aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp1824
1 files changed, 912 insertions, 912 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp
index d1872c90f8..62ad1fc0f5 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst/generic.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2022-2023 Arm Limited.
+ * Copyright (c) 2022-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -45,7 +45,7 @@ void a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
{
struct Params
{
- long unsigned int n_channels;
+ uint64_t n_channels;
const void *weights;
const int32_t *bias;
const arm_gemm::Requantize32 *requant;
@@ -55,7 +55,7 @@ void a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
const uint8_t *inptrs[16];
Params(
- long unsigned int n_channels,
+ uint64_t n_channels,
const uint8_t *const *inptrs_raw,
const void *const weights,
const int32_t *const bias,
@@ -91,1070 +91,1070 @@ void a64_u8qa_nhwc_3x3_s1_output2x2_mla_depthfirst_impl(
requant_muls, requant_shifts, outptrs);
__asm__ __volatile__(
- "ldr x16, [%x[params], %[offsetof_Params_n_channels]]\n"
+ "ldr x17, [%x[params], %[offsetof_Params_n_channels]]\n"
"ldr x23, [%x[params], %[offsetof_Params_requant]]\n"
- "lsr x15, x16, #0x3\n"
- "add x20, x23, %[offsetof_Requantize32_b_offset]\n"
- "ld1r { v18.16b }, [x20]\n"
+ "mov x16, #0x0\n"
+ "mov x15, #0x0\n"
"ldr x22, [%x[params], %[offsetof_Params_outptrs]]\n"
+ "add x14, %x[params], %[offsetof_Params_inptrs]\n"
+ "ldr x13, [%x[params], %[offsetof_Params_weights]]\n"
+ "ldr x12, [%x[params], %[offsetof_Params_requant_muls]]\n"
+ "ldr x11, [%x[params], %[offsetof_Params_requant_shifts]]\n"
+ "lsr x10, x17, #0x3\n"
+ "add x20, x23, %[offsetof_Requantize32_b_offset]\n"
+ "ld1r { v29.16b }, [x20]\n"
"add x21, x23, %[offsetof_Requantize32_c_offset]\n"
"add x20, x23, %[offsetof_Requantize32_minval]\n"
- "ld1r { v5.8h }, [x21]\n"
- "ld1r { v14.8h }, [x20]\n"
+ "ld1r { v12.8h }, [x21]\n"
+ "ld1r { v15.8h }, [x20]\n"
"add x20, x23, %[offsetof_Requantize32_maxval]\n"
- "mov x14, #0x0\n"
- "ld1r { v12.8h }, [x20]\n"
- "mov x13, #0x0\n"
- "add x12, %x[params], %[offsetof_Params_inptrs]\n"
- "ldr x11, [%x[params], %[offsetof_Params_weights]]\n"
- "ldr x10, [%x[params], %[offsetof_Params_requant_muls]]\n"
- "ldr x9, [%x[params], %[offsetof_Params_requant_shifts]]\n"
- "ldp x28, x27, [x22, #0x0]\n"
- "ldp x26, x25, [x22, #0x10]\n"
- "cbz x15, 3f\n"
- "ldr d19, [x11, #0x0]\n"
- "ldr d7, [x11, #0x8]\n"
- "subs x15, x15, #0x1\n"
- "usubl v19.8h, v19.8b, v18.8b\n"
- "ldr d1, [x11, #0x10]\n"
- "ldr d17, [x11, #0x18]\n"
- "usubl v7.8h, v7.8b, v18.8b\n"
- "usubl v1.8h, v1.8b, v18.8b\n"
- "ldr d8, [x11, #0x20]\n"
- "ldr d31, [x11, #0x28]\n"
- "usubl v17.8h, v17.8b, v18.8b\n"
- "usubl v8.8h, v8.8b, v18.8b\n"
- "ldr d29, [x11, #0x30]\n"
- "ldr d16, [x11, #0x38]\n"
- "usubl v31.8h, v31.8b, v18.8b\n"
- "usubl v29.8h, v29.8b, v18.8b\n"
- "ldr d4, [x11, #0x40]\n"
+ "ld1r { v17.8h }, [x20]\n"
+ "ldp x9, x28, [x22, #0x0]\n"
+ "ldp x27, x26, [x22, #0x10]\n"
+ "cbz x10, 3f\n"
+ "ldr d22, [x13, #0x0]\n"
+ "ldr d24, [x13, #0x8]\n"
+ "subs x10, x10, #0x1\n"
+ "ldr d9, [x13, #0x10]\n"
+ "ldr d7, [x13, #0x18]\n"
+ "ldr d25, [x13, #0x20]\n"
+ "ldr d4, [x13, #0x28]\n"
+ "ldr d13, [x13, #0x30]\n"
+ "ldr d14, [x13, #0x38]\n"
+ "usubl v22.8h, v22.8b, v29.8b\n"
+ "usubl v24.8h, v24.8b, v29.8b\n"
+ "ldr d2, [x13, #0x40]\n"
"ldr x20, [%x[params], %[offsetof_Params_bias]]\n"
- "usubl v16.8h, v16.8b, v18.8b\n"
- "usubl v4.8h, v4.8b, v18.8b\n"
- "ldr q28, [x20, #0x0]\n"
- "ldr q9, [x20, #0x10]\n"
+ "usubl v9.8h, v9.8b, v29.8b\n"
+ "usubl v7.8h, v7.8b, v29.8b\n"
+ "usubl v25.8h, v25.8b, v29.8b\n"
+ "usubl v4.8h, v4.8b, v29.8b\n"
+ "usubl v13.8h, v13.8b, v29.8b\n"
+ "usubl v14.8h, v14.8b, v29.8b\n"
+ "ldr q20, [x20, #0x0]\n"
+ "ldr q1, [x20, #0x10]\n"
+ "ldp x23, x22, [x14, #0x0]\n"
"add x20, x20, #0x20\n"
+ "usubl v2.8h, v2.8b, v29.8b\n"
"str x20, [%x[params], %[offsetof_Params_bias]]\n"
- "ldp x23, x22, [x12, #0x0]\n"
- "ldp x21, x20, [x12, #0x10]\n"
- "mov v3.16b, v28.16b\n"
- "mov v30.16b, v9.16b\n"
- "ldr d23, [x23, x14]\n"
- "ldr d10, [x22, x14]\n"
- "mov v0.16b, v28.16b\n"
- "mov v22.16b, v9.16b\n"
- "ldr d11, [x21, x14]\n"
- "ldr d13, [x20, x14]\n"
- "mov v6.16b, v28.16b\n"
- "mov v2.16b, v9.16b\n"
- "ldr x20, [x12, #0x20]\n"
- "ldr d27, [x20, x14]\n"
- "ushll v23.8h, v23.8b, #0x0\n"
- "ushll v10.8h, v10.8b, #0x0\n"
+ "mov v8.16b, v20.16b\n"
+ "mov v21.16b, v1.16b\n"
+ "mov v3.16b, v20.16b\n"
+ "mov v30.16b, v1.16b\n"
+ "ldp x21, x20, [x14, #0x10]\n"
+ "mov v10.16b, v20.16b\n"
+ "mov v27.16b, v1.16b\n"
+ "ldr d11, [x23, x16]\n"
+ "ldr d28, [x22, x16]\n"
+ "ldr d18, [x21, x16]\n"
+ "ldr d19, [x20, x16]\n"
+ "ldr x20, [x14, #0x20]\n"
"ushll v11.8h, v11.8b, #0x0\n"
- "ushll v13.8h, v13.8b, #0x0\n"
- "ushll v27.8h, v27.8b, #0x0\n"
+ "ushll v28.8h, v28.8b, #0x0\n"
+ "ushll v18.8h, v18.8b, #0x0\n"
+ "ushll v19.8h, v19.8b, #0x0\n"
+ "ldr d23, [x20, x16]\n"
+ "ushll v23.8h, v23.8b, #0x0\n"
"beq 2f\n"
"1:" // Loop
- "ldr q24, [x10, #0x0]\n"
- "ldr q25, [x9, #0x0]\n"
- "smlal v28.4s, v23.4h, v8.4h\n"
- "smlal2 v9.4s, v23.8h, v8.8h\n"
- "ldr q20, [x10, #0x10]\n"
- "ldr q26, [x9, #0x10]\n"
- "smlal v28.4s, v10.4h, v19.4h\n"
- "smlal v3.4s, v23.4h, v17.4h\n"
- "ldr x20, [x12, #0x28]\n"
- "ldr d21, [x20, x14]\n"
- "smlal v0.4s, v23.4h, v7.4h\n"
- "smlal v6.4s, v23.4h, v19.4h\n"
- "smlal2 v9.4s, v10.8h, v19.8h\n"
- "ldr x20, [x12, #0x38]\n"
- "ldr d10, [x20, x14]\n"
- "smlal v28.4s, v13.4h, v31.4h\n"
- "smlal2 v30.4s, v23.8h, v17.8h\n"
- "smlal2 v22.4s, v23.8h, v7.8h\n"
- "ldr x20, [x12, #0x30]\n"
- "ldr d15, [x20, x14]\n"
- "smlal2 v2.4s, v23.8h, v19.8h\n"
- "smlal v3.4s, v11.4h, v1.4h\n"
- "ushll v21.8h, v21.8b, #0x0\n"
- "ldr x20, [x12, #0x40]\n"
- "ldr d23, [x20, x14]\n"
- "smlal v0.4s, v13.4h, v1.4h\n"
- "smlal v6.4s, v13.4h, v7.4h\n"
- "ushll v10.8h, v10.8b, #0x0\n"
- "smlal2 v9.4s, v13.8h, v31.8h\n"
- "smlal v28.4s, v27.4h, v16.4h\n"
- "ldr x20, [x12, #0x48]\n"
- "ushll v15.8h, v15.8b, #0x0\n"
- "smlal2 v30.4s, v11.8h, v1.8h\n"
- "ldr d11, [x20, x14]\n"
- "smlal2 v22.4s, v13.8h, v1.8h\n"
- "ushll v23.8h, v23.8b, #0x0\n"
- "smlal2 v2.4s, v13.8h, v7.8h\n"
- "smlal v3.4s, v13.4h, v8.4h\n"
- "ldr x21, [x12, #0x50]\n"
- "ldr x20, [x12, #0x58]\n"
- "smlal v0.4s, v21.4h, v29.4h\n"
- "smlal v6.4s, v27.4h, v17.4h\n"
+ "ldr q26, [x12, #0x0]\n"
+ "ldr q6, [x11, #0x0]\n"
+ "smlal v20.4s, v11.4h, v25.4h\n"
+ "smlal2 v1.4s, v11.8h, v25.8h\n"
+ "ldr q0, [x12, #0x10]\n"
+ "ldr q5, [x11, #0x10]\n"
+ "smlal v8.4s, v11.4h, v7.4h\n"
+ "smlal v3.4s, v11.4h, v24.4h\n"
+ "ldr x25, [x14, #0x28]\n"
+ "smlal v10.4s, v11.4h, v22.4h\n"
+ "ldr x24, [x14, #0x38]\n"
+ "smlal2 v21.4s, v11.8h, v7.8h\n"
+ "smlal2 v30.4s, v11.8h, v24.8h\n"
+ "smlal2 v27.4s, v11.8h, v22.8h\n"
+ "ldr x23, [x14, #0x30]\n"
+ "ldr x22, [x14, #0x40]\n"
+ "smlal v20.4s, v28.4h, v22.4h\n"
+ "smlal2 v1.4s, v28.8h, v22.8h\n"
+ "ldr x20, [x14, #0x48]\n"
+ "ldr x21, [x14, #0x50]\n"
+ "ldr d16, [x25, x16]\n"
+ "ldr d28, [x24, x16]\n"
+ "smlal v8.4s, v18.4h, v9.4h\n"
+ "smlal v3.4s, v19.4h, v9.4h\n"
+ "ldr d31, [x23, x16]\n"
+ "ldr d11, [x22, x16]\n"
+ "smlal v10.4s, v19.4h, v24.4h\n"
+ "smlal2 v21.4s, v18.8h, v9.8h\n"
+ "ldr d18, [x20, x16]\n"
+ "smlal2 v30.4s, v19.8h, v9.8h\n"
+ "smlal2 v27.4s, v19.8h, v24.8h\n"
+ "ldr x20, [x14, #0x58]\n"
+ "smlal v20.4s, v19.4h, v4.4h\n"
+ "smlal2 v1.4s, v19.8h, v4.8h\n"
+ "ushll v16.8h, v16.8b, #0x0\n"
+ "ldr x24, [x14, #0x60]\n"
+ "ushll v28.8h, v28.8b, #0x0\n"
+ "smlal v8.4s, v19.4h, v25.4h\n"
+ "ldr x23, [x14, #0x68]\n"
+ "ldr x22, [x14, #0x70]\n"
+ "smlal v10.4s, v23.4h, v7.4h\n"
+ "ushll v31.8h, v31.8b, #0x0\n"
+ "smlal2 v21.4s, v19.8h, v25.8h\n"
+ "ldr d19, [x21, x16]\n"
+ "smlal v3.4s, v16.4h, v13.4h\n"
+ "smlal2 v30.4s, v16.8h, v13.8h\n"
+ "ldr d16, [x20, x16]\n"
"ushll v11.8h, v11.8b, #0x0\n"
- "ldr x24, [x12, #0x60]\n"
- "smlal2 v9.4s, v27.8h, v16.8h\n"
- "smlal v28.4s, v10.4h, v7.4h\n"
- "ldr x23, [x12, #0x68]\n"
- "ldr x22, [x12, #0x70]\n"
- "smlal2 v30.4s, v13.8h, v8.8h\n"
- "ldr d13, [x21, x14]\n"
- "smlal2 v22.4s, v21.8h, v29.8h\n"
- "ldr d21, [x20, x14]\n"
- "smlal2 v2.4s, v27.8h, v17.8h\n"
- "smlal v3.4s, v27.4h, v29.4h\n"
- "ushll v13.8h, v13.8b, #0x0\n"
- "ldr x21, [x12, #0x78]\n"
- "smlal v0.4s, v27.4h, v8.4h\n"
- "smlal v6.4s, v15.4h, v4.4h\n"
- "ushll v21.8h, v21.8b, #0x0\n"
+ "smlal v20.4s, v23.4h, v14.4h\n"
+ "smlal2 v1.4s, v23.8h, v14.8h\n"
+ "ushll v18.8h, v18.8b, #0x0\n"
+ "ldr x21, [x14, #0x78]\n"
+ "smlal2 v27.4s, v23.8h, v7.8h\n"
+ "smlal v8.4s, v23.4h, v13.4h\n"
+ "ushll v19.8h, v19.8b, #0x0\n"
"ldr x20, [%x[params], %[offsetof_Params_bias]]\n"
- "smlal2 v9.4s, v10.8h, v7.8h\n"
- "smlal v28.4s, v23.4h, v1.4h\n"
- "add x11, x11, #0x48\n"
- "subs x15, x15, #0x1\n"
- "smlal2 v30.4s, v27.8h, v29.8h\n"
- "smlal2 v22.4s, v27.8h, v8.8h\n"
- "ldr d27, [x24, x14]\n"
- "ushll v27.8h, v27.8b, #0x0\n"
- "smlal2 v2.4s, v15.8h, v4.8h\n"
- "ldr d15, [x23, x14]\n"
- "smlal v3.4s, v10.4h, v19.4h\n"
- "ushll v15.8h, v15.8b, #0x0\n"
- "smlal v0.4s, v11.4h, v31.4h\n"
- "smlal v6.4s, v11.4h, v8.4h\n"
- "add x10, x10, #0x20\n"
- "add x9, x9, #0x20\n"
- "smlal2 v9.4s, v23.8h, v1.8h\n"
- "smlal v28.4s, v11.4h, v4.4h\n"
- "smlal2 v30.4s, v10.8h, v19.8h\n"
- "ldr d10, [x22, x14]\n"
- "smlal2 v22.4s, v11.8h, v31.8h\n"
- "ushll v10.8h, v10.8b, #0x0\n"
- "smlal2 v2.4s, v11.8h, v8.8h\n"
- "ldr d8, [x21, x14]\n"
+ "smlal v10.4s, v31.4h, v2.4h\n"
+ "smlal2 v21.4s, v23.8h, v13.8h\n"
+ "ushll v16.8h, v16.8b, #0x0\n"
+ "add x13, x13, #0x48\n"
+ "smlal v3.4s, v23.4h, v25.4h\n"
+ "smlal2 v30.4s, v23.8h, v25.8h\n"
+ "ldr d23, [x24, x16]\n"
+ "subs x10, x10, #0x1\n"
+ "smlal v20.4s, v28.4h, v24.4h\n"
+ "smlal2 v1.4s, v28.8h, v24.8h\n"
+ "add x12, x12, #0x20\n"
+ "add x11, x11, #0x20\n"
+ "smlal2 v27.4s, v31.8h, v2.8h\n"
+ "ldr d31, [x23, x16]\n"
+ "smlal v8.4s, v28.4h, v22.4h\n"
+ "smlal v10.4s, v18.4h, v25.4h\n"
+ "smlal2 v21.4s, v28.8h, v22.8h\n"
+ "ldr d28, [x22, x16]\n"
+ "ushll v23.8h, v23.8b, #0x0\n"
+ "smlal v3.4s, v18.4h, v4.4h\n"
+ "smlal2 v30.4s, v18.8h, v4.8h\n"
+ "smlal v20.4s, v11.4h, v9.4h\n"
+ "smlal2 v1.4s, v11.8h, v9.8h\n"
+ "ushll v31.8h, v31.8b, #0x0\n"
+ "smlal2 v27.4s, v18.8h, v25.8h\n"
+ "ldr d25, [x21, x16]\n"
+ "smlal v8.4s, v11.4h, v24.4h\n"
+ "ushll v28.8h, v28.8b, #0x0\n"
+ "smlal v10.4s, v16.4h, v9.4h\n"
+ "smlal2 v21.4s, v11.8h, v24.8h\n"
+ "add x16, x16, #0x8\n"
+ "smlal v3.4s, v19.4h, v22.4h\n"
+ "smlal2 v30.4s, v19.8h, v22.8h\n"
+ "smlal v20.4s, v18.4h, v2.4h\n"
+ "smlal2 v1.4s, v18.8h, v2.8h\n"
+ "ushll v25.8h, v25.8b, #0x0\n"
+ "smlal2 v27.4s, v16.8h, v9.8h\n"
+ "smlal v8.4s, v18.4h, v14.4h\n"
+ "smlal v10.4s, v31.4h, v4.4h\n"
+ "smlal2 v21.4s, v18.8h, v14.8h\n"
"smlal v3.4s, v23.4h, v7.4h\n"
- "ushll v8.8h, v8.8b, #0x0\n"
- "smlal v0.4s, v13.4h, v19.4h\n"
- "smlal v6.4s, v21.4h, v1.4h\n"
- "add x14, x14, #0x8\n"
- "smlal2 v9.4s, v11.8h, v4.8h\n"
- "smlal v28.4s, v13.4h, v17.4h\n"
"smlal2 v30.4s, v23.8h, v7.8h\n"
- "smlal2 v22.4s, v13.8h, v19.8h\n"
- "smlal2 v2.4s, v21.8h, v1.8h\n"
- "smlal v3.4s, v11.4h, v16.4h\n"
- "smlal v0.4s, v27.4h, v17.4h\n"
- "smlal v6.4s, v15.4h, v31.4h\n"
- "smlal2 v9.4s, v13.8h, v17.8h\n"
- "smlal v28.4s, v27.4h, v29.4h\n"
- "sqrdmulh v28.4s, v28.4s, v24.4s\n"
- "smlal2 v30.4s, v11.8h, v16.8h\n"
- "smlal2 v22.4s, v27.8h, v17.8h\n"
- "and v17.16b, v28.16b, v25.16b\n"
- "smlal2 v2.4s, v15.8h, v31.8h\n"
- "smlal v3.4s, v21.4h, v31.4h\n"
- "sshr v17.4s, v17.4s, #0x1f\n"
- "smlal v0.4s, v10.4h, v16.4h\n"
- "smlal v6.4s, v10.4h, v29.4h\n"
- "sqadd v28.4s, v28.4s, v17.4s\n"
- "smlal2 v9.4s, v27.8h, v29.8h\n"
- "smlal2 v30.4s, v21.8h, v31.8h\n"
- "sqrdmulh v9.4s, v9.4s, v20.4s\n"
- "smlal2 v22.4s, v10.8h, v16.8h\n"
- "smlal2 v2.4s, v10.8h, v29.8h\n"
- "and v23.16b, v9.16b, v26.16b\n"
- "smlal v3.4s, v15.4h, v4.4h\n"
- "smlal v0.4s, v8.4h, v4.4h\n"
- "sqrdmulh v3.4s, v3.4s, v24.4s\n"
- "smlal v6.4s, v8.4h, v16.4h\n"
- "smlal2 v30.4s, v15.8h, v4.8h\n"
- "sqrdmulh v0.4s, v0.4s, v24.4s\n"
- "smlal2 v22.4s, v8.8h, v4.8h\n"
- "smlal2 v2.4s, v8.8h, v16.8h\n"
- "sqrdmulh v6.4s, v6.4s, v24.4s\n"
- "sshr v23.4s, v23.4s, #0x1f\n"
- "and v8.16b, v3.16b, v25.16b\n"
- "sqrdmulh v30.4s, v30.4s, v20.4s\n"
- "and v11.16b, v0.16b, v25.16b\n"
- "sqrdmulh v22.4s, v22.4s, v20.4s\n"
- "and v29.16b, v6.16b, v25.16b\n"
- "sqrdmulh v2.4s, v2.4s, v20.4s\n"
- "sqadd v9.4s, v9.4s, v23.4s\n"
- "sshr v8.4s, v8.4s, #0x1f\n"
- "and v13.16b, v30.16b, v26.16b\n"
- "sshr v11.4s, v11.4s, #0x1f\n"
- "and v21.16b, v22.16b, v26.16b\n"
- "sshr v29.4s, v29.4s, #0x1f\n"
- "and v23.16b, v2.16b, v26.16b\n"
- "sqadd v3.4s, v3.4s, v8.4s\n"
+ "smlal v20.4s, v19.4h, v7.4h\n"
+ "smlal2 v1.4s, v19.8h, v7.8h\n"
+ "smlal2 v27.4s, v31.8h, v4.8h\n"
+ "smlal v8.4s, v16.4h, v4.4h\n"
+ "smlal v10.4s, v28.4h, v13.4h\n"
+ "smlal2 v21.4s, v16.8h, v4.8h\n"
+ "smlal v3.4s, v28.4h, v14.4h\n"
+ "smlal2 v30.4s, v28.8h, v14.8h\n"
+ "smlal v20.4s, v23.4h, v13.4h\n"
+ "smlal2 v1.4s, v23.8h, v13.8h\n"
+ "smlal2 v27.4s, v28.8h, v13.8h\n"
+ "smlal v8.4s, v31.4h, v2.4h\n"
+ "smlal v10.4s, v25.4h, v14.4h\n"
+ "smlal2 v21.4s, v31.8h, v2.8h\n"
+ "smlal v3.4s, v25.4h, v2.4h\n"
+ "smlal2 v30.4s, v25.8h, v2.8h\n"
+ "sqrdmulh v20.4s, v20.4s, v26.4s\n"
+ "sqrdmulh v1.4s, v1.4s, v0.4s\n"
+ "smlal2 v27.4s, v25.8h, v14.8h\n"
+ "sqrdmulh v8.4s, v8.4s, v26.4s\n"
+ "and v19.16b, v20.16b, v6.16b\n"
+ "sqrdmulh v10.4s, v10.4s, v26.4s\n"
+ "sqrdmulh v21.4s, v21.4s, v0.4s\n"
+ "and v9.16b, v1.16b, v5.16b\n"
+ "sqrdmulh v3.4s, v3.4s, v26.4s\n"
+ "sshr v19.4s, v19.4s, #0x1f\n"
+ "and v16.16b, v8.16b, v6.16b\n"
+ "sqrdmulh v30.4s, v30.4s, v0.4s\n"
+ "sqrdmulh v27.4s, v27.4s, v0.4s\n"
+ "sshr v9.4s, v9.4s, #0x1f\n"
+ "and v22.16b, v3.16b, v6.16b\n"
+ "sqadd v20.4s, v20.4s, v19.4s\n"
+ "and v13.16b, v10.16b, v6.16b\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "and v28.16b, v21.16b, v5.16b\n"
+ "sqadd v1.4s, v1.4s, v9.4s\n"
+ "sshr v22.4s, v22.4s, #0x1f\n"
+ "and v0.16b, v30.16b, v5.16b\n"
"sshr v13.4s, v13.4s, #0x1f\n"
- "sqadd v0.4s, v0.4s, v11.4s\n"
- "sshr v21.4s, v21.4s, #0x1f\n"
- "sqadd v6.4s, v6.4s, v29.4s\n"
- "sshr v23.4s, v23.4s, #0x1f\n"
- "srshl v28.4s, v28.4s, v25.4s\n"
- "srshl v3.4s, v3.4s, v25.4s\n"
- "sqadd v30.4s, v30.4s, v13.4s\n"
- "srshl v0.4s, v0.4s, v25.4s\n"
- "sqadd v22.4s, v22.4s, v21.4s\n"
- "srshl v6.4s, v6.4s, v25.4s\n"
- "sqadd v2.4s, v2.4s, v23.4s\n"
- "srshl v9.4s, v9.4s, v26.4s\n"
- "sqxtn v28.4h, v28.4s\n"
- "srshl v30.4s, v30.4s, v26.4s\n"
+ "and v18.16b, v27.16b, v5.16b\n"
+ "sqadd v8.4s, v8.4s, v16.4s\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
+ "sqadd v3.4s, v3.4s, v22.4s\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqadd v10.4s, v10.4s, v13.4s\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "srshl v20.4s, v20.4s, v6.4s\n"
+ "srshl v8.4s, v8.4s, v6.4s\n"
+ "sqadd v21.4s, v21.4s, v28.4s\n"
+ "srshl v3.4s, v3.4s, v6.4s\n"
+ "sqadd v30.4s, v30.4s, v0.4s\n"
+ "srshl v10.4s, v10.4s, v6.4s\n"
+ "sqadd v27.4s, v27.4s, v18.4s\n"
+ "srshl v1.4s, v1.4s, v5.4s\n"
+ "sqxtn v20.4h, v20.4s\n"
+ "srshl v21.4s, v21.4s, v5.4s\n"
+ "sqxtn v8.4h, v8.4s\n"
+ "srshl v30.4s, v30.4s, v5.4s\n"
"sqxtn v3.4h, v3.4s\n"
- "srshl v22.4s, v22.4s, v26.4s\n"
- "sqxtn v0.4h, v0.4s\n"
- "srshl v2.4s, v2.4s, v26.4s\n"
- "sqxtn v6.4h, v6.4s\n"
- "sqxtn2 v28.8h, v9.4s\n"
+ "srshl v27.4s, v27.4s, v5.4s\n"
+ "sqxtn v10.4h, v10.4s\n"
+ "sqxtn2 v20.8h, v1.4s\n"
+ "sqxtn2 v8.8h, v21.4s\n"
"sqxtn2 v3.8h, v30.4s\n"
- "sqxtn2 v0.8h, v22.4s\n"
- "sqxtn2 v6.8h, v2.4s\n"
- "sqadd v28.8h, v28.8h, v5.8h\n"
- "sqadd v3.8h, v3.8h, v5.8h\n"
- "sqadd v0.8h, v0.8h, v5.8h\n"
- "sqadd v6.8h, v6.8h, v5.8h\n"
- "smax v28.8h, v28.8h, v14.8h\n"
- "smax v3.8h, v3.8h, v14.8h\n"
- "smax v0.8h, v0.8h, v14.8h\n"
- "smax v6.8h, v6.8h, v14.8h\n"
- "smin v28.8h, v28.8h, v12.8h\n"
- "smin v3.8h, v3.8h, v12.8h\n"
- "smin v0.8h, v0.8h, v12.8h\n"
- "smin v6.8h, v6.8h, v12.8h\n"
- "uzp1 v28.16b, v28.16b, v28.16b\n"
- "str d28, [x28, x13]\n"
+ "sqxtn2 v10.8h, v27.4s\n"
+ "sqadd v20.8h, v20.8h, v12.8h\n"
+ "sqadd v8.8h, v8.8h, v12.8h\n"
+ "sqadd v3.8h, v3.8h, v12.8h\n"
+ "sqadd v10.8h, v10.8h, v12.8h\n"
+ "smax v20.8h, v20.8h, v15.8h\n"
+ "smax v8.8h, v8.8h, v15.8h\n"
+ "smax v3.8h, v3.8h, v15.8h\n"
+ "smax v10.8h, v10.8h, v15.8h\n"
+ "smin v20.8h, v20.8h, v17.8h\n"
+ "smin v8.8h, v8.8h, v17.8h\n"
+ "smin v3.8h, v3.8h, v17.8h\n"
+ "smin v10.8h, v10.8h, v17.8h\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "uzp1 v8.16b, v8.16b, v8.16b\n"
"uzp1 v3.16b, v3.16b, v3.16b\n"
- "uzp1 v0.16b, v0.16b, v0.16b\n"
- "str d3, [x27, x13]\n"
- "uzp1 v6.16b, v6.16b, v6.16b\n"
- "str d0, [x26, x13]\n"
- "str d6, [x25, x13]\n"
- "ldr q28, [x20, #0x0]\n"
- "ldr q9, [x20, #0x10]\n"
+ "uzp1 v10.16b, v10.16b, v10.16b\n"
+ "str d20, [x9, x15]\n"
+ "str d8, [x28, x15]\n"
+ "str d3, [x27, x15]\n"
+ "str d10, [x26, x15]\n"
+ "add x15, x15, #0x8\n"
+ "ldr q20, [x20, #0x0]\n"
+ "ldr q1, [x20, #0x10]\n"
"add x20, x20, #0x20\n"
- "ldr d19, [x11, #0x0]\n"
- "ldr d7, [x11, #0x8]\n"
- "add x13, x13, #0x8\n"
"str x20, [%x[params], %[offsetof_Params_bias]]\n"
- "ldr d1, [x11, #0x10]\n"
- "ldr d17, [x11, #0x18]\n"
- "mov v3.16b, v28.16b\n"
- "mov v30.16b, v9.16b\n"
- "ldr d8, [x11, #0x20]\n"
- "ldr d31, [x11, #0x28]\n"
- "mov v0.16b, v28.16b\n"
- "mov v22.16b, v9.16b\n"
- "ldr d29, [x11, #0x30]\n"
- "ldr d16, [x11, #0x38]\n"
- "mov v6.16b, v28.16b\n"
- "mov v2.16b, v9.16b\n"
- "ldr d4, [x11, #0x40]\n"
- "ldp x23, x22, [x12, #0x0]\n"
- "usubl v19.8h, v19.8b, v18.8b\n"
- "usubl v7.8h, v7.8b, v18.8b\n"
- "ldp x21, x20, [x12, #0x10]\n"
- "ldr d23, [x23, x14]\n"
- "usubl v1.8h, v1.8b, v18.8b\n"
- "usubl v17.8h, v17.8b, v18.8b\n"
- "ldr d10, [x22, x14]\n"
- "ldr d11, [x21, x14]\n"
- "usubl v8.8h, v8.8b, v18.8b\n"
- "usubl v31.8h, v31.8b, v18.8b\n"
- "ldr d13, [x20, x14]\n"
- "ldr x20, [x12, #0x20]\n"
- "usubl v29.8h, v29.8b, v18.8b\n"
- "usubl v16.8h, v16.8b, v18.8b\n"
- "ldr d27, [x20, x14]\n"
- "usubl v4.8h, v4.8b, v18.8b\n"
- "ushll v23.8h, v23.8b, #0x0\n"
- "ushll v10.8h, v10.8b, #0x0\n"
+ "ldr d22, [x13, #0x0]\n"
+ "ldr d24, [x13, #0x8]\n"
+ "ldr d9, [x13, #0x10]\n"
+ "ldr d7, [x13, #0x18]\n"
+ "ldr d25, [x13, #0x20]\n"
+ "ldr d4, [x13, #0x28]\n"
+ "mov v8.16b, v20.16b\n"
+ "mov v21.16b, v1.16b\n"
+ "ldr d13, [x13, #0x30]\n"
+ "ldr d14, [x13, #0x38]\n"
+ "mov v3.16b, v20.16b\n"
+ "mov v30.16b, v1.16b\n"
+ "ldr d2, [x13, #0x40]\n"
+ "ldp x23, x22, [x14, #0x0]\n"
+ "mov v10.16b, v20.16b\n"
+ "mov v27.16b, v1.16b\n"
+ "usubl v22.8h, v22.8b, v29.8b\n"
+ "usubl v24.8h, v24.8b, v29.8b\n"
+ "usubl v9.8h, v9.8b, v29.8b\n"
+ "usubl v7.8h, v7.8b, v29.8b\n"
+ "ldp x21, x20, [x14, #0x10]\n"
+ "usubl v25.8h, v25.8b, v29.8b\n"
+ "usubl v4.8h, v4.8b, v29.8b\n"
+ "usubl v13.8h, v13.8b, v29.8b\n"
+ "usubl v14.8h, v14.8b, v29.8b\n"
+ "ldr d11, [x23, x16]\n"
+ "ldr d28, [x22, x16]\n"
+ "ldr d18, [x21, x16]\n"
+ "ldr d19, [x20, x16]\n"
+ "usubl v2.8h, v2.8b, v29.8b\n"
+ "ldr x20, [x14, #0x20]\n"
"ushll v11.8h, v11.8b, #0x0\n"
- "ushll v13.8h, v13.8b, #0x0\n"
- "ushll v27.8h, v27.8b, #0x0\n"
+ "ushll v28.8h, v28.8b, #0x0\n"
+ "ushll v18.8h, v18.8b, #0x0\n"
+ "ushll v19.8h, v19.8b, #0x0\n"
+ "ldr d23, [x20, x16]\n"
+ "ushll v23.8h, v23.8b, #0x0\n"
"bgt 1b\n"
"2:" // Tail
- "ldr q26, [x10, #0x0]\n"
- "ldr q25, [x9, #0x0]\n"
- "smlal v28.4s, v23.4h, v8.4h\n"
- "smlal2 v9.4s, v23.8h, v8.8h\n"
- "ldr q24, [x10, #0x10]\n"
- "ldr q20, [x9, #0x10]\n"
- "smlal v28.4s, v10.4h, v19.4h\n"
- "smlal v3.4s, v23.4h, v17.4h\n"
- "ldr x20, [x12, #0x28]\n"
- "ldr d21, [x20, x14]\n"
- "smlal v0.4s, v23.4h, v7.4h\n"
- "smlal v6.4s, v23.4h, v19.4h\n"
- "smlal2 v9.4s, v10.8h, v19.8h\n"
- "ldr x20, [x12, #0x38]\n"
- "ldr d15, [x20, x14]\n"
- "smlal v28.4s, v13.4h, v31.4h\n"
- "smlal2 v30.4s, v23.8h, v17.8h\n"
- "smlal2 v22.4s, v23.8h, v7.8h\n"
- "ldr x20, [x12, #0x30]\n"
- "ldr d10, [x20, x14]\n"
- "smlal2 v2.4s, v23.8h, v19.8h\n"
- "smlal v3.4s, v11.4h, v1.4h\n"
- "ushll v21.8h, v21.8b, #0x0\n"
- "ldr x20, [x12, #0x40]\n"
- "ldr d23, [x20, x14]\n"
- "smlal v0.4s, v13.4h, v1.4h\n"
- "smlal v6.4s, v13.4h, v7.4h\n"
- "ushll v15.8h, v15.8b, #0x0\n"
- "smlal2 v9.4s, v13.8h, v31.8h\n"
- "smlal v28.4s, v27.4h, v16.4h\n"
- "ldr x20, [x12, #0x48]\n"
- "ushll v10.8h, v10.8b, #0x0\n"
- "smlal2 v30.4s, v11.8h, v1.8h\n"
- "ldr d11, [x20, x14]\n"
- "smlal2 v22.4s, v13.8h, v1.8h\n"
+ "ldr q16, [x12, #0x0]\n"
+ "ldr q5, [x11, #0x0]\n"
+ "smlal v20.4s, v11.4h, v25.4h\n"
+ "smlal2 v1.4s, v11.8h, v25.8h\n"
+ "ldr q6, [x12, #0x10]\n"
+ "ldr q31, [x11, #0x10]\n"
+ "smlal v8.4s, v11.4h, v7.4h\n"
+ "smlal v3.4s, v11.4h, v24.4h\n"
+ "ldr x25, [x14, #0x28]\n"
+ "smlal v10.4s, v11.4h, v22.4h\n"
+ "ldr x23, [x14, #0x38]\n"
+ "smlal2 v21.4s, v11.8h, v7.8h\n"
+ "smlal2 v30.4s, v11.8h, v24.8h\n"
+ "smlal2 v27.4s, v11.8h, v22.8h\n"
+ "ldr x22, [x14, #0x30]\n"
+ "ldr x21, [x14, #0x40]\n"
+ "smlal v20.4s, v28.4h, v22.4h\n"
+ "smlal2 v1.4s, v28.8h, v22.8h\n"
+ "ldr x20, [x14, #0x48]\n"
+ "ldr x24, [x14, #0x50]\n"
+ "ldr d28, [x25, x16]\n"
+ "ldr d26, [x23, x16]\n"
+ "smlal v8.4s, v18.4h, v9.4h\n"
+ "smlal v3.4s, v19.4h, v9.4h\n"
+ "ldr d11, [x22, x16]\n"
+ "ldr d0, [x21, x16]\n"
+ "smlal v10.4s, v19.4h, v24.4h\n"
+ "smlal2 v21.4s, v18.8h, v9.8h\n"
+ "ldr d18, [x20, x16]\n"
+ "smlal2 v30.4s, v19.8h, v9.8h\n"
+ "smlal2 v27.4s, v19.8h, v24.8h\n"
+ "ldr x20, [x14, #0x58]\n"
+ "smlal v20.4s, v19.4h, v4.4h\n"
+ "smlal2 v1.4s, v19.8h, v4.8h\n"
+ "ushll v28.8h, v28.8b, #0x0\n"
+ "ldr x23, [x14, #0x60]\n"
+ "ushll v26.8h, v26.8b, #0x0\n"
+ "smlal v8.4s, v19.4h, v25.4h\n"
+ "ldr x22, [x14, #0x68]\n"
+ "ldr x21, [x14, #0x70]\n"
+ "smlal v10.4s, v23.4h, v7.4h\n"
+ "ushll v11.8h, v11.8b, #0x0\n"
+ "smlal2 v21.4s, v19.8h, v25.8h\n"
+ "ldr d19, [x24, x16]\n"
+ "smlal v3.4s, v28.4h, v13.4h\n"
+ "smlal2 v30.4s, v28.8h, v13.8h\n"
+ "ldr d28, [x20, x16]\n"
+ "ushll v0.8h, v0.8b, #0x0\n"
+ "smlal v20.4s, v23.4h, v14.4h\n"
+ "smlal2 v1.4s, v23.8h, v14.8h\n"
+ "ushll v18.8h, v18.8b, #0x0\n"
+ "ldr x20, [x14, #0x78]\n"
+ "smlal2 v27.4s, v23.8h, v7.8h\n"
+ "smlal v8.4s, v23.4h, v13.4h\n"
+ "ushll v19.8h, v19.8b, #0x0\n"
+ "tst x17, #0x7\n"
+ "smlal v10.4s, v11.4h, v2.4h\n"
+ "smlal2 v21.4s, v23.8h, v13.8h\n"
+ "ushll v28.8h, v28.8b, #0x0\n"
+ "add x12, x12, #0x20\n"
+ "smlal v3.4s, v23.4h, v25.4h\n"
+ "smlal2 v30.4s, v23.8h, v25.8h\n"
+ "ldr d23, [x23, x16]\n"
+ "add x11, x11, #0x20\n"
+ "smlal v20.4s, v26.4h, v24.4h\n"
+ "smlal2 v1.4s, v26.8h, v24.8h\n"
+ "smlal2 v27.4s, v11.8h, v2.8h\n"
+ "ldr d11, [x22, x16]\n"
+ "smlal v8.4s, v26.4h, v22.4h\n"
+ "smlal v10.4s, v18.4h, v25.4h\n"
+ "smlal2 v21.4s, v26.8h, v22.8h\n"
+ "ldr d26, [x21, x16]\n"
"ushll v23.8h, v23.8b, #0x0\n"
- "smlal2 v2.4s, v13.8h, v7.8h\n"
- "smlal v3.4s, v13.4h, v8.4h\n"
- "ldr x24, [x12, #0x50]\n"
- "ldr x20, [x12, #0x58]\n"
- "smlal v0.4s, v21.4h, v29.4h\n"
- "smlal v6.4s, v27.4h, v17.4h\n"
+ "smlal v3.4s, v18.4h, v4.4h\n"
+ "smlal2 v30.4s, v18.8h, v4.8h\n"
+ "smlal v20.4s, v0.4h, v9.4h\n"
+ "smlal2 v1.4s, v0.8h, v9.8h\n"
"ushll v11.8h, v11.8b, #0x0\n"
- "ldr x23, [x12, #0x60]\n"
- "smlal2 v9.4s, v27.8h, v16.8h\n"
- "smlal v28.4s, v15.4h, v7.4h\n"
- "ldr x22, [x12, #0x68]\n"
- "ldr x21, [x12, #0x70]\n"
- "smlal2 v30.4s, v13.8h, v8.8h\n"
- "ldr d13, [x24, x14]\n"
- "smlal2 v22.4s, v21.8h, v29.8h\n"
- "ldr d21, [x20, x14]\n"
- "smlal2 v2.4s, v27.8h, v17.8h\n"
- "smlal v3.4s, v27.4h, v29.4h\n"
- "ushll v13.8h, v13.8b, #0x0\n"
- "ldr x20, [x12, #0x78]\n"
- "smlal v0.4s, v27.4h, v8.4h\n"
- "smlal v6.4s, v10.4h, v4.4h\n"
- "ushll v21.8h, v21.8b, #0x0\n"
- "tst x16, #0x7\n"
- "smlal2 v9.4s, v15.8h, v7.8h\n"
- "smlal v28.4s, v23.4h, v1.4h\n"
- "add x10, x10, #0x20\n"
- "add x9, x9, #0x20\n"
- "smlal2 v30.4s, v27.8h, v29.8h\n"
- "smlal2 v22.4s, v27.8h, v8.8h\n"
- "ldr d27, [x23, x14]\n"
- "ushll v27.8h, v27.8b, #0x0\n"
- "smlal2 v2.4s, v10.8h, v4.8h\n"
- "ldr d10, [x22, x14]\n"
- "smlal v3.4s, v15.4h, v19.4h\n"
- "ushll v10.8h, v10.8b, #0x0\n"
- "smlal v0.4s, v11.4h, v31.4h\n"
- "smlal v6.4s, v11.4h, v8.4h\n"
- "smlal2 v9.4s, v23.8h, v1.8h\n"
- "smlal v28.4s, v11.4h, v4.4h\n"
- "smlal2 v30.4s, v15.8h, v19.8h\n"
- "ldr d15, [x21, x14]\n"
- "smlal2 v22.4s, v11.8h, v31.8h\n"
- "ushll v15.8h, v15.8b, #0x0\n"
- "smlal2 v2.4s, v11.8h, v8.8h\n"
- "ldr d8, [x20, x14]\n"
+ "smlal2 v27.4s, v18.8h, v25.8h\n"
+ "ldr d25, [x20, x16]\n"
+ "smlal v8.4s, v0.4h, v24.4h\n"
+ "ushll v26.8h, v26.8b, #0x0\n"
+ "smlal v10.4s, v28.4h, v9.4h\n"
+ "smlal2 v21.4s, v0.8h, v24.8h\n"
+ "add x16, x16, #0x8\n"
+ "smlal v3.4s, v19.4h, v22.4h\n"
+ "smlal2 v30.4s, v19.8h, v22.8h\n"
+ "smlal v20.4s, v18.4h, v2.4h\n"
+ "smlal2 v1.4s, v18.8h, v2.8h\n"
+ "ushll v25.8h, v25.8b, #0x0\n"
+ "smlal2 v27.4s, v28.8h, v9.8h\n"
+ "smlal v8.4s, v18.4h, v14.4h\n"
+ "smlal v10.4s, v11.4h, v4.4h\n"
+ "smlal2 v21.4s, v18.8h, v14.8h\n"
"smlal v3.4s, v23.4h, v7.4h\n"
- "ushll v8.8h, v8.8b, #0x0\n"
- "smlal v0.4s, v13.4h, v19.4h\n"
- "smlal v6.4s, v21.4h, v1.4h\n"
- "add x14, x14, #0x8\n"
- "smlal2 v9.4s, v11.8h, v4.8h\n"
- "smlal v28.4s, v13.4h, v17.4h\n"
"smlal2 v30.4s, v23.8h, v7.8h\n"
- "smlal2 v22.4s, v13.8h, v19.8h\n"
- "smlal2 v2.4s, v21.8h, v1.8h\n"
- "smlal v3.4s, v11.4h, v16.4h\n"
- "smlal v0.4s, v27.4h, v17.4h\n"
- "smlal v6.4s, v10.4h, v31.4h\n"
- "smlal2 v9.4s, v13.8h, v17.8h\n"
- "smlal v28.4s, v27.4h, v29.4h\n"
- "sqrdmulh v28.4s, v28.4s, v26.4s\n"
- "smlal2 v30.4s, v11.8h, v16.8h\n"
- "smlal2 v22.4s, v27.8h, v17.8h\n"
- "and v1.16b, v28.16b, v25.16b\n"
- "smlal2 v2.4s, v10.8h, v31.8h\n"
- "smlal v3.4s, v21.4h, v31.4h\n"
- "sshr v1.4s, v1.4s, #0x1f\n"
- "smlal v0.4s, v15.4h, v16.4h\n"
- "smlal v6.4s, v15.4h, v29.4h\n"
- "sqadd v28.4s, v28.4s, v1.4s\n"
- "smlal2 v9.4s, v27.8h, v29.8h\n"
- "smlal2 v30.4s, v21.8h, v31.8h\n"
- "sqrdmulh v9.4s, v9.4s, v24.4s\n"
- "smlal2 v22.4s, v15.8h, v16.8h\n"
- "smlal2 v2.4s, v15.8h, v29.8h\n"
- "and v27.16b, v9.16b, v20.16b\n"
- "smlal v3.4s, v10.4h, v4.4h\n"
- "smlal v0.4s, v8.4h, v4.4h\n"
- "sqrdmulh v3.4s, v3.4s, v26.4s\n"
- "smlal v6.4s, v8.4h, v16.4h\n"
- "smlal2 v30.4s, v10.8h, v4.8h\n"
- "sqrdmulh v0.4s, v0.4s, v26.4s\n"
- "smlal2 v22.4s, v8.8h, v4.8h\n"
- "smlal2 v2.4s, v8.8h, v16.8h\n"
- "sqrdmulh v6.4s, v6.4s, v26.4s\n"
- "sshr v27.4s, v27.4s, #0x1f\n"
- "and v16.16b, v3.16b, v25.16b\n"
- "sqrdmulh v30.4s, v30.4s, v24.4s\n"
- "and v4.16b, v0.16b, v25.16b\n"
- "sqrdmulh v22.4s, v22.4s, v24.4s\n"
- "and v17.16b, v6.16b, v25.16b\n"
- "sqrdmulh v2.4s, v2.4s, v24.4s\n"
- "sqadd v9.4s, v9.4s, v27.4s\n"
- "sshr v16.4s, v16.4s, #0x1f\n"
- "and v8.16b, v30.16b, v20.16b\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "and v26.16b, v22.16b, v20.16b\n"
- "sshr v17.4s, v17.4s, #0x1f\n"
- "and v11.16b, v2.16b, v20.16b\n"
- "sqadd v3.4s, v3.4s, v16.4s\n"
- "sshr v8.4s, v8.4s, #0x1f\n"
- "sqadd v0.4s, v0.4s, v4.4s\n"
+ "smlal v20.4s, v19.4h, v7.4h\n"
+ "smlal2 v1.4s, v19.8h, v7.8h\n"
+ "smlal2 v27.4s, v11.8h, v4.8h\n"
+ "smlal v8.4s, v28.4h, v4.4h\n"
+ "smlal v10.4s, v26.4h, v13.4h\n"
+ "smlal2 v21.4s, v28.8h, v4.8h\n"
+ "smlal v3.4s, v26.4h, v14.4h\n"
+ "smlal2 v30.4s, v26.8h, v14.8h\n"
+ "smlal v20.4s, v23.4h, v13.4h\n"
+ "smlal2 v1.4s, v23.8h, v13.8h\n"
+ "smlal2 v27.4s, v26.8h, v13.8h\n"
+ "smlal v8.4s, v11.4h, v2.4h\n"
+ "smlal v10.4s, v25.4h, v14.4h\n"
+ "smlal2 v21.4s, v11.8h, v2.8h\n"
+ "smlal v3.4s, v25.4h, v2.4h\n"
+ "smlal2 v30.4s, v25.8h, v2.8h\n"
+ "sqrdmulh v20.4s, v20.4s, v16.4s\n"
+ "sqrdmulh v1.4s, v1.4s, v6.4s\n"
+ "smlal2 v27.4s, v25.8h, v14.8h\n"
+ "sqrdmulh v8.4s, v8.4s, v16.4s\n"
+ "and v0.16b, v20.16b, v5.16b\n"
+ "sqrdmulh v10.4s, v10.4s, v16.4s\n"
+ "sqrdmulh v21.4s, v21.4s, v6.4s\n"
+ "and v24.16b, v1.16b, v31.16b\n"
+ "sqrdmulh v3.4s, v3.4s, v16.4s\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "and v22.16b, v8.16b, v5.16b\n"
+ "sqrdmulh v30.4s, v30.4s, v6.4s\n"
+ "sqrdmulh v27.4s, v27.4s, v6.4s\n"
+ "sshr v24.4s, v24.4s, #0x1f\n"
+ "and v26.16b, v3.16b, v5.16b\n"
+ "sqadd v20.4s, v20.4s, v0.4s\n"
+ "and v16.16b, v10.16b, v5.16b\n"
+ "sshr v22.4s, v22.4s, #0x1f\n"
+ "and v23.16b, v21.16b, v31.16b\n"
+ "sqadd v1.4s, v1.4s, v24.4s\n"
"sshr v26.4s, v26.4s, #0x1f\n"
- "sqadd v6.4s, v6.4s, v17.4s\n"
- "sshr v11.4s, v11.4s, #0x1f\n"
- "srshl v28.4s, v28.4s, v25.4s\n"
- "srshl v3.4s, v3.4s, v25.4s\n"
- "sqadd v30.4s, v30.4s, v8.4s\n"
- "srshl v0.4s, v0.4s, v25.4s\n"
- "sqadd v22.4s, v22.4s, v26.4s\n"
- "srshl v6.4s, v6.4s, v25.4s\n"
- "sqadd v2.4s, v2.4s, v11.4s\n"
- "srshl v9.4s, v9.4s, v20.4s\n"
- "sqxtn v28.4h, v28.4s\n"
- "srshl v30.4s, v30.4s, v20.4s\n"
+ "and v7.16b, v30.16b, v31.16b\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "and v9.16b, v27.16b, v31.16b\n"
+ "sqadd v8.4s, v8.4s, v22.4s\n"
+ "sshr v23.4s, v23.4s, #0x1f\n"
+ "sqadd v3.4s, v3.4s, v26.4s\n"
+ "sshr v7.4s, v7.4s, #0x1f\n"
+ "sqadd v10.4s, v10.4s, v16.4s\n"
+ "sshr v9.4s, v9.4s, #0x1f\n"
+ "srshl v20.4s, v20.4s, v5.4s\n"
+ "srshl v8.4s, v8.4s, v5.4s\n"
+ "sqadd v21.4s, v21.4s, v23.4s\n"
+ "srshl v3.4s, v3.4s, v5.4s\n"
+ "sqadd v30.4s, v30.4s, v7.4s\n"
+ "srshl v10.4s, v10.4s, v5.4s\n"
+ "sqadd v27.4s, v27.4s, v9.4s\n"
+ "srshl v1.4s, v1.4s, v31.4s\n"
+ "sqxtn v20.4h, v20.4s\n"
+ "srshl v21.4s, v21.4s, v31.4s\n"
+ "sqxtn v8.4h, v8.4s\n"
+ "srshl v30.4s, v30.4s, v31.4s\n"
"sqxtn v3.4h, v3.4s\n"
- "srshl v22.4s, v22.4s, v20.4s\n"
- "sqxtn v0.4h, v0.4s\n"
- "srshl v2.4s, v2.4s, v20.4s\n"
- "sqxtn v6.4h, v6.4s\n"
- "sqxtn2 v28.8h, v9.4s\n"
+ "srshl v27.4s, v27.4s, v31.4s\n"
+ "sqxtn v10.4h, v10.4s\n"
+ "sqxtn2 v20.8h, v1.4s\n"
+ "sqxtn2 v8.8h, v21.4s\n"
"sqxtn2 v3.8h, v30.4s\n"
- "sqxtn2 v0.8h, v22.4s\n"
- "sqxtn2 v6.8h, v2.4s\n"
- "sqadd v28.8h, v28.8h, v5.8h\n"
- "sqadd v3.8h, v3.8h, v5.8h\n"
- "sqadd v0.8h, v0.8h, v5.8h\n"
- "sqadd v6.8h, v6.8h, v5.8h\n"
- "smax v28.8h, v28.8h, v14.8h\n"
- "smax v3.8h, v3.8h, v14.8h\n"
- "smax v0.8h, v0.8h, v14.8h\n"
- "smax v6.8h, v6.8h, v14.8h\n"
- "smin v28.8h, v28.8h, v12.8h\n"
- "smin v3.8h, v3.8h, v12.8h\n"
- "smin v0.8h, v0.8h, v12.8h\n"
- "smin v6.8h, v6.8h, v12.8h\n"
- "uzp1 v28.16b, v28.16b, v28.16b\n"
- "str d28, [x28, x13]\n"
+ "sqxtn2 v10.8h, v27.4s\n"
+ "sqadd v20.8h, v20.8h, v12.8h\n"
+ "sqadd v8.8h, v8.8h, v12.8h\n"
+ "sqadd v3.8h, v3.8h, v12.8h\n"
+ "sqadd v10.8h, v10.8h, v12.8h\n"
+ "smax v20.8h, v20.8h, v15.8h\n"
+ "smax v8.8h, v8.8h, v15.8h\n"
+ "smax v3.8h, v3.8h, v15.8h\n"
+ "smax v10.8h, v10.8h, v15.8h\n"
+ "smin v20.8h, v20.8h, v17.8h\n"
+ "smin v8.8h, v8.8h, v17.8h\n"
+ "smin v3.8h, v3.8h, v17.8h\n"
+ "smin v10.8h, v10.8h, v17.8h\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "uzp1 v8.16b, v8.16b, v8.16b\n"
"uzp1 v3.16b, v3.16b, v3.16b\n"
- "uzp1 v0.16b, v0.16b, v0.16b\n"
- "str d3, [x27, x13]\n"
- "uzp1 v6.16b, v6.16b, v6.16b\n"
- "str d0, [x26, x13]\n"
- "str d6, [x25, x13]\n"
- "add x13, x13, #0x8\n"
+ "uzp1 v10.16b, v10.16b, v10.16b\n"
+ "str d20, [x9, x15]\n"
+ "str d8, [x28, x15]\n"
+ "str d3, [x27, x15]\n"
+ "str d10, [x26, x15]\n"
+ "add x15, x15, #0x8\n"
"beq 64f\n"
- "add x11, x11, #0x48\n"
+ "add x13, x13, #0x48\n"
"3:" // Oddments
"ldr x20, [%x[params], %[offsetof_Params_bias]]\n"
- "tbz x16, #2, 5f\n"
- "ld1 { v28.4s }, [x20], #0x10\n"
- "tbz x16, #1, 4f\n"
- "ld1 { v9.d }[0], [x20], #0x8\n"
- "tbz x16, #0, 7f\n"
- "ld1 { v9.s }[2], [x20]\n"
+ "tbz x17, #2, 5f\n"
+ "ld1 { v20.4s }, [x20], #0x10\n"
+ "tbz x17, #1, 4f\n"
+ "ld1 { v1.d }[0], [x20], #0x8\n"
+ "tbz x17, #0, 7f\n"
+ "ld1 { v1.s }[2], [x20]\n"
"b 7f\n"
"4:" // Oddments: Load bias: Bit 2: Bit 1: Unset
- "tbz x16, #0, 7f\n"
- "ld1 { v9.s }[0], [x20]\n"
+ "tbz x17, #0, 7f\n"
+ "ld1 { v1.s }[0], [x20]\n"
"b 7f\n"
"5:" // Oddments: Load bias: Bit 2: Unset
- "tbz x16, #1, 6f\n"
- "ld1 { v28.d }[0], [x20], #0x8\n"
- "tbz x16, #0, 7f\n"
- "ld1 { v28.s }[2], [x20]\n"
+ "tbz x17, #1, 6f\n"
+ "ld1 { v20.d }[0], [x20], #0x8\n"
+ "tbz x17, #0, 7f\n"
+ "ld1 { v20.s }[2], [x20]\n"
"b 7f\n"
"6:" // Oddments: Load bias: Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 7f\n"
- "ld1 { v28.s }[0], [x20]\n"
+ "tbz x17, #0, 7f\n"
+ "ld1 { v20.s }[0], [x20]\n"
"7:" // Oddments: Load bias: Bit 2: End
- "ldr d19, [x11, #0x0]\n"
- "ldr d7, [x11, #0x8]\n"
- "mov v3.16b, v28.16b\n"
- "mov v30.16b, v9.16b\n"
- "ldr d1, [x11, #0x10]\n"
- "ldr d17, [x11, #0x18]\n"
- "mov v0.16b, v28.16b\n"
- "mov v22.16b, v9.16b\n"
- "ldr d8, [x11, #0x20]\n"
- "ldr d31, [x11, #0x28]\n"
- "mov v6.16b, v28.16b\n"
- "mov v2.16b, v9.16b\n"
- "ldr d29, [x11, #0x30]\n"
- "ldr d16, [x11, #0x38]\n"
- "usubl v19.8h, v19.8b, v18.8b\n"
- "usubl v7.8h, v7.8b, v18.8b\n"
- "ldr d4, [x11, #0x40]\n"
- "ldp x24, x23, [x12, #0x0]\n"
- "usubl v1.8h, v1.8b, v18.8b\n"
- "usubl v17.8h, v17.8b, v18.8b\n"
- "ldp x22, x21, [x12, #0x10]\n"
- "ldr x20, [x12, #0x20]\n"
- "usubl v8.8h, v8.8b, v18.8b\n"
- "usubl v31.8h, v31.8b, v18.8b\n"
- "usubl v29.8h, v29.8b, v18.8b\n"
- "usubl v16.8h, v16.8b, v18.8b\n"
- "usubl v4.8h, v4.8b, v18.8b\n"
- "add x24, x24, x14\n"
- "add x23, x23, x14\n"
- "add x22, x22, x14\n"
- "add x21, x21, x14\n"
- "add x20, x20, x14\n"
- "tbz x16, #2, 9f\n"
- "ld1 { v23.s }[0], [x24], #0x4\n"
- "ld1 { v10.s }[0], [x23], #0x4\n"
- "ld1 { v11.s }[0], [x22], #0x4\n"
- "ld1 { v13.s }[0], [x21], #0x4\n"
- "ld1 { v27.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 8f\n"
- "ld1 { v23.h }[2], [x24], #0x2\n"
- "ld1 { v10.h }[2], [x23], #0x2\n"
- "ld1 { v11.h }[2], [x22], #0x2\n"
- "ld1 { v13.h }[2], [x21], #0x2\n"
- "ld1 { v27.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 11f\n"
- "ld1 { v23.b }[6], [x24]\n"
- "ld1 { v10.b }[6], [x23]\n"
- "ld1 { v11.b }[6], [x22]\n"
- "ld1 { v13.b }[6], [x21]\n"
- "ld1 { v27.b }[6], [x20]\n"
+ "ldr d22, [x13, #0x0]\n"
+ "ldr d24, [x13, #0x8]\n"
+ "mov v8.16b, v20.16b\n"
+ "mov v21.16b, v1.16b\n"
+ "ldr d9, [x13, #0x10]\n"
+ "ldr d7, [x13, #0x18]\n"
+ "mov v3.16b, v20.16b\n"
+ "mov v30.16b, v1.16b\n"
+ "ldr d25, [x13, #0x20]\n"
+ "ldr d4, [x13, #0x28]\n"
+ "mov v10.16b, v20.16b\n"
+ "mov v27.16b, v1.16b\n"
+ "ldr d13, [x13, #0x30]\n"
+ "ldr d14, [x13, #0x38]\n"
+ "usubl v22.8h, v22.8b, v29.8b\n"
+ "usubl v24.8h, v24.8b, v29.8b\n"
+ "ldr d2, [x13, #0x40]\n"
+ "ldp x24, x23, [x14, #0x0]\n"
+ "usubl v9.8h, v9.8b, v29.8b\n"
+ "usubl v7.8h, v7.8b, v29.8b\n"
+ "usubl v25.8h, v25.8b, v29.8b\n"
+ "usubl v4.8h, v4.8b, v29.8b\n"
+ "usubl v13.8h, v13.8b, v29.8b\n"
+ "usubl v14.8h, v14.8b, v29.8b\n"
+ "ldp x22, x21, [x14, #0x10]\n"
+ "usubl v2.8h, v2.8b, v29.8b\n"
+ "add x24, x24, x16\n"
+ "add x23, x23, x16\n"
+ "ldr x20, [x14, #0x20]\n"
+ "add x22, x22, x16\n"
+ "add x21, x21, x16\n"
+ "add x20, x20, x16\n"
+ "tbz x17, #2, 9f\n"
+ "ld1 { v11.s }[0], [x24], #0x4\n"
+ "ld1 { v28.s }[0], [x23], #0x4\n"
+ "ld1 { v18.s }[0], [x22], #0x4\n"
+ "ld1 { v19.s }[0], [x21], #0x4\n"
+ "ld1 { v23.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 8f\n"
+ "ld1 { v11.h }[2], [x24], #0x2\n"
+ "ld1 { v28.h }[2], [x23], #0x2\n"
+ "ld1 { v18.h }[2], [x22], #0x2\n"
+ "ld1 { v19.h }[2], [x21], #0x2\n"
+ "ld1 { v23.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 11f\n"
+ "ld1 { v11.b }[6], [x24]\n"
+ "ld1 { v28.b }[6], [x23]\n"
+ "ld1 { v18.b }[6], [x22]\n"
+ "ld1 { v19.b }[6], [x21]\n"
+ "ld1 { v23.b }[6], [x20]\n"
"b 11f\n"
"8:" // Oddments: Initial loads: Bit 2: Bit 1: Unset
- "tbz x16, #0, 11f\n"
- "ld1 { v23.b }[4], [x24]\n"
- "ld1 { v10.b }[4], [x23]\n"
- "ld1 { v11.b }[4], [x22]\n"
- "ld1 { v13.b }[4], [x21]\n"
- "ld1 { v27.b }[4], [x20]\n"
+ "tbz x17, #0, 11f\n"
+ "ld1 { v11.b }[4], [x24]\n"
+ "ld1 { v28.b }[4], [x23]\n"
+ "ld1 { v18.b }[4], [x22]\n"
+ "ld1 { v19.b }[4], [x21]\n"
+ "ld1 { v23.b }[4], [x20]\n"
"b 11f\n"
"9:" // Oddments: Initial loads: Bit 2: Unset
- "tbz x16, #1, 10f\n"
- "ld1 { v23.h }[0], [x24], #0x2\n"
- "ld1 { v10.h }[0], [x23], #0x2\n"
- "ld1 { v11.h }[0], [x22], #0x2\n"
- "ld1 { v13.h }[0], [x21], #0x2\n"
- "ld1 { v27.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 11f\n"
- "ld1 { v23.b }[2], [x24]\n"
- "ld1 { v10.b }[2], [x23]\n"
- "ld1 { v11.b }[2], [x22]\n"
- "ld1 { v13.b }[2], [x21]\n"
- "ld1 { v27.b }[2], [x20]\n"
+ "tbz x17, #1, 10f\n"
+ "ld1 { v11.h }[0], [x24], #0x2\n"
+ "ld1 { v28.h }[0], [x23], #0x2\n"
+ "ld1 { v18.h }[0], [x22], #0x2\n"
+ "ld1 { v19.h }[0], [x21], #0x2\n"
+ "ld1 { v23.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 11f\n"
+ "ld1 { v11.b }[2], [x24]\n"
+ "ld1 { v28.b }[2], [x23]\n"
+ "ld1 { v18.b }[2], [x22]\n"
+ "ld1 { v19.b }[2], [x21]\n"
+ "ld1 { v23.b }[2], [x20]\n"
"b 11f\n"
"10:" // Oddments: Initial loads: Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 11f\n"
- "ld1 { v23.b }[0], [x24]\n"
- "ld1 { v10.b }[0], [x23]\n"
- "ld1 { v11.b }[0], [x22]\n"
- "ld1 { v13.b }[0], [x21]\n"
- "ld1 { v27.b }[0], [x20]\n"
+ "tbz x17, #0, 11f\n"
+ "ld1 { v11.b }[0], [x24]\n"
+ "ld1 { v28.b }[0], [x23]\n"
+ "ld1 { v18.b }[0], [x22]\n"
+ "ld1 { v19.b }[0], [x21]\n"
+ "ld1 { v23.b }[0], [x20]\n"
"11:" // Oddments: Initial loads: Bit 2: End
- "ushll v23.8h, v23.8b, #0x0\n"
- "smlal v28.4s, v23.4h, v8.4h\n"
- "smlal2 v9.4s, v23.8h, v8.8h\n"
- "ldr x20, [x12, #0x28]\n"
- "smlal v3.4s, v23.4h, v17.4h\n"
- "smlal2 v30.4s, v23.8h, v17.8h\n"
- "ushll v10.8h, v10.8b, #0x0\n"
"ushll v11.8h, v11.8b, #0x0\n"
- "smlal v0.4s, v23.4h, v7.4h\n"
- "smlal2 v22.4s, v23.8h, v7.8h\n"
- "add x20, x20, x14\n"
- "smlal v6.4s, v23.4h, v19.4h\n"
- "smlal2 v2.4s, v23.8h, v19.8h\n"
- "ushll v13.8h, v13.8b, #0x0\n"
- "smlal v28.4s, v10.4h, v19.4h\n"
- "smlal2 v9.4s, v10.8h, v19.8h\n"
- "ushll v27.8h, v27.8b, #0x0\n"
- "smlal v3.4s, v11.4h, v1.4h\n"
- "smlal2 v30.4s, v11.8h, v1.8h\n"
- "smlal v28.4s, v13.4h, v31.4h\n"
- "smlal2 v9.4s, v13.8h, v31.8h\n"
- "smlal v3.4s, v13.4h, v8.4h\n"
- "smlal2 v30.4s, v13.8h, v8.8h\n"
- "smlal v0.4s, v13.4h, v1.4h\n"
- "smlal2 v22.4s, v13.8h, v1.8h\n"
- "smlal v6.4s, v13.4h, v7.4h\n"
- "smlal2 v2.4s, v13.8h, v7.8h\n"
- "tbz x16, #2, 13f\n"
- "ld1 { v26.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 12f\n"
- "ld1 { v26.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 15f\n"
- "ld1 { v26.b }[6], [x20]\n"
+ "ushll v28.8h, v28.8b, #0x0\n"
+ "ldr x20, [x14, #0x28]\n"
+ "ushll v18.8h, v18.8b, #0x0\n"
+ "ushll v19.8h, v19.8b, #0x0\n"
+ "ushll v23.8h, v23.8b, #0x0\n"
+ "smlal v20.4s, v11.4h, v25.4h\n"
+ "smlal2 v1.4s, v11.8h, v25.8h\n"
+ "smlal v8.4s, v11.4h, v7.4h\n"
+ "smlal2 v21.4s, v11.8h, v7.8h\n"
+ "add x20, x20, x16\n"
+ "smlal v3.4s, v11.4h, v24.4h\n"
+ "smlal2 v30.4s, v11.8h, v24.8h\n"
+ "smlal v10.4s, v11.4h, v22.4h\n"
+ "smlal2 v27.4s, v11.8h, v22.8h\n"
+ "smlal v20.4s, v28.4h, v22.4h\n"
+ "smlal2 v1.4s, v28.8h, v22.8h\n"
+ "smlal v8.4s, v18.4h, v9.4h\n"
+ "smlal2 v21.4s, v18.8h, v9.8h\n"
+ "smlal v3.4s, v19.4h, v9.4h\n"
+ "smlal2 v30.4s, v19.8h, v9.8h\n"
+ "smlal v10.4s, v19.4h, v24.4h\n"
+ "smlal2 v27.4s, v19.8h, v24.8h\n"
+ "smlal v20.4s, v19.4h, v4.4h\n"
+ "smlal2 v1.4s, v19.8h, v4.8h\n"
+ "smlal v8.4s, v19.4h, v25.4h\n"
+ "smlal2 v21.4s, v19.8h, v25.8h\n"
+ "tbz x17, #2, 13f\n"
+ "ld1 { v6.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 12f\n"
+ "ld1 { v6.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 15f\n"
+ "ld1 { v6.b }[6], [x20]\n"
"b 15f\n"
"12:" // Oddments: Load (3, 0): Bit 2: Bit 1: Unset
- "tbz x16, #0, 15f\n"
- "ld1 { v26.b }[4], [x20]\n"
+ "tbz x17, #0, 15f\n"
+ "ld1 { v6.b }[4], [x20]\n"
"b 15f\n"
"13:" // Oddments: Load (3, 0): Bit 2: Unset
- "tbz x16, #1, 14f\n"
- "ld1 { v26.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 15f\n"
- "ld1 { v26.b }[2], [x20]\n"
+ "tbz x17, #1, 14f\n"
+ "ld1 { v6.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 15f\n"
+ "ld1 { v6.b }[2], [x20]\n"
"b 15f\n"
"14:" // Oddments: Load (3, 0): Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 15f\n"
- "ld1 { v26.b }[0], [x20]\n"
+ "tbz x17, #0, 15f\n"
+ "ld1 { v6.b }[0], [x20]\n"
"15:" // Oddments: Load (3, 0): Bit 2: End
- "ushll v26.8h, v26.8b, #0x0\n"
- "smlal v0.4s, v26.4h, v29.4h\n"
- "smlal2 v22.4s, v26.8h, v29.8h\n"
- "ldr x20, [x12, #0x30]\n"
- "smlal v28.4s, v27.4h, v16.4h\n"
- "smlal2 v9.4s, v27.8h, v16.8h\n"
- "add x20, x20, x14\n"
- "smlal v3.4s, v27.4h, v29.4h\n"
- "smlal2 v30.4s, v27.8h, v29.8h\n"
- "smlal v0.4s, v27.4h, v8.4h\n"
- "smlal2 v22.4s, v27.8h, v8.8h\n"
- "smlal v6.4s, v27.4h, v17.4h\n"
- "smlal2 v2.4s, v27.8h, v17.8h\n"
- "tbz x16, #2, 17f\n"
- "ld1 { v23.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 16f\n"
- "ld1 { v23.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 19f\n"
- "ld1 { v23.b }[6], [x20]\n"
+ "ushll v6.8h, v6.8b, #0x0\n"
+ "ldr x20, [x14, #0x30]\n"
+ "smlal v20.4s, v23.4h, v14.4h\n"
+ "smlal2 v1.4s, v23.8h, v14.8h\n"
+ "smlal v8.4s, v23.4h, v13.4h\n"
+ "smlal2 v21.4s, v23.8h, v13.8h\n"
+ "smlal v10.4s, v23.4h, v7.4h\n"
+ "smlal2 v27.4s, v23.8h, v7.8h\n"
+ "smlal v3.4s, v6.4h, v13.4h\n"
+ "smlal2 v30.4s, v6.8h, v13.8h\n"
+ "add x20, x20, x16\n"
+ "smlal v3.4s, v23.4h, v25.4h\n"
+ "smlal2 v30.4s, v23.8h, v25.8h\n"
+ "tbz x17, #2, 17f\n"
+ "ld1 { v26.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 16f\n"
+ "ld1 { v26.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 19f\n"
+ "ld1 { v26.b }[6], [x20]\n"
"b 19f\n"
"16:" // Oddments: Load (3, 3): Bit 2: Bit 1: Unset
- "tbz x16, #0, 19f\n"
- "ld1 { v23.b }[4], [x20]\n"
+ "tbz x17, #0, 19f\n"
+ "ld1 { v26.b }[4], [x20]\n"
"b 19f\n"
"17:" // Oddments: Load (3, 3): Bit 2: Unset
- "tbz x16, #1, 18f\n"
- "ld1 { v23.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 19f\n"
- "ld1 { v23.b }[2], [x20]\n"
+ "tbz x17, #1, 18f\n"
+ "ld1 { v26.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 19f\n"
+ "ld1 { v26.b }[2], [x20]\n"
"b 19f\n"
"18:" // Oddments: Load (3, 3): Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 19f\n"
- "ld1 { v23.b }[0], [x20]\n"
+ "tbz x17, #0, 19f\n"
+ "ld1 { v26.b }[0], [x20]\n"
"19:" // Oddments: Load (3, 3): Bit 2: End
- "ushll v23.8h, v23.8b, #0x0\n"
- "ldr x20, [x12, #0x38]\n"
- "smlal v6.4s, v23.4h, v4.4h\n"
- "smlal2 v2.4s, v23.8h, v4.8h\n"
- "add x20, x20, x14\n"
- "tbz x16, #2, 21f\n"
- "ld1 { v21.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 20f\n"
- "ld1 { v21.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 23f\n"
- "ld1 { v21.b }[6], [x20]\n"
+ "ushll v26.8h, v26.8b, #0x0\n"
+ "ldr x20, [x14, #0x38]\n"
+ "smlal v10.4s, v26.4h, v2.4h\n"
+ "smlal2 v27.4s, v26.8h, v2.8h\n"
+ "add x20, x20, x16\n"
+ "tbz x17, #2, 21f\n"
+ "ld1 { v31.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 20f\n"
+ "ld1 { v31.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 23f\n"
+ "ld1 { v31.b }[6], [x20]\n"
"b 23f\n"
"20:" // Oddments: Load (0, 1): Bit 2: Bit 1: Unset
- "tbz x16, #0, 23f\n"
- "ld1 { v21.b }[4], [x20]\n"
+ "tbz x17, #0, 23f\n"
+ "ld1 { v31.b }[4], [x20]\n"
"b 23f\n"
"21:" // Oddments: Load (0, 1): Bit 2: Unset
- "tbz x16, #1, 22f\n"
- "ld1 { v21.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 23f\n"
- "ld1 { v21.b }[2], [x20]\n"
+ "tbz x17, #1, 22f\n"
+ "ld1 { v31.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 23f\n"
+ "ld1 { v31.b }[2], [x20]\n"
"b 23f\n"
"22:" // Oddments: Load (0, 1): Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 23f\n"
- "ld1 { v21.b }[0], [x20]\n"
+ "tbz x17, #0, 23f\n"
+ "ld1 { v31.b }[0], [x20]\n"
"23:" // Oddments: Load (0, 1): Bit 2: End
- "ushll v21.8h, v21.8b, #0x0\n"
- "ldr x20, [x12, #0x40]\n"
- "smlal v28.4s, v21.4h, v7.4h\n"
- "smlal2 v9.4s, v21.8h, v7.8h\n"
- "smlal v3.4s, v21.4h, v19.4h\n"
- "smlal2 v30.4s, v21.8h, v19.8h\n"
- "add x20, x20, x14\n"
- "tbz x16, #2, 25f\n"
- "ld1 { v18.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 24f\n"
- "ld1 { v18.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 27f\n"
- "ld1 { v18.b }[6], [x20]\n"
+ "ushll v31.8h, v31.8b, #0x0\n"
+ "ldr x20, [x14, #0x40]\n"
+ "smlal v20.4s, v31.4h, v24.4h\n"
+ "smlal2 v1.4s, v31.8h, v24.8h\n"
+ "smlal v8.4s, v31.4h, v22.4h\n"
+ "smlal2 v21.4s, v31.8h, v22.8h\n"
+ "add x20, x20, x16\n"
+ "tbz x17, #2, 25f\n"
+ "ld1 { v26.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 24f\n"
+ "ld1 { v26.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 27f\n"
+ "ld1 { v26.b }[6], [x20]\n"
"b 27f\n"
"24:" // Oddments: Load (0, 2): Bit 2: Bit 1: Unset
- "tbz x16, #0, 27f\n"
- "ld1 { v18.b }[4], [x20]\n"
+ "tbz x17, #0, 27f\n"
+ "ld1 { v26.b }[4], [x20]\n"
"b 27f\n"
"25:" // Oddments: Load (0, 2): Bit 2: Unset
- "tbz x16, #1, 26f\n"
- "ld1 { v18.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 27f\n"
- "ld1 { v18.b }[2], [x20]\n"
+ "tbz x17, #1, 26f\n"
+ "ld1 { v26.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 27f\n"
+ "ld1 { v26.b }[2], [x20]\n"
"b 27f\n"
"26:" // Oddments: Load (0, 2): Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 27f\n"
- "ld1 { v18.b }[0], [x20]\n"
+ "tbz x17, #0, 27f\n"
+ "ld1 { v26.b }[0], [x20]\n"
"27:" // Oddments: Load (0, 2): Bit 2: End
- "ushll v18.8h, v18.8b, #0x0\n"
- "ldr x20, [x12, #0x48]\n"
- "smlal v28.4s, v18.4h, v1.4h\n"
- "smlal2 v9.4s, v18.8h, v1.8h\n"
- "smlal v3.4s, v18.4h, v7.4h\n"
- "smlal2 v30.4s, v18.8h, v7.8h\n"
- "add x20, x20, x14\n"
- "tbz x16, #2, 29f\n"
- "ld1 { v15.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 28f\n"
- "ld1 { v15.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 31f\n"
- "ld1 { v15.b }[6], [x20]\n"
+ "ushll v26.8h, v26.8b, #0x0\n"
+ "ldr x20, [x14, #0x48]\n"
+ "smlal v20.4s, v26.4h, v9.4h\n"
+ "smlal2 v1.4s, v26.8h, v9.8h\n"
+ "smlal v8.4s, v26.4h, v24.4h\n"
+ "smlal2 v21.4s, v26.8h, v24.8h\n"
+ "add x20, x20, x16\n"
+ "tbz x17, #2, 29f\n"
+ "ld1 { v18.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 28f\n"
+ "ld1 { v18.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 31f\n"
+ "ld1 { v18.b }[6], [x20]\n"
"b 31f\n"
"28:" // Oddments: Load (2, 2): Bit 2: Bit 1: Unset
- "tbz x16, #0, 31f\n"
- "ld1 { v15.b }[4], [x20]\n"
+ "tbz x17, #0, 31f\n"
+ "ld1 { v18.b }[4], [x20]\n"
"b 31f\n"
"29:" // Oddments: Load (2, 2): Bit 2: Unset
- "tbz x16, #1, 30f\n"
- "ld1 { v15.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 31f\n"
- "ld1 { v15.b }[2], [x20]\n"
+ "tbz x17, #1, 30f\n"
+ "ld1 { v18.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 31f\n"
+ "ld1 { v18.b }[2], [x20]\n"
"b 31f\n"
"30:" // Oddments: Load (2, 2): Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 31f\n"
- "ld1 { v15.b }[0], [x20]\n"
+ "tbz x17, #0, 31f\n"
+ "ld1 { v18.b }[0], [x20]\n"
"31:" // Oddments: Load (2, 2): Bit 2: End
- "ushll v15.8h, v15.8b, #0x0\n"
- "ldr x20, [x12, #0x50]\n"
- "smlal v28.4s, v15.4h, v4.4h\n"
- "smlal2 v9.4s, v15.8h, v4.8h\n"
- "smlal v3.4s, v15.4h, v16.4h\n"
- "smlal2 v30.4s, v15.8h, v16.8h\n"
- "add x20, x20, x14\n"
- "smlal v0.4s, v15.4h, v31.4h\n"
- "smlal2 v22.4s, v15.8h, v31.8h\n"
- "smlal v6.4s, v15.4h, v8.4h\n"
- "smlal2 v2.4s, v15.8h, v8.8h\n"
- "tbz x16, #2, 33f\n"
- "ld1 { v20.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 32f\n"
- "ld1 { v20.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 35f\n"
- "ld1 { v20.b }[6], [x20]\n"
+ "ushll v18.8h, v18.8b, #0x0\n"
+ "ldr x20, [x14, #0x50]\n"
+ "smlal v20.4s, v18.4h, v2.4h\n"
+ "smlal2 v1.4s, v18.8h, v2.8h\n"
+ "smlal v8.4s, v18.4h, v14.4h\n"
+ "smlal2 v21.4s, v18.8h, v14.8h\n"
+ "smlal v3.4s, v18.4h, v4.4h\n"
+ "smlal2 v30.4s, v18.8h, v4.8h\n"
+ "smlal v10.4s, v18.4h, v25.4h\n"
+ "add x20, x20, x16\n"
+ "smlal2 v27.4s, v18.8h, v25.8h\n"
+ "tbz x17, #2, 33f\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 32f\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 35f\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 35f\n"
"32:" // Oddments: Load (1, 0): Bit 2: Bit 1: Unset
- "tbz x16, #0, 35f\n"
- "ld1 { v20.b }[4], [x20]\n"
+ "tbz x17, #0, 35f\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 35f\n"
"33:" // Oddments: Load (1, 0): Bit 2: Unset
- "tbz x16, #1, 34f\n"
- "ld1 { v20.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 35f\n"
- "ld1 { v20.b }[2], [x20]\n"
+ "tbz x17, #1, 34f\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 35f\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 35f\n"
"34:" // Oddments: Load (1, 0): Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 35f\n"
- "ld1 { v20.b }[0], [x20]\n"
+ "tbz x17, #0, 35f\n"
+ "ld1 { v16.b }[0], [x20]\n"
"35:" // Oddments: Load (1, 0): Bit 2: End
- "ushll v20.8h, v20.8b, #0x0\n"
- "ldr x20, [x12, #0x58]\n"
- "smlal v28.4s, v20.4h, v17.4h\n"
- "smlal2 v9.4s, v20.8h, v17.8h\n"
- "smlal v0.4s, v20.4h, v19.4h\n"
- "smlal2 v22.4s, v20.8h, v19.8h\n"
- "add x20, x20, x14\n"
- "tbz x16, #2, 37f\n"
- "ld1 { v11.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 36f\n"
- "ld1 { v11.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 39f\n"
- "ld1 { v11.b }[6], [x20]\n"
+ "ushll v16.8h, v16.8b, #0x0\n"
+ "ldr x20, [x14, #0x58]\n"
+ "smlal v20.4s, v16.4h, v7.4h\n"
+ "smlal2 v1.4s, v16.8h, v7.8h\n"
+ "smlal v3.4s, v16.4h, v22.4h\n"
+ "smlal2 v30.4s, v16.8h, v22.8h\n"
+ "add x20, x20, x16\n"
+ "tbz x17, #2, 37f\n"
+ "ld1 { v26.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 36f\n"
+ "ld1 { v26.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 39f\n"
+ "ld1 { v26.b }[6], [x20]\n"
"b 39f\n"
"36:" // Oddments: Load (1, 3): Bit 2: Bit 1: Unset
- "tbz x16, #0, 39f\n"
- "ld1 { v11.b }[4], [x20]\n"
+ "tbz x17, #0, 39f\n"
+ "ld1 { v26.b }[4], [x20]\n"
"b 39f\n"
"37:" // Oddments: Load (1, 3): Bit 2: Unset
- "tbz x16, #1, 38f\n"
- "ld1 { v11.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 39f\n"
- "ld1 { v11.b }[2], [x20]\n"
+ "tbz x17, #1, 38f\n"
+ "ld1 { v26.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 39f\n"
+ "ld1 { v26.b }[2], [x20]\n"
"b 39f\n"
"38:" // Oddments: Load (1, 3): Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 39f\n"
- "ld1 { v11.b }[0], [x20]\n"
+ "tbz x17, #0, 39f\n"
+ "ld1 { v26.b }[0], [x20]\n"
"39:" // Oddments: Load (1, 3): Bit 2: End
- "ushll v11.8h, v11.8b, #0x0\n"
- "ldr x20, [x12, #0x60]\n"
- "smlal v3.4s, v11.4h, v31.4h\n"
- "smlal2 v30.4s, v11.8h, v31.8h\n"
- "smlal v6.4s, v11.4h, v1.4h\n"
- "smlal2 v2.4s, v11.8h, v1.8h\n"
- "add x20, x20, x14\n"
- "tbz x16, #2, 41f\n"
- "ld1 { v23.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 40f\n"
- "ld1 { v23.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 43f\n"
- "ld1 { v23.b }[6], [x20]\n"
+ "ushll v26.8h, v26.8b, #0x0\n"
+ "ldr x20, [x14, #0x60]\n"
+ "smlal v8.4s, v26.4h, v4.4h\n"
+ "smlal2 v21.4s, v26.8h, v4.8h\n"
+ "smlal v10.4s, v26.4h, v9.4h\n"
+ "smlal2 v27.4s, v26.8h, v9.8h\n"
+ "add x20, x20, x16\n"
+ "tbz x17, #2, 41f\n"
+ "ld1 { v5.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 40f\n"
+ "ld1 { v5.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 43f\n"
+ "ld1 { v5.b }[6], [x20]\n"
"b 43f\n"
"40:" // Oddments: Load (2, 0): Bit 2: Bit 1: Unset
- "tbz x16, #0, 43f\n"
- "ld1 { v23.b }[4], [x20]\n"
+ "tbz x17, #0, 43f\n"
+ "ld1 { v5.b }[4], [x20]\n"
"b 43f\n"
"41:" // Oddments: Load (2, 0): Bit 2: Unset
- "tbz x16, #1, 42f\n"
- "ld1 { v23.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 43f\n"
- "ld1 { v23.b }[2], [x20]\n"
+ "tbz x17, #1, 42f\n"
+ "ld1 { v5.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 43f\n"
+ "ld1 { v5.b }[2], [x20]\n"
"b 43f\n"
"42:" // Oddments: Load (2, 0): Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 43f\n"
- "ld1 { v23.b }[0], [x20]\n"
+ "tbz x17, #0, 43f\n"
+ "ld1 { v5.b }[0], [x20]\n"
"43:" // Oddments: Load (2, 0): Bit 2: End
- "ushll v23.8h, v23.8b, #0x0\n"
- "ldr x20, [x12, #0x68]\n"
- "smlal v28.4s, v23.4h, v29.4h\n"
- "smlal2 v9.4s, v23.8h, v29.8h\n"
- "smlal v0.4s, v23.4h, v17.4h\n"
- "smlal2 v22.4s, v23.8h, v17.8h\n"
- "add x20, x20, x14\n"
- "tbz x16, #2, 45f\n"
- "ld1 { v20.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 44f\n"
- "ld1 { v20.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 47f\n"
- "ld1 { v20.b }[6], [x20]\n"
+ "ushll v5.8h, v5.8b, #0x0\n"
+ "ldr x20, [x14, #0x68]\n"
+ "smlal v20.4s, v5.4h, v13.4h\n"
+ "smlal2 v1.4s, v5.8h, v13.8h\n"
+ "smlal v3.4s, v5.4h, v7.4h\n"
+ "smlal2 v30.4s, v5.8h, v7.8h\n"
+ "add x20, x20, x16\n"
+ "tbz x17, #2, 45f\n"
+ "ld1 { v19.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 44f\n"
+ "ld1 { v19.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 47f\n"
+ "ld1 { v19.b }[6], [x20]\n"
"b 47f\n"
"44:" // Oddments: Load (2, 3): Bit 2: Bit 1: Unset
- "tbz x16, #0, 47f\n"
- "ld1 { v20.b }[4], [x20]\n"
+ "tbz x17, #0, 47f\n"
+ "ld1 { v19.b }[4], [x20]\n"
"b 47f\n"
"45:" // Oddments: Load (2, 3): Bit 2: Unset
- "tbz x16, #1, 46f\n"
- "ld1 { v20.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 47f\n"
- "ld1 { v20.b }[2], [x20]\n"
+ "tbz x17, #1, 46f\n"
+ "ld1 { v19.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 47f\n"
+ "ld1 { v19.b }[2], [x20]\n"
"b 47f\n"
"46:" // Oddments: Load (2, 3): Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 47f\n"
- "ld1 { v20.b }[0], [x20]\n"
+ "tbz x17, #0, 47f\n"
+ "ld1 { v19.b }[0], [x20]\n"
"47:" // Oddments: Load (2, 3): Bit 2: End
- "ushll v20.8h, v20.8b, #0x0\n"
- "ldr x20, [x12, #0x70]\n"
- "smlal v3.4s, v20.4h, v4.4h\n"
- "smlal2 v30.4s, v20.8h, v4.8h\n"
- "smlal v6.4s, v20.4h, v31.4h\n"
- "smlal2 v2.4s, v20.8h, v31.8h\n"
- "add x20, x20, x14\n"
- "tbz x16, #2, 49f\n"
- "ld1 { v8.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 48f\n"
- "ld1 { v8.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 51f\n"
- "ld1 { v8.b }[6], [x20]\n"
+ "ushll v19.8h, v19.8b, #0x0\n"
+ "ldr x20, [x14, #0x70]\n"
+ "smlal v8.4s, v19.4h, v2.4h\n"
+ "smlal2 v21.4s, v19.8h, v2.8h\n"
+ "smlal v10.4s, v19.4h, v4.4h\n"
+ "smlal2 v27.4s, v19.8h, v4.8h\n"
+ "add x20, x20, x16\n"
+ "tbz x17, #2, 49f\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 48f\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 51f\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 51f\n"
"48:" // Oddments: Load (3, 1): Bit 2: Bit 1: Unset
- "tbz x16, #0, 51f\n"
- "ld1 { v8.b }[4], [x20]\n"
+ "tbz x17, #0, 51f\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 51f\n"
"49:" // Oddments: Load (3, 1): Bit 2: Unset
- "tbz x16, #1, 50f\n"
- "ld1 { v8.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 51f\n"
- "ld1 { v8.b }[2], [x20]\n"
+ "tbz x17, #1, 50f\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 51f\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 51f\n"
"50:" // Oddments: Load (3, 1): Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 51f\n"
- "ld1 { v8.b }[0], [x20]\n"
+ "tbz x17, #0, 51f\n"
+ "ld1 { v16.b }[0], [x20]\n"
"51:" // Oddments: Load (3, 1): Bit 2: End
- "ushll v8.8h, v8.8b, #0x0\n"
- "ldr x20, [x12, #0x78]\n"
- "smlal v0.4s, v8.4h, v16.4h\n"
- "smlal2 v22.4s, v8.8h, v16.8h\n"
- "smlal v6.4s, v8.4h, v29.4h\n"
- "smlal2 v2.4s, v8.8h, v29.8h\n"
- "add x20, x20, x14\n"
- "tbz x16, #2, 53f\n"
- "ld1 { v8.s }[0], [x20], #0x4\n"
- "tbz x16, #1, 52f\n"
- "ld1 { v8.h }[2], [x20], #0x2\n"
- "tbz x16, #0, 55f\n"
- "ld1 { v8.b }[6], [x20]\n"
+ "ushll v16.8h, v16.8b, #0x0\n"
+ "ldr x20, [x14, #0x78]\n"
+ "smlal v3.4s, v16.4h, v14.4h\n"
+ "smlal2 v30.4s, v16.8h, v14.8h\n"
+ "smlal v10.4s, v16.4h, v13.4h\n"
+ "smlal2 v27.4s, v16.8h, v13.8h\n"
+ "add x20, x20, x16\n"
+ "tbz x17, #2, 53f\n"
+ "ld1 { v28.s }[0], [x20], #0x4\n"
+ "tbz x17, #1, 52f\n"
+ "ld1 { v28.h }[2], [x20], #0x2\n"
+ "tbz x17, #0, 55f\n"
+ "ld1 { v28.b }[6], [x20]\n"
"b 55f\n"
"52:" // Oddments: Load (3, 2): Bit 2: Bit 1: Unset
- "tbz x16, #0, 55f\n"
- "ld1 { v8.b }[4], [x20]\n"
+ "tbz x17, #0, 55f\n"
+ "ld1 { v28.b }[4], [x20]\n"
"b 55f\n"
"53:" // Oddments: Load (3, 2): Bit 2: Unset
- "tbz x16, #1, 54f\n"
- "ld1 { v8.h }[0], [x20], #0x2\n"
- "tbz x16, #0, 55f\n"
- "ld1 { v8.b }[2], [x20]\n"
+ "tbz x17, #1, 54f\n"
+ "ld1 { v28.h }[0], [x20], #0x2\n"
+ "tbz x17, #0, 55f\n"
+ "ld1 { v28.b }[2], [x20]\n"
"b 55f\n"
"54:" // Oddments: Load (3, 2): Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 55f\n"
- "ld1 { v8.b }[0], [x20]\n"
+ "tbz x17, #0, 55f\n"
+ "ld1 { v28.b }[0], [x20]\n"
"55:" // Oddments: Load (3, 2): Bit 2: End
- "ushll v8.8h, v8.8b, #0x0\n"
- "smlal v0.4s, v8.4h, v4.4h\n"
- "smlal2 v22.4s, v8.8h, v4.8h\n"
- "smlal v6.4s, v8.4h, v16.4h\n"
- "smlal2 v2.4s, v8.8h, v16.8h\n"
- "tbz x16, #2, 57f\n"
- "ld1 { v7.4s }, [x10], #0x10\n"
- "ld1 { v23.4s }, [x9], #0x10\n"
- "tbz x16, #1, 56f\n"
- "ld1 { v11.d }[0], [x10], #0x8\n"
- "ld1 { v27.d }[0], [x9], #0x8\n"
- "tbz x16, #0, 59f\n"
- "ld1 { v11.s }[2], [x10]\n"
- "ld1 { v27.s }[2], [x9]\n"
+ "ushll v28.8h, v28.8b, #0x0\n"
+ "smlal v3.4s, v28.4h, v2.4h\n"
+ "smlal2 v30.4s, v28.8h, v2.8h\n"
+ "smlal v10.4s, v28.4h, v14.4h\n"
+ "smlal2 v27.4s, v28.8h, v14.8h\n"
+ "tbz x17, #2, 57f\n"
+ "ld1 { v29.4s }, [x12], #0x10\n"
+ "ld1 { v19.4s }, [x11], #0x10\n"
+ "tbz x17, #1, 56f\n"
+ "ld1 { v7.d }[0], [x12], #0x8\n"
+ "ld1 { v18.d }[0], [x11], #0x8\n"
+ "tbz x17, #0, 59f\n"
+ "ld1 { v7.s }[2], [x12]\n"
+ "ld1 { v18.s }[2], [x11]\n"
"b 59f\n"
"56:" // Oddments: Load requant params: Bit 2: Bit 1: Unset
- "tbz x16, #0, 59f\n"
- "ld1 { v11.s }[0], [x10]\n"
- "ld1 { v27.s }[0], [x9]\n"
+ "tbz x17, #0, 59f\n"
+ "ld1 { v7.s }[0], [x12]\n"
+ "ld1 { v18.s }[0], [x11]\n"
"b 59f\n"
"57:" // Oddments: Load requant params: Bit 2: Unset
- "tbz x16, #1, 58f\n"
- "ld1 { v7.d }[0], [x10], #0x8\n"
- "ld1 { v23.d }[0], [x9], #0x8\n"
- "tbz x16, #0, 59f\n"
- "ld1 { v7.s }[2], [x10]\n"
- "ld1 { v23.s }[2], [x9]\n"
+ "tbz x17, #1, 58f\n"
+ "ld1 { v29.d }[0], [x12], #0x8\n"
+ "ld1 { v19.d }[0], [x11], #0x8\n"
+ "tbz x17, #0, 59f\n"
+ "ld1 { v29.s }[2], [x12]\n"
+ "ld1 { v19.s }[2], [x11]\n"
"b 59f\n"
"58:" // Oddments: Load requant params: Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 59f\n"
- "ld1 { v7.s }[0], [x10]\n"
- "ld1 { v23.s }[0], [x9]\n"
+ "tbz x17, #0, 59f\n"
+ "ld1 { v29.s }[0], [x12]\n"
+ "ld1 { v19.s }[0], [x11]\n"
"59:" // Oddments: Load requant params: Bit 2: End
- "sqrdmulh v28.4s, v28.4s, v7.4s\n"
- "and v20.16b, v28.16b, v23.16b\n"
- "add x28, x28, x13\n"
- "add x27, x27, x13\n"
- "sqrdmulh v9.4s, v9.4s, v11.4s\n"
- "sshr v20.4s, v20.4s, #0x1f\n"
- "add x26, x26, x13\n"
- "add x25, x25, x13\n"
- "and v4.16b, v9.16b, v27.16b\n"
- "sqrdmulh v3.4s, v3.4s, v7.4s\n"
- "sqrdmulh v0.4s, v0.4s, v7.4s\n"
- "sqrdmulh v6.4s, v6.4s, v7.4s\n"
- "sqadd v28.4s, v28.4s, v20.4s\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "and v19.16b, v3.16b, v23.16b\n"
- "sqrdmulh v30.4s, v30.4s, v11.4s\n"
- "and v29.16b, v0.16b, v23.16b\n"
- "sqrdmulh v22.4s, v22.4s, v11.4s\n"
- "and v26.16b, v6.16b, v23.16b\n"
- "sqrdmulh v2.4s, v2.4s, v11.4s\n"
- "sqadd v9.4s, v9.4s, v4.4s\n"
- "sshr v19.4s, v19.4s, #0x1f\n"
- "and v17.16b, v30.16b, v27.16b\n"
- "sshr v29.4s, v29.4s, #0x1f\n"
- "and v8.16b, v22.16b, v27.16b\n"
+ "sqrdmulh v20.4s, v20.4s, v29.4s\n"
+ "sqrdmulh v1.4s, v1.4s, v7.4s\n"
+ "add x9, x9, x15\n"
+ "add x28, x28, x15\n"
+ "sqrdmulh v8.4s, v8.4s, v29.4s\n"
+ "sqrdmulh v3.4s, v3.4s, v29.4s\n"
+ "add x27, x27, x15\n"
+ "add x26, x26, x15\n"
+ "sqrdmulh v10.4s, v10.4s, v29.4s\n"
+ "sqrdmulh v21.4s, v21.4s, v7.4s\n"
+ "and v24.16b, v20.16b, v19.16b\n"
+ "and v28.16b, v1.16b, v18.16b\n"
+ "and v26.16b, v8.16b, v19.16b\n"
+ "and v0.16b, v3.16b, v19.16b\n"
+ "sqrdmulh v30.4s, v30.4s, v7.4s\n"
+ "sqrdmulh v27.4s, v27.4s, v7.4s\n"
+ "sshr v24.4s, v24.4s, #0x1f\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
"sshr v26.4s, v26.4s, #0x1f\n"
- "and v13.16b, v2.16b, v27.16b\n"
- "sqadd v3.4s, v3.4s, v19.4s\n"
- "sshr v17.4s, v17.4s, #0x1f\n"
- "sqadd v0.4s, v0.4s, v29.4s\n"
- "sshr v8.4s, v8.4s, #0x1f\n"
- "sqadd v6.4s, v6.4s, v26.4s\n"
- "sshr v13.4s, v13.4s, #0x1f\n"
- "srshl v28.4s, v28.4s, v23.4s\n"
- "srshl v3.4s, v3.4s, v23.4s\n"
- "sqadd v30.4s, v30.4s, v17.4s\n"
- "srshl v0.4s, v0.4s, v23.4s\n"
- "sqadd v22.4s, v22.4s, v8.4s\n"
- "srshl v6.4s, v6.4s, v23.4s\n"
- "sqadd v2.4s, v2.4s, v13.4s\n"
- "srshl v9.4s, v9.4s, v27.4s\n"
- "sqxtn v28.4h, v28.4s\n"
- "srshl v30.4s, v30.4s, v27.4s\n"
+ "and v11.16b, v21.16b, v18.16b\n"
+ "sshr v0.4s, v0.4s, #0x1f\n"
+ "and v25.16b, v30.16b, v18.16b\n"
+ "sqadd v20.4s, v20.4s, v24.4s\n"
+ "and v5.16b, v10.16b, v19.16b\n"
+ "sqadd v1.4s, v1.4s, v28.4s\n"
+ "and v4.16b, v27.16b, v18.16b\n"
+ "sqadd v8.4s, v8.4s, v26.4s\n"
+ "sshr v11.4s, v11.4s, #0x1f\n"
+ "sshr v5.4s, v5.4s, #0x1f\n"
+ "sqadd v3.4s, v3.4s, v0.4s\n"
+ "sshr v25.4s, v25.4s, #0x1f\n"
+ "sshr v4.4s, v4.4s, #0x1f\n"
+ "srshl v20.4s, v20.4s, v19.4s\n"
+ "srshl v8.4s, v8.4s, v19.4s\n"
+ "sqadd v10.4s, v10.4s, v5.4s\n"
+ "sqadd v21.4s, v21.4s, v11.4s\n"
+ "srshl v3.4s, v3.4s, v19.4s\n"
+ "sqadd v30.4s, v30.4s, v25.4s\n"
+ "sqadd v27.4s, v27.4s, v4.4s\n"
+ "srshl v1.4s, v1.4s, v18.4s\n"
+ "srshl v10.4s, v10.4s, v19.4s\n"
+ "sqxtn v20.4h, v20.4s\n"
+ "srshl v21.4s, v21.4s, v18.4s\n"
+ "sqxtn v8.4h, v8.4s\n"
+ "srshl v30.4s, v30.4s, v18.4s\n"
"sqxtn v3.4h, v3.4s\n"
- "srshl v22.4s, v22.4s, v27.4s\n"
- "sqxtn v0.4h, v0.4s\n"
- "srshl v2.4s, v2.4s, v27.4s\n"
- "sqxtn v6.4h, v6.4s\n"
- "sqxtn2 v28.8h, v9.4s\n"
+ "srshl v27.4s, v27.4s, v18.4s\n"
+ "sqxtn v10.4h, v10.4s\n"
+ "sqxtn2 v20.8h, v1.4s\n"
+ "sqxtn2 v8.8h, v21.4s\n"
"sqxtn2 v3.8h, v30.4s\n"
- "sqxtn2 v0.8h, v22.4s\n"
- "sqxtn2 v6.8h, v2.4s\n"
- "sqadd v28.8h, v28.8h, v5.8h\n"
- "sqadd v3.8h, v3.8h, v5.8h\n"
- "sqadd v0.8h, v0.8h, v5.8h\n"
- "sqadd v6.8h, v6.8h, v5.8h\n"
- "smax v28.8h, v28.8h, v14.8h\n"
- "smax v3.8h, v3.8h, v14.8h\n"
- "smax v0.8h, v0.8h, v14.8h\n"
- "smax v6.8h, v6.8h, v14.8h\n"
- "smin v28.8h, v28.8h, v12.8h\n"
- "smin v3.8h, v3.8h, v12.8h\n"
- "smin v0.8h, v0.8h, v12.8h\n"
- "smin v6.8h, v6.8h, v12.8h\n"
- "uzp1 v28.16b, v28.16b, v28.16b\n"
+ "sqxtn2 v10.8h, v27.4s\n"
+ "sqadd v20.8h, v20.8h, v12.8h\n"
+ "sqadd v8.8h, v8.8h, v12.8h\n"
+ "sqadd v3.8h, v3.8h, v12.8h\n"
+ "sqadd v10.8h, v10.8h, v12.8h\n"
+ "smax v20.8h, v20.8h, v15.8h\n"
+ "smax v8.8h, v8.8h, v15.8h\n"
+ "smax v3.8h, v3.8h, v15.8h\n"
+ "smax v10.8h, v10.8h, v15.8h\n"
+ "smin v20.8h, v20.8h, v17.8h\n"
+ "smin v8.8h, v8.8h, v17.8h\n"
+ "smin v3.8h, v3.8h, v17.8h\n"
+ "smin v10.8h, v10.8h, v17.8h\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "uzp1 v8.16b, v8.16b, v8.16b\n"
"uzp1 v3.16b, v3.16b, v3.16b\n"
- "uzp1 v0.16b, v0.16b, v0.16b\n"
- "uzp1 v6.16b, v6.16b, v6.16b\n"
- "tbz x16, #2, 61f\n"
- "st1 { v28.s }[0], [x28], #0x4\n"
+ "uzp1 v10.16b, v10.16b, v10.16b\n"
+ "tbz x17, #2, 61f\n"
+ "st1 { v20.s }[0], [x9], #0x4\n"
+ "st1 { v8.s }[0], [x28], #0x4\n"
"st1 { v3.s }[0], [x27], #0x4\n"
- "st1 { v0.s }[0], [x26], #0x4\n"
- "st1 { v6.s }[0], [x25], #0x4\n"
- "tbz x16, #1, 60f\n"
- "st1 { v28.h }[2], [x28], #0x2\n"
+ "st1 { v10.s }[0], [x26], #0x4\n"
+ "tbz x17, #1, 60f\n"
+ "st1 { v20.h }[2], [x9], #0x2\n"
+ "st1 { v8.h }[2], [x28], #0x2\n"
"st1 { v3.h }[2], [x27], #0x2\n"
- "st1 { v0.h }[2], [x26], #0x2\n"
- "st1 { v6.h }[2], [x25], #0x2\n"
- "tbz x16, #0, 63f\n"
- "st1 { v28.b }[6], [x28], #0x1\n"
+ "st1 { v10.h }[2], [x26], #0x2\n"
+ "tbz x17, #0, 63f\n"
+ "st1 { v20.b }[6], [x9], #0x1\n"
+ "st1 { v8.b }[6], [x28], #0x1\n"
"st1 { v3.b }[6], [x27], #0x1\n"
- "st1 { v0.b }[6], [x26], #0x1\n"
- "st1 { v6.b }[6], [x25], #0x1\n"
+ "st1 { v10.b }[6], [x26], #0x1\n"
"b 63f\n"
"60:" // Oddments: Bit 2: Bit 1: Unset
- "tbz x16, #0, 63f\n"
- "st1 { v28.b }[4], [x28], #0x1\n"
+ "tbz x17, #0, 63f\n"
+ "st1 { v20.b }[4], [x9], #0x1\n"
+ "st1 { v8.b }[4], [x28], #0x1\n"
"st1 { v3.b }[4], [x27], #0x1\n"
- "st1 { v0.b }[4], [x26], #0x1\n"
- "st1 { v6.b }[4], [x25], #0x1\n"
+ "st1 { v10.b }[4], [x26], #0x1\n"
"b 63f\n"
"61:" // Oddments: Bit 2: Unset
- "tbz x16, #1, 62f\n"
- "st1 { v28.h }[0], [x28], #0x2\n"
+ "tbz x17, #1, 62f\n"
+ "st1 { v20.h }[0], [x9], #0x2\n"
+ "st1 { v8.h }[0], [x28], #0x2\n"
"st1 { v3.h }[0], [x27], #0x2\n"
- "st1 { v0.h }[0], [x26], #0x2\n"
- "st1 { v6.h }[0], [x25], #0x2\n"
- "tbz x16, #0, 63f\n"
- "st1 { v28.b }[2], [x28], #0x1\n"
+ "st1 { v10.h }[0], [x26], #0x2\n"
+ "tbz x17, #0, 63f\n"
+ "st1 { v20.b }[2], [x9], #0x1\n"
+ "st1 { v8.b }[2], [x28], #0x1\n"
"st1 { v3.b }[2], [x27], #0x1\n"
- "st1 { v0.b }[2], [x26], #0x1\n"
- "st1 { v6.b }[2], [x25], #0x1\n"
+ "st1 { v10.b }[2], [x26], #0x1\n"
"b 63f\n"
"62:" // Oddments: Bit 2: Unset: Bit 1: Unset
- "tbz x16, #0, 63f\n"
- "st1 { v28.b }[0], [x28], #0x1\n"
+ "tbz x17, #0, 63f\n"
+ "st1 { v20.b }[0], [x9], #0x1\n"
+ "st1 { v8.b }[0], [x28], #0x1\n"
"st1 { v3.b }[0], [x27], #0x1\n"
- "st1 { v0.b }[0], [x26], #0x1\n"
- "st1 { v6.b }[0], [x25], #0x1\n"
+ "st1 { v10.b }[0], [x26], #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_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", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+ : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
);
}