "usubl v29.8h, v29.8b, v21.8b\n"
- "ldr x23, [x14, #0x58]\n"
- "smlal v11.4s, v29.4h, v3.4h\n"
- "add x23, x23, x17\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "ldr x23, [x12, #0x58]\n"
+ "smlal v13.4s, v29.4h, v3.4h\n"
"smlal2 v26.4s, v29.8h, v3.8h\n"
- "smlal v12.4s, v29.4h, v0.4h\n"
- "smlal2 v22.4s, v29.8h, v0.8h\n"
+ "smlal v18.4s, v29.4h, v0.4h\n"
+ "smlal2 v24.4s, v29.8h, v0.8h\n"
+ "add x23, x23, x15\n"
"tbz x8, #2, 37f\n"
"ld1 { v28.s }[0], [x23], #0x4\n"
"tbz x8, #1, 36f\n"
@@ -914,13 +898,13 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"tbz x8, #0, 39f\n"
"ld1 { v28.b }[0], [x23]\n"
"39:" // Oddments: Load (1, 3): Bit 2: End
- "usubl v28.8h, v28.8b, v21.8b\n"
- "ldr x22, [x14, #0x60]\n"
- "smlal v23.4s, v28.4h, v5.4h\n"
- "add x22, x22, x17\n"
- "smlal2 v9.4s, v28.8h, v5.8h\n"
- "smlal v24.4s, v28.4h, v2.4h\n"
- "smlal2 v10.4s, v28.8h, v2.8h\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "ldr x22, [x12, #0x60]\n"
+ "smlal v19.4s, v28.4h, v5.4h\n"
+ "smlal2 v11.4s, v28.8h, v5.8h\n"
+ "smlal v9.4s, v28.4h, v2.4h\n"
+ "smlal2 v23.4s, v28.8h, v2.8h\n"
+ "add x22, x22, x15\n"
"tbz x8, #2, 41f\n"
"ld1 { v31.s }[0], [x22], #0x4\n"
"tbz x8, #1, 40f\n"
@@ -942,13 +926,13 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"tbz x8, #0, 43f\n"
"ld1 { v31.b }[0], [x22]\n"
"43:" // Oddments: Load (2, 0): Bit 2: End
- "usubl v31.8h, v31.8b, v21.8b\n"
- "ldr x21, [x14, #0x68]\n"
- "smlal v11.4s, v31.4h, v6.4h\n"
- "add x21, x21, x17\n"
+ "usubl v31.8h, v31.8b, v22.8b\n"
+ "ldr x21, [x12, #0x68]\n"
+ "smlal v13.4s, v31.4h, v6.4h\n"
"smlal2 v26.4s, v31.8h, v6.8h\n"
- "smlal v12.4s, v31.4h, v3.4h\n"
- "smlal2 v22.4s, v31.8h, v3.8h\n"
+ "smlal v18.4s, v31.4h, v3.4h\n"
+ "smlal2 v24.4s, v31.8h, v3.8h\n"
+ "add x21, x21, x15\n"
"tbz x8, #2, 45f\n"
"ld1 { v30.s }[0], [x21], #0x4\n"
"tbz x8, #1, 44f\n"
@@ -970,13 +954,13 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"tbz x8, #0, 47f\n"
"ld1 { v30.b }[0], [x21]\n"
"47:" // Oddments: Load (2, 3): Bit 2: End
- "usubl v30.8h, v30.8b, v21.8b\n"
- "ldr x20, [x14, #0x70]\n"
- "smlal v23.4s, v30.4h, v8.4h\n"
- "add x20, x20, x17\n"
- "smlal2 v9.4s, v30.8h, v8.8h\n"
- "smlal v24.4s, v30.4h, v5.4h\n"
- "smlal2 v10.4s, v30.8h, v5.8h\n"
+ "usubl v30.8h, v30.8b, v22.8b\n"
+ "ldr x20, [x12, #0x70]\n"
+ "smlal v19.4s, v30.4h, v8.4h\n"
+ "smlal2 v11.4s, v30.8h, v8.8h\n"
+ "smlal v9.4s, v30.4h, v5.4h\n"
+ "smlal2 v23.4s, v30.8h, v5.8h\n"
+ "add x20, x20, x15\n"
"tbz x8, #2, 49f\n"
"ld1 { v29.s }[0], [x20], #0x4\n"
"tbz x8, #1, 48f\n"
@@ -998,13 +982,13 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"tbz x8, #0, 51f\n"
"ld1 { v29.b }[0], [x20]\n"
"51:" // Oddments: Load (3, 1): Bit 2: End
- "usubl v29.8h, v29.8b, v21.8b\n"
- "ldr x19, [x14, #0x78]\n"
- "smlal v12.4s, v29.4h, v7.4h\n"
- "add x19, x19, x17\n"
- "smlal2 v22.4s, v29.8h, v7.8h\n"
- "smlal v24.4s, v29.4h, v6.4h\n"
- "smlal2 v10.4s, v29.8h, v6.8h\n"
+ "usubl v29.8h, v29.8b, v22.8b\n"
+ "ldr x19, [x12, #0x78]\n"
+ "smlal v18.4s, v29.4h, v7.4h\n"
+ "smlal2 v24.4s, v29.8h, v7.8h\n"
+ "smlal v9.4s, v29.4h, v6.4h\n"
+ "smlal2 v23.4s, v29.8h, v6.8h\n"
+ "add x19, x19, x15\n"
"tbz x8, #2, 53f\n"
"ld1 { v28.s }[0], [x19], #0x4\n"
"tbz x8, #1, 52f\n"
@@ -1026,160 +1010,150 @@ void a64_u8s8u8q_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
"tbz x8, #0, 55f\n"
"ld1 { v28.b }[0], [x19]\n"
"55:" // Oddments: Load (3, 2): Bit 2: End
- "usubl v28.8h, v28.8b, v21.8b\n"
- "smlal v12.4s, v28.4h, v8.4h\n"
- "smlal2 v22.4s, v28.8h, v8.8h\n"
- "smlal v24.4s, v28.4h, v7.4h\n"
- "smlal2 v10.4s, v28.8h, v7.8h\n"
+ "usubl v28.8h, v28.8b, v22.8b\n"
+ "smlal v18.4s, v28.4h, v8.4h\n"
+ "smlal2 v24.4s, v28.8h, v8.8h\n"
+ "smlal v9.4s, v28.4h, v7.4h\n"
+ "smlal2 v23.4s, v28.8h, v7.8h\n"
"tbz x8, #2, 57f\n"
- "ld1 { v25.4s }, [x13], #0x10\n"
- "ld1 { v18.4s }, [x11], #0x10\n"
+ "ld1 { v21.4s }, [x13], #0x10\n"
+ "ld1 { v25.4s }, [x11], #0x10\n"
"tbz x8, #1, 56f\n"
- "ld1 { v16.d }[0], [x13], #0x8\n"
- "ld1 { v20.d }[0], [x11], #0x8\n"
+ "ld1 { v10.d }[0], [x13], #0x8\n"
+ "ld1 { v16.d }[0], [x11], #0x8\n"
"tbz x8, #0, 59f\n"
- "ld1 { v16.s }[2], [x13]\n"
- "ld1 { v20.s }[2], [x11]\n"
+ "ld1 { v10.s }[2], [x13]\n"
+ "ld1 { v16.s }[2], [x11]\n"
"b 59f\n"
"56:" // Oddments: Load requant params: Bit 2: Bit 1: Unset
"tbz x8, #0, 59f\n"
- "ld1 { v16.s }[0], [x13]\n"
- "ld1 { v20.s }[0], [x11]\n"
+ "ld1 { v10.s }[0], [x13]\n"
+ "ld1 { v16.s }[0], [x11]\n"
"b 59f\n"
"57:" // Oddments: Load requant params: Bit 2: Unset
"tbz x8, #1, 58f\n"
- "ld1 { v25.d }[0], [x13], #0x8\n"
- "ld1 { v18.d }[0], [x11], #0x8\n"
+ "ld1 { v21.d }[0], [x13], #0x8\n"
+ "ld1 { v25.d }[0], [x11], #0x8\n"
"tbz x8, #0, 59f\n"
- "ld1 { v25.s }[2], [x13]\n"
- "ld1 { v18.s }[2], [x11]\n"
+ "ld1 { v21.s }[2], [x13]\n"
+ "ld1 { v25.s }[2], [x11]\n"
"b 59f\n"
"58:" // Oddments: Load requant params: Bit 2: Unset: Bit 1: Unset
"tbz x8, #0, 59f\n"
- "ld1 { v25.s }[0], [x13]\n"
- "ld1 { v18.s }[0], [x11]\n"
+ "ld1 { v21.s }[0], [x13]\n"
+ "ld1 { v25.s }[0], [x11]\n"
"59:" // Oddments: Load requant params: Bit 2: End
- "sqrdmulh v11.4s, v11.4s, v25.4s\n"
- "add x10, x10, x15\n"
- "sqrdmulh v26.4s, v26.4s, v16.4s\n"
- "add x9, x9, x15\n"
- "sqrdmulh v23.4s, v23.4s, v25.4s\n"
- "add x28, x28, x15\n"
- "sqrdmulh v9.4s, v9.4s, v16.4s\n"
- "add x27, x27, x15\n"
- "sqrdmulh v12.4s, v12.4s, v25.4s\n"
- "and v19.16b, v11.16b, v18.16b\n"
- "and v5.16b, v26.16b, v20.16b\n"
- "and v28.16b, v23.16b, v18.16b\n"
- "sshr v19.4s, v19.4s, #0x1f\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sshr v28.4s, v28.4s, #0x1f\n"
- "sqadd v11.4s, v11.4s, v19.4s\n"
- "sqadd v26.4s, v26.4s, v5.4s\n"
- "sqadd v23.4s, v23.4s, v28.4s\n"
- "and v8.16b, v9.16b, v20.16b\n"
- "srshl v11.4s, v11.4s, v18.4s\n"
- "srshl v26.4s, v26.4s, v20.4s\n"
- "srshl v23.4s, v23.4s, v18.4s\n"
- "sshr v8.4s, v8.4s, #0x1f\n"
- "add v11.4s, v11.4s, v13.4s\n"
- "add v26.4s, v26.4s, v13.4s\n"
- "add v23.4s, v23.4s, v13.4s\n"
- "smin v11.4s, v11.4s, v14.4s\n"
- "smin v26.4s, v26.4s, v14.4s\n"
- "smin v23.4s, v23.4s, v14.4s\n"
- "smax v11.4s, v11.4s, v15.4s\n"
- "smax v26.4s, v26.4s, v15.4s\n"
- "smax v23.4s, v23.4s, v15.4s\n"
- "sqadd v9.4s, v9.4s, v8.4s\n"
- "uzp1 v11.16b, v11.16b, v26.16b\n"
- "and v1.16b, v12.16b, v18.16b\n"
- "uzp1 v11.16b, v11.16b, v11.16b\n"
- "srshl v9.4s, v9.4s, v20.4s\n"
- "sshr v1.4s, v1.4s, #0x1f\n"
- "sqrdmulh v22.4s, v22.4s, v16.4s\n"
- "sqrdmulh v24.4s, v24.4s, v25.4s\n"
- "add v9.4s, v9.4s, v13.4s\n"
- "sqadd v12.4s, v12.4s, v1.4s\n"
- "and v0.16b, v22.16b, v20.16b\n"
- "smin v9.4s, v9.4s, v14.4s\n"
- "and v26.16b, v24.16b, v18.16b\n"
- "srshl v12.4s, v12.4s, v18.4s\n"
- "smax v9.4s, v9.4s, v15.4s\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "sshr v26.4s, v26.4s, #0x1f\n"
- "uzp1 v23.16b, v23.16b, v9.16b\n"
- "add v12.4s, v12.4s, v13.4s\n"
- "uzp1 v23.16b, v23.16b, v23.16b\n"
- "sqadd v22.4s, v22.4s, v0.4s\n"
- "smin v12.4s, v12.4s, v14.4s\n"
- "sqadd v24.4s, v24.4s, v26.4s\n"
- "sqrdmulh v10.4s, v10.4s, v16.4s\n"
- "smax v12.4s, v12.4s, v15.4s\n"
- "srshl v22.4s, v22.4s, v20.4s\n"
- "srshl v24.4s, v24.4s, v18.4s\n"
- "and v16.16b, v10.16b, v20.16b\n"
- "add v22.4s, v22.4s, v13.4s\n"
- "add v24.4s, v24.4s, v13.4s\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "smin v22.4s, v22.4s, v14.4s\n"
- "smin v24.4s, v24.4s, v14.4s\n"
- "sqadd v10.4s, v10.4s, v16.4s\n"
- "smax v22.4s, v22.4s, v15.4s\n"
- "smax v24.4s, v24.4s, v15.4s\n"
- "srshl v10.4s, v10.4s, v20.4s\n"
- "uzp1 v12.16b, v12.16b, v22.16b\n"
- "uzp1 v12.16b, v12.16b, v12.16b\n"
- "add v10.4s, v10.4s, v13.4s\n"
- "smin v10.4s, v10.4s, v14.4s\n"
- "smax v10.4s, v10.4s, v15.4s\n"
- "uzp1 v24.16b, v24.16b, v10.16b\n"
- "uzp1 v24.16b, v24.16b, v24.16b\n"
+ "sqdmulh v13.4s, v13.4s, v21.4s\n"
+ "sqdmulh v19.4s, v19.4s, v21.4s\n"
+ "add x10, x10, x14\n"
+ "add x9, x9, x14\n"
+ "sqdmulh v18.4s, v18.4s, v21.4s\n"
+ "sqdmulh v9.4s, v9.4s, v21.4s\n"
+ "add x28, x28, x14\n"
+ "add x27, x27, x14\n"
+ "and v7.16b, v13.16b, v25.16b\n"
+ "sqdmulh v26.4s, v26.4s, v10.4s\n"
+ "and v4.16b, v19.16b, v25.16b\n"
+ "sqdmulh v11.4s, v11.4s, v10.4s\n"
+ "and v21.16b, v18.16b, v25.16b\n"
+ "sqdmulh v24.4s, v24.4s, v10.4s\n"
+ "and v20.16b, v9.16b, v25.16b\n"
+ "sqdmulh v23.4s, v23.4s, v10.4s\n"
+ "sshr v7.4s, v7.4s, #0x1f\n"
+ "and v29.16b, v26.16b, v16.16b\n"
+ "sshr v4.4s, v4.4s, #0x1f\n"
+ "and v10.16b, v11.16b, v16.16b\n"
+ "sshr v21.4s, v21.4s, #0x1f\n"
+ "and v31.16b, v24.16b, v16.16b\n"
+ "sshr v20.4s, v20.4s, #0x1f\n"
+ "and v30.16b, v23.16b, v16.16b\n"
+ "sqadd v13.4s, v13.4s, v7.4s\n"
+ "sshr v29.4s, v29.4s, #0x1f\n"
+ "sqadd v19.4s, v19.4s, v4.4s\n"
+ "sshr v10.4s, v10.4s, #0x1f\n"
+ "sqadd v18.4s, v18.4s, v21.4s\n"
+ "sshr v31.4s, v31.4s, #0x1f\n"
+ "sqadd v9.4s, v9.4s, v20.4s\n"
+ "sshr v30.4s, v30.4s, #0x1f\n"
+ "srshl v13.4s, v13.4s, v25.4s\n"
+ "sqadd v26.4s, v26.4s, v29.4s\n"
+ "srshl v19.4s, v19.4s, v25.4s\n"
+ "sqadd v11.4s, v11.4s, v10.4s\n"
+ "srshl v18.4s, v18.4s, v25.4s\n"
+ "sqadd v24.4s, v24.4s, v31.4s\n"
+ "srshl v9.4s, v9.4s, v25.4s\n"
+ "sqadd v23.4s, v23.4s, v30.4s\n"
+ "srshl v26.4s, v26.4s, v16.4s\n"
+ "sqxtn v13.4h, v13.4s\n"
+ "srshl v11.4s, v11.4s, v16.4s\n"
+ "sqxtn v19.4h, v19.4s\n"
+ "srshl v24.4s, v24.4s, v16.4s\n"
+ "sqxtn v18.4h, v18.4s\n"
+ "srshl v23.4s, v23.4s, v16.4s\n"
+ "sqxtn v9.4h, v9.4s\n"
+ "sqxtn2 v13.8h, v26.4s\n"
+ "sqxtn2 v19.8h, v11.4s\n"
+ "sqxtn2 v18.8h, v24.4s\n"
+ "sqxtn2 v9.8h, v23.4s\n"
+ "sqadd v13.8h, v13.8h, v14.8h\n"
+ "sqadd v19.8h, v19.8h, v14.8h\n"
+ "sqadd v18.8h, v18.8h, v14.8h\n"
+ "sqadd v9.8h, v9.8h, v14.8h\n"
+ "smax v13.8h, v13.8h, v17.8h\n"
+ "smax v19.8h, v19.8h, v17.8h\n"
+ "smax v18.8h, v18.8h, v17.8h\n"
+ "smax v9.8h, v9.8h, v17.8h\n"
+ "smin v13.8h, v13.8h, v15.8h\n"
+ "smin v19.8h, v19.8h, v15.8h\n"
+ "smin v18.8h, v18.8h, v15.8h\n"
+ "smin v9.8h, v9.8h, v15.8h\n"
+ "uzp1 v13.16b, v13.16b, v13.16b\n"
+ "uzp1 v19.16b, v19.16b, v19.16b\n"
+ "uzp1 v18.16b, v18.16b, v18.16b\n"
+ "uzp1 v9.16b, v9.16b, v9.16b\n"
"tbz x8, #2, 61f\n"
- "st1 { v11.s }[0], [x10], #0x4\n"
- "st1 { v23.s }[0], [x9], #0x4\n"
- "st1 { v12.s }[0], [x28], #0x4\n"
- "st1 { v24.s }[0], [x27], #0x4\n"
+ "st1 { v13.s }[0], [x10], #0x4\n"
+ "st1 { v19.s }[0], [x9], #0x4\n"
+ "st1 { v18.s }[0], [x28], #0x4\n"
+ "st1 { v9.s }[0], [x27], #0x4\n"
"tbz x8, #1, 60f\n"
- "st1 { v11.h }[2], [x10], #0x2\n"
- "st1 { v23.h }[2], [x9], #0x2\n"
- "st1 { v12.h }[2], [x28], #0x2\n"
- "st1 { v24.h }[2], [x27], #0x2\n"
+ "st1 { v13.h }[2], [x10], #0x2\n"
+ "st1 { v19.h }[2], [x9], #0x2\n"
+ "st1 { v18.h }[2], [x28], #0x2\n"
+ "st1 { v9.h }[2], [x27], #0x2\n"
"tbz x8, #0, 63f\n"
- "st1 { v11.b }[6], [x10], #0x1\n"
- "st1 { v23.b }[6], [x9], #0x1\n"
- "st1 { v12.b }[6], [x28], #0x1\n"
- "st1 { v24.b }[6], [x27], #0x1\n"
+ "st1 { v13.b }[6], [x10], #0x1\n"
+ "st1 { v19.b }[6], [x9], #0x1\n"
+ "st1 { v18.b }[6], [x28], #0x1\n"
+ "st1 { v9.b }[6], [x27], #0x1\n"
"b 63f\n"
"60:" // Oddments: Bit 2: Bit 1: Unset
"tbz x8, #0, 63f\n"
- "st1 { v11.b }[4], [x10], #0x1\n"
- "st1 { v23.b }[4], [x9], #0x1\n"
- "st1 { v12.b }[4], [x28], #0x1\n"
- "st1 { v24.b }[4], [x27], #0x1\n"
+ "st1 { v13.b }[4], [x10], #0x1\n"
+ "st1 { v19.b }[4], [x9], #0x1\n"
+ "st1 { v18.b }[4], [x28], #0x1\n"
+ "st1 { v9.b }[4], [x27], #0x1\n"
"b 63f\n"
"61:" // Oddments: Bit 2: Unset
"tbz x8, #1, 62f\n"
- "st1 { v11.h }[0], [x10], #0x2\n"
- "st1 { v23.h }[0], [x9], #0x2\n"
- "st1 { v12.h }[0], [x28], #0x2\n"
- "st1 { v24.h }[0], [x27], #0x2\n"
+ "st1 { v13.h }[0], [x10], #0x2\n"
+ "st1 { v19.h }[0], [x9], #0x2\n"
+ "st1 { v18.h }[0], [x28], #0x2\n"
+ "st1 { v9.h }[0], [x27], #0x2\n"
"tbz x8, #0, 63f\n"
- "st1 { v11.b }[2], [x10], #0x1\n"
- "st1 { v23.b }[2], [x9], #0x1\n"
- "st1 { v12.b }[2], [x28], #0x1\n"
- "st1 { v24.b }[2], [x27], #0x1\n"
+ "st1 { v13.b }[2], [x10], #0x1\n"
+ "st1 { v19.b }[2], [x9], #0x1\n"
+ "st1 { v18.b }[2], [x28], #0x1\n"
+ "st1 { v9.b }[2], [x27], #0x1\n"
"b 63f\n"
"62:" // Oddments: Bit 2: Unset: Bit 1: Unset
"tbz x8, #0, 63f\n"
- "st1 { v11.b }[0], [x10], #0x1\n"
- "st1 { v23.b }[0], [x9], #0x1\n"
- "st1 { v12.b }[0], [x28], #0x1\n"
- "st1 { v24.b }[0], [x27], #0x1\n"
+ "st1 { v13.b }[0], [x10], #0x1\n"
+ "st1 { v19.b }[0], [x9], #0x1\n"
+ "st1 { v18.b }[0], [x28], #0x1\n"
+ "st1 { v9.b }[0], [x27], #0x1\n"
"63:" // Oddments: Bit 2: End
-
"64:" // End
-
:
: [offsetof_Params_bias] "I" (offsetof(Params, bias)), [offsetof_Params_inptrs] "I" (offsetof(Params, inptrs)), [offsetof_Params_n_channels] "I" (offsetof(Params, n_channels)), [offsetof_Params_outptrs] "I" (offsetof(Params, outptrs)), [offsetof_Params_requant] "I" (offsetof(Params, requant)), [offsetof_Params_requant_muls] "I" (offsetof(Params, requant_muls)), [offsetof_Params_requant_shifts] "I" (offsetof(Params, requant_shifts)), [offsetof_Params_weights] "I" (offsetof(Params, weights)), [offsetof_Requantize32_a_offset] "I" (offsetof(arm_gemm::Requantize32, a_offset)), [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)), [params] "r" (&params)
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"