aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp1798
1 files changed, 899 insertions, 899 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
index de072a7d55..fb533893a6 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst/generic.cpp
@@ -104,16 +104,16 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
"ldr x23, [%x[params], %[offsetof_Params_requant]]\n"
"lsr x8, x7, #0x3\n"
"add x20, x23, %[offsetof_Requantize32_a_offset]\n"
- "ld1r { v12.16b }, [x20]\n"
+ "ld1r { v6.16b }, [x20]\n"
"ldr x22, [%x[params], %[offsetof_Params_outptrs]]\n"
"add x21, x23, %[offsetof_Requantize32_b_offset]\n"
"add x20, x23, %[offsetof_Requantize32_c_offset]\n"
- "ld1r { v13.16b }, [x21]\n"
- "ld1r { v11.8h }, [x20]\n"
+ "ld1r { v15.16b }, [x21]\n"
+ "ld1r { v13.8h }, [x20]\n"
"add x21, x23, %[offsetof_Requantize32_minval]\n"
"add x20, x23, %[offsetof_Requantize32_maxval]\n"
- "ld1r { v16.8h }, [x21]\n"
- "ld1r { v14.8h }, [x20]\n"
+ "ld1r { v17.8h }, [x21]\n"
+ "ld1r { v24.8h }, [x20]\n"
"mov x17, #0x0\n"
"mov x16, #0x0\n"
"add x15, %x[params], %[offsetof_Params_inptrs]\n"
@@ -123,563 +123,563 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
"ldp x11, x10, [x22, #0x0]\n"
"ldp x9, x28, [x22, #0x10]\n"
"cbz x8, 3f\n"
- "ldr d0, [x14, #0x0]\n"
- "ldr d1, [x14, #0x8]\n"
+ "ldr d11, [x14, #0x0]\n"
+ "ldr d22, [x14, #0x8]\n"
"subs x8, x8, #0x1\n"
- "usubl v0.8h, v0.8b, v13.8b\n"
- "ldr d2, [x14, #0x10]\n"
- "ldr d3, [x14, #0x18]\n"
- "usubl v1.8h, v1.8b, v13.8b\n"
- "usubl v2.8h, v2.8b, v13.8b\n"
- "ldr d4, [x14, #0x20]\n"
- "ldr d5, [x14, #0x28]\n"
- "usubl v3.8h, v3.8b, v13.8b\n"
- "usubl v4.8h, v4.8b, v13.8b\n"
- "ldr d6, [x14, #0x30]\n"
+ "usubl v11.8h, v11.8b, v15.8b\n"
+ "ldr d14, [x14, #0x10]\n"
+ "ldr d28, [x14, #0x18]\n"
+ "usubl v22.8h, v22.8b, v15.8b\n"
+ "usubl v14.8h, v14.8b, v15.8b\n"
+ "ldr d18, [x14, #0x20]\n"
+ "ldr d9, [x14, #0x28]\n"
+ "usubl v28.8h, v28.8b, v15.8b\n"
+ "usubl v18.8h, v18.8b, v15.8b\n"
+ "ldr d26, [x14, #0x30]\n"
"ldr d7, [x14, #0x38]\n"
- "usubl v5.8h, v5.8b, v13.8b\n"
- "usubl v6.8h, v6.8b, v13.8b\n"
- "ldr d8, [x14, #0x40]\n"
- "ldr x24, [%x[params], %[offsetof_Params_bias]]\n"
- "usubl v7.8h, v7.8b, v13.8b\n"
- "usubl v8.8h, v8.8b, v13.8b\n"
- "ldr q15, [x24, #0x0]\n"
- "ldr q17, [x24, #0x10]\n"
- "add x24, x24, #0x20\n"
- "str x24, [%x[params], %[offsetof_Params_bias]]\n"
+ "usubl v9.8h, v9.8b, v15.8b\n"
+ "usubl v26.8h, v26.8b, v15.8b\n"
+ "ldr d4, [x14, #0x40]\n"
+ "ldr x20, [%x[params], %[offsetof_Params_bias]]\n"
+ "usubl v7.8h, v7.8b, v15.8b\n"
+ "usubl v4.8h, v4.8b, v15.8b\n"
+ "ldr q5, [x20, #0x0]\n"
+ "ldr q3, [x20, #0x10]\n"
+ "add x20, x20, #0x20\n"
+ "str x20, [%x[params], %[offsetof_Params_bias]]\n"
"ldp x27, x26, [x15, #0x0]\n"
"ldp x25, x24, [x15, #0x10]\n"
- "mov v10.16b, v15.16b\n"
- "mov v20.16b, v17.16b\n"
+ "mov v21.16b, v5.16b\n"
+ "mov v8.16b, v3.16b\n"
"ldp x23, x22, [x15, #0x20]\n"
"ldp x21, x20, [x15, #0x30]\n"
- "mov v9.16b, v15.16b\n"
- "mov v23.16b, v17.16b\n"
- "ldr d31, [x27, x17]\n"
- "ldr d30, [x26, x17]\n"
- "mov v21.16b, v15.16b\n"
- "mov v22.16b, v17.16b\n"
- "ldr d29, [x25, x17]\n"
- "ldr d28, [x24, x17]\n"
- "usubl v31.8h, v31.8b, v12.8b\n"
- "usubl v30.8h, v30.8b, v12.8b\n"
- "ldr d27, [x23, x17]\n"
- "ldr d26, [x22, x17]\n"
- "usubl v29.8h, v29.8b, v12.8b\n"
- "usubl v28.8h, v28.8b, v12.8b\n"
- "ldr d25, [x21, x17]\n"
- "ldr d24, [x20, x17]\n"
- "usubl v27.8h, v27.8b, v12.8b\n"
- "usubl v26.8h, v26.8b, v12.8b\n"
- "usubl v25.8h, v25.8b, v12.8b\n"
- "usubl v24.8h, v24.8b, v12.8b\n"
+ "mov v20.16b, v5.16b\n"
+ "mov v0.16b, v3.16b\n"
+ "ldr d25, [x27, x17]\n"
+ "ldr d27, [x26, x17]\n"
+ "mov v19.16b, v5.16b\n"
+ "mov v31.16b, v3.16b\n"
+ "ldr d1, [x25, x17]\n"
+ "ldr d2, [x24, x17]\n"
+ "usubl v25.8h, v25.8b, v6.8b\n"
+ "usubl v27.8h, v27.8b, v6.8b\n"
+ "ldr d12, [x23, x17]\n"
+ "ldr d16, [x22, x17]\n"
+ "usubl v1.8h, v1.8b, v6.8b\n"
+ "usubl v2.8h, v2.8b, v6.8b\n"
+ "ldr d23, [x21, x17]\n"
+ "ldr d10, [x20, x17]\n"
+ "usubl v12.8h, v12.8b, v6.8b\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "usubl v23.8h, v23.8b, v6.8b\n"
+ "usubl v10.8h, v10.8b, v6.8b\n"
"beq 2f\n"
"1:" // Loop
- "smlal v15.4s, v31.4h, v8.4h\n"
- "smlal2 v17.4s, v31.8h, v8.8h\n"
- "ldr x24, [x15, #0x40]\n"
- "ldr x22, [x15, #0x48]\n"
- "smlal v10.4s, v31.4h, v6.4h\n"
- "smlal2 v20.4s, v31.8h, v6.8h\n"
- "ldr x21, [x15, #0x50]\n"
- "ldr x20, [x15, #0x58]\n"
- "smlal v15.4s, v30.4h, v0.4h\n"
- "smlal2 v17.4s, v30.8h, v0.8h\n"
- "ldr q19, [x13, #0x0]\n"
- "ldr x23, [x15, #0x78]\n"
- "smlal v10.4s, v28.4h, v1.4h\n"
- "smlal2 v20.4s, v28.8h, v1.8h\n"
- "ldr d28, [x22, x17]\n"
- "usubl v28.8h, v28.8b, v12.8b\n"
- "smlal v15.4s, v29.4h, v1.4h\n"
- "smlal2 v17.4s, v29.8h, v1.8h\n"
- "ldr d29, [x24, x17]\n"
- "usubl v29.8h, v29.8b, v12.8b\n"
- "smlal v10.4s, v27.4h, v2.4h\n"
- "smlal2 v20.4s, v27.8h, v2.8h\n"
+ "ldr q30, [x13, #0x0]\n"
+ "ldr q29, [x12, #0x0]\n"
+ "smlal v5.4s, v25.4h, v4.4h\n"
+ "smlal2 v3.4s, v25.8h, v4.8h\n"
+ "ldr x21, [x15, #0x58]\n"
+ "ldr x20, [x15, #0x78]\n"
+ "smlal v5.4s, v27.4h, v11.4h\n"
+ "smlal v21.4s, v25.4h, v26.4h\n"
+ "ldr x25, [x15, #0x60]\n"
+ "ldr x24, [x15, #0x80]\n"
+ "smlal v20.4s, v25.4h, v14.4h\n"
+ "smlal v19.4s, v25.4h, v11.4h\n"
+ "smlal2 v3.4s, v27.8h, v11.8h\n"
"ldr d27, [x21, x17]\n"
- "usubl v27.8h, v27.8b, v12.8b\n"
- "smlal v15.4s, v26.4h, v3.4h\n"
- "smlal2 v17.4s, v26.8h, v3.8h\n"
- "ldr d26, [x20, x17]\n"
- "ldr x20, [x15, #0x60]\n"
- "smlal v10.4s, v24.4h, v0.4h\n"
- "smlal2 v20.4s, v24.8h, v0.8h\n"
- "usubl v26.8h, v26.8b, v12.8b\n"
- "ldr x21, [x15, #0x80]\n"
- "smlal v15.4s, v25.4h, v4.4h\n"
- "smlal2 v17.4s, v25.8h, v4.8h\n"
- "ldr d25, [x20, x17]\n"
- "ldr x20, [x15, #0x68]\n"
- "smlal v10.4s, v29.4h, v4.4h\n"
- "smlal2 v20.4s, v29.8h, v4.8h\n"
- "ldr d29, [x20, x17]\n"
- "usubl v25.8h, v25.8b, v12.8b\n"
- "smlal v15.4s, v24.4h, v2.4h\n"
- "smlal2 v17.4s, v24.8h, v2.8h\n"
- "ldr q18, [x12, #0x0]\n"
+ "usubl v27.8h, v27.8b, v6.8b\n"
+ "smlal v5.4s, v1.4h, v22.4h\n"
+ "smlal2 v8.4s, v25.8h, v26.8h\n"
+ "smlal2 v0.4s, v25.8h, v14.8h\n"
+ "ldr x23, [x15, #0x68]\n"
"ldr x22, [x15, #0x88]\n"
- "smlal v10.4s, v28.4h, v5.4h\n"
- "smlal2 v20.4s, v28.8h, v5.8h\n"
- "ldr d28, [x21, x17]\n"
- "ldr x21, [x15, #0x70]\n"
- "smlal v9.4s, v31.4h, v2.4h\n"
- "smlal2 v23.4s, v31.8h, v2.8h\n"
- "usubl v28.8h, v28.8b, v12.8b\n"
+ "smlal2 v31.4s, v25.8h, v11.8h\n"
+ "ldr d25, [x20, x17]\n"
+ "usubl v25.8h, v25.8b, v6.8b\n"
+ "smlal v21.4s, v2.4h, v22.4h\n"
+ "smlal v20.4s, v27.4h, v28.4h\n"
+ "smlal v19.4s, v25.4h, v18.4h\n"
+ "ldr x21, [x15, #0x40]\n"
+ "ldr x20, [x15, #0x70]\n"
+ "smlal2 v3.4s, v1.8h, v22.8h\n"
+ "ldr d1, [x25, x17]\n"
+ "usubl v1.8h, v1.8b, v6.8b\n"
+ "smlal v5.4s, v16.4h, v28.4h\n"
+ "smlal2 v8.4s, v2.8h, v22.8h\n"
+ "ldr d2, [x24, x17]\n"
+ "usubl v2.8h, v2.8b, v6.8b\n"
+ "smlal2 v0.4s, v27.8h, v28.8h\n"
+ "ldr d27, [x23, x17]\n"
+ "smlal2 v31.4s, v25.8h, v18.8h\n"
+ "ldr d25, [x22, x17]\n"
+ "smlal v21.4s, v12.4h, v14.4h\n"
"ldr x25, [x15, #0x98]\n"
- "smlal v15.4s, v27.4h, v5.4h\n"
- "smlal2 v17.4s, v27.8h, v5.8h\n"
- "usubl v29.8h, v29.8b, v12.8b\n"
- "ldr x24, [x15, #0x90]\n"
- "smlal v10.4s, v27.4h, v3.4h\n"
- "smlal2 v20.4s, v27.8h, v3.8h\n"
+ "smlal v20.4s, v1.4h, v11.4h\n"
+ "smlal v19.4s, v2.4h, v22.4h\n"
+ "ldr x24, [x15, #0x50]\n"
+ "smlal2 v3.4s, v16.8h, v28.8h\n"
+ "ldr d16, [x21, x17]\n"
+ "usubl v27.8h, v27.8b, v6.8b\n"
+ "smlal v5.4s, v23.4h, v18.4h\n"
+ "usubl v25.8h, v25.8b, v6.8b\n"
+ "smlal2 v8.4s, v12.8h, v14.8h\n"
+ "ldr d12, [x20, x17]\n"
+ "ldr x23, [x15, #0x48]\n"
+ "smlal2 v0.4s, v1.8h, v11.8h\n"
+ "smlal2 v31.4s, v2.8h, v22.8h\n"
+ "ldr x21, [x15, #0x90]\n"
+ "ldr x20, [x15, #0xa8]\n"
+ "smlal v21.4s, v10.4h, v11.4h\n"
+ "smlal v20.4s, v27.4h, v18.4h\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x22, [x15, #0xa0]\n"
+ "smlal v19.4s, v25.4h, v9.4h\n"
+ "smlal2 v3.4s, v23.8h, v18.8h\n"
+ "ldr d23, [x25, x17]\n"
+ "usubl v12.8h, v12.8b, v6.8b\n"
+ "usubl v23.8h, v23.8b, v6.8b\n"
+ "smlal v5.4s, v10.4h, v14.4h\n"
+ "smlal2 v8.4s, v10.8h, v11.8h\n"
+ "ldr d11, [x24, x17]\n"
+ "usubl v11.8h, v11.8b, v6.8b\n"
+ "smlal2 v0.4s, v27.8h, v18.8h\n"
"ldr d27, [x23, x17]\n"
- "usubl v27.8h, v27.8b, v12.8b\n"
- "smlal v21.4s, v31.4h, v0.4h\n"
- "smlal v9.4s, v26.4h, v3.4h\n"
- "ldr x23, [x15, #0xa8]\n"
- "ldr x20, [x15, #0xa0]\n"
- "smlal2 v23.4s, v26.8h, v3.8h\n"
- "ldr d26, [x22, x17]\n"
- "smlal2 v22.4s, v31.8h, v0.8h\n"
- "ldr d24, [x21, x17]\n"
- "smlal v21.4s, v27.4h, v4.4h\n"
- "smlal v9.4s, v25.4h, v0.4h\n"
- "usubl v26.8h, v26.8b, v12.8b\n"
- "ldr x22, [x15, #0xb0]\n"
- "smlal2 v23.4s, v25.8h, v0.8h\n"
- "ldr q30, [x13, #0x10]\n"
- "smlal2 v22.4s, v27.8h, v4.8h\n"
- "ldr d27, [x20, x17]\n"
- "smlal v21.4s, v28.4h, v1.4h\n"
- "smlal v15.4s, v25.4h, v6.4h\n"
- "usubl v24.8h, v24.8b, v12.8b\n"
- "ldr x21, [x15, #0xb8]\n"
- "smlal2 v17.4s, v25.8h, v6.8h\n"
- "ldr d25, [x24, x17]\n"
- "smlal v9.4s, v29.4h, v4.4h\n"
- "usubl v25.8h, v25.8b, v12.8b\n"
- "smlal2 v23.4s, v29.8h, v4.8h\n"
- "ldr d29, [x25, x17]\n"
- "ldr q31, [x12, #0x10]\n"
- "smlal2 v22.4s, v28.8h, v1.8h\n"
- "smlal v21.4s, v26.4h, v5.4h\n"
- "usubl v29.8h, v29.8b, v12.8b\n"
- "smlal v15.4s, v24.4h, v7.4h\n"
- "ldr x20, [x15, #0xc0]\n"
- "smlal2 v17.4s, v24.8h, v7.8h\n"
- "smlal v9.4s, v24.4h, v1.4h\n"
- "usubl v27.8h, v27.8b, v12.8b\n"
- "ldr x24, [%x[params], %[offsetof_Params_bias]]\n"
- "smlal2 v23.4s, v24.8h, v1.8h\n"
- "ldr d24, [x23, x17]\n"
- "smlal2 v22.4s, v26.8h, v5.8h\n"
- "ldr d26, [x22, x17]\n"
- "smlal v21.4s, v29.4h, v2.4h\n"
- "usubl v24.8h, v24.8b, v12.8b\n"
- "smlal2 v22.4s, v29.8h, v2.8h\n"
+ "smlal2 v31.4s, v25.8h, v9.8h\n"
+ "ldr d25, [x21, x17]\n"
+ "ldr x21, [x15, #0xb0]\n"
+ "smlal v21.4s, v16.4h, v18.4h\n"
+ "smlal v20.4s, v12.4h, v22.4h\n"
+ "smlal v19.4s, v23.4h, v14.4h\n"
+ "smlal2 v3.4s, v10.8h, v14.8h\n"
+ "ldr d10, [x20, x17]\n"
+ "usubl v27.8h, v27.8b, v6.8b\n"
+ "usubl v25.8h, v25.8b, v6.8b\n"
+ "usubl v10.8h, v10.8b, v6.8b\n"
+ "smlal v5.4s, v11.4h, v9.4h\n"
+ "ldr x20, [x15, #0xb8]\n"
+ "smlal2 v8.4s, v16.8h, v18.8h\n"
+ "ldr d18, [x22, x17]\n"
+ "ldr d16, [x21, x17]\n"
+ "smlal2 v0.4s, v12.8h, v22.8h\n"
+ "ldr d22, [x20, x17]\n"
+ "smlal2 v31.4s, v23.8h, v14.8h\n"
+ "ldr q14, [x13, #0x10]\n"
+ "smlal v21.4s, v27.4h, v9.4h\n"
+ "smlal v20.4s, v25.4h, v26.4h\n"
+ "smlal v19.4s, v10.4h, v28.4h\n"
+ "usubl v18.8h, v18.8b, v6.8b\n"
+ "ldr x21, [x15, #0xc0]\n"
+ "smlal2 v3.4s, v11.8h, v9.8h\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "smlal v5.4s, v1.4h, v26.4h\n"
+ "ldr x20, [%x[params], %[offsetof_Params_bias]]\n"
+ "smlal2 v8.4s, v27.8h, v9.8h\n"
+ "ldr d27, [x21, x17]\n"
+ "smlal2 v0.4s, v25.8h, v26.8h\n"
+ "ldr q25, [x12, #0x10]\n"
+ "smlal2 v31.4s, v10.8h, v28.8h\n"
+ "smlal v21.4s, v11.4h, v28.4h\n"
+ "usubl v22.8h, v22.8b, v6.8b\n"
"add x14, x14, #0x48\n"
- "smlal v9.4s, v25.4h, v6.4h\n"
- "smlal v21.4s, v24.4h, v3.4h\n"
- "usubl v26.8h, v26.8b, v12.8b\n"
+ "smlal v20.4s, v18.4h, v7.4h\n"
+ "smlal v19.4s, v16.4h, v7.4h\n"
+ "usubl v27.8h, v27.8b, v6.8b\n"
+ "add x17, x17, #0x8\n"
+ "smlal2 v3.4s, v1.8h, v26.8h\n"
+ "smlal v5.4s, v12.4h, v7.4h\n"
+ "sqrdmulh v5.4s, v5.4s, v30.4s\n"
"subs x8, x8, #0x1\n"
- "smlal v10.4s, v28.4h, v7.4h\n"
- "smlal2 v20.4s, v28.8h, v7.8h\n"
- "sqrdmulh v15.4s, v15.4s, v19.4s\n"
+ "smlal2 v8.4s, v11.8h, v28.8h\n"
+ "smlal2 v0.4s, v18.8h, v7.8h\n"
+ "and v28.16b, v5.16b, v29.16b\n"
"add x13, x13, #0x20\n"
- "smlal2 v23.4s, v25.8h, v6.8h\n"
- "ldr d25, [x21, x17]\n"
- "smlal2 v22.4s, v24.8h, v3.8h\n"
- "usubl v25.8h, v25.8b, v12.8b\n"
- "smlal v9.4s, v27.4h, v7.4h\n"
- "smlal v21.4s, v26.4h, v7.4h\n"
- "and v0.16b, v15.16b, v18.16b\n"
+ "smlal2 v31.4s, v16.8h, v7.8h\n"
+ "smlal v21.4s, v2.4h, v7.4h\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
"add x12, x12, #0x20\n"
- "smlal v10.4s, v29.4h, v8.4h\n"
- "smlal2 v20.4s, v29.8h, v8.8h\n"
- "ldr d29, [x20, x17]\n"
- "usubl v29.8h, v29.8b, v12.8b\n"
- "smlal2 v23.4s, v27.8h, v7.8h\n"
- "smlal2 v22.4s, v26.8h, v7.8h\n"
- "sqrdmulh v17.4s, v17.4s, v30.4s\n"
- "add x17, x17, #0x8\n"
- "smlal v9.4s, v24.4h, v5.4h\n"
- "smlal v21.4s, v25.4h, v6.4h\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "smlal2 v23.4s, v24.8h, v5.8h\n"
- "smlal2 v22.4s, v25.8h, v6.8h\n"
- "and v7.16b, v17.16b, v31.16b\n"
- "smlal v9.4s, v25.4h, v8.4h\n"
- "smlal v21.4s, v29.4h, v8.4h\n"
- "sqrdmulh v10.4s, v10.4s, v19.4s\n"
- "smlal2 v23.4s, v25.8h, v8.8h\n"
- "smlal2 v22.4s, v29.8h, v8.8h\n"
- "sqrdmulh v9.4s, v9.4s, v19.4s\n"
- "sqrdmulh v21.4s, v21.4s, v19.4s\n"
- "sqadd v15.4s, v15.4s, v0.4s\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "and v19.16b, v10.16b, v18.16b\n"
+ "smlal v20.4s, v10.4h, v9.4h\n"
+ "smlal v19.4s, v22.4h, v26.4h\n"
+ "sqadd v5.4s, v5.4s, v28.4s\n"
+ "smlal2 v3.4s, v12.8h, v7.8h\n"
+ "smlal2 v8.4s, v2.8h, v7.8h\n"
+ "sqrdmulh v3.4s, v3.4s, v14.4s\n"
+ "smlal2 v0.4s, v10.8h, v9.8h\n"
+ "smlal2 v31.4s, v22.8h, v26.8h\n"
+ "and v16.16b, v3.16b, v25.16b\n"
+ "smlal v21.4s, v23.4h, v4.4h\n"
+ "smlal v20.4s, v22.4h, v4.4h\n"
+ "sqrdmulh v21.4s, v21.4s, v30.4s\n"
+ "smlal v19.4s, v27.4h, v4.4h\n"
+ "smlal2 v8.4s, v23.8h, v4.8h\n"
"sqrdmulh v20.4s, v20.4s, v30.4s\n"
- "and v27.16b, v9.16b, v18.16b\n"
- "sqrdmulh v23.4s, v23.4s, v30.4s\n"
- "and v0.16b, v21.16b, v18.16b\n"
- "sqrdmulh v22.4s, v22.4s, v30.4s\n"
- "sqadd v17.4s, v17.4s, v7.4s\n"
- "sshr v19.4s, v19.4s, #0x1f\n"
- "and v5.16b, v20.16b, v31.16b\n"
- "sshr v27.4s, v27.4s, #0x1f\n"
- "and v4.16b, v23.16b, v31.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "and v7.16b, v22.16b, v31.16b\n"
- "sqadd v10.4s, v10.4s, v19.4s\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sqadd v9.4s, v9.4s, v27.4s\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sqadd v21.4s, v21.4s, v0.4s\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "srshl v15.4s, v15.4s, v18.4s\n"
- "srshl v10.4s, v10.4s, v18.4s\n"
- "sqadd v20.4s, v20.4s, v5.4s\n"
- "srshl v9.4s, v9.4s, v18.4s\n"
- "sqadd v23.4s, v23.4s, v4.4s\n"
- "srshl v21.4s, v21.4s, v18.4s\n"
- "sqadd v22.4s, v22.4s, v7.4s\n"
- "srshl v17.4s, v17.4s, v31.4s\n"
- "sqxtn v15.4h, v15.4s\n"
- "srshl v20.4s, v20.4s, v31.4s\n"
- "sqxtn v10.4h, v10.4s\n"
- "srshl v23.4s, v23.4s, v31.4s\n"
- "sqxtn v9.4h, v9.4s\n"
- "srshl v22.4s, v22.4s, v31.4s\n"
+ "smlal2 v0.4s, v22.8h, v4.8h\n"
+ "smlal2 v31.4s, v27.8h, v4.8h\n"
+ "sqrdmulh v19.4s, v19.4s, v30.4s\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "and v12.16b, v21.16b, v29.16b\n"
+ "sqrdmulh v8.4s, v8.4s, v14.4s\n"
+ "and v23.16b, v20.16b, v29.16b\n"
+ "sqrdmulh v0.4s, v0.4s, v14.4s\n"
+ "and v9.16b, v19.16b, v29.16b\n"
+ "sqrdmulh v31.4s, v31.4s, v14.4s\n"
+ "sqadd v3.4s, v3.4s, v16.4s\n"
+ "sshr v12.4s, v12.4s, #0x1f\n"
+ "and v18.16b, v8.16b, v25.16b\n"
+ "sshr v23.4s, v23.4s, #0x1f\n"
+ "and v22.16b, v0.16b, v25.16b\n"
+ "sshr v9.4s, v9.4s, #0x1f\n"
+ "and v16.16b, v31.16b, v25.16b\n"
+ "sqadd v21.4s, v21.4s, v12.4s\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sqadd v20.4s, v20.4s, v23.4s\n"
+ "sshr v22.4s, v22.4s, #0x1f\n"
+ "sqadd v19.4s, v19.4s, v9.4s\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "srshl v5.4s, v5.4s, v29.4s\n"
+ "srshl v21.4s, v21.4s, v29.4s\n"
+ "sqadd v8.4s, v8.4s, v18.4s\n"
+ "srshl v20.4s, v20.4s, v29.4s\n"
+ "sqadd v0.4s, v0.4s, v22.4s\n"
+ "srshl v19.4s, v19.4s, v29.4s\n"
+ "sqadd v31.4s, v31.4s, v16.4s\n"
+ "srshl v3.4s, v3.4s, v25.4s\n"
+ "sqxtn v5.4h, v5.4s\n"
+ "srshl v8.4s, v8.4s, v25.4s\n"
"sqxtn v21.4h, v21.4s\n"
- "sqxtn2 v15.8h, v17.4s\n"
- "sqxtn2 v10.8h, v20.4s\n"
- "sqxtn2 v9.8h, v23.4s\n"
- "sqxtn2 v21.8h, v22.4s\n"
- "sqadd v15.8h, v15.8h, v11.8h\n"
- "sqadd v10.8h, v10.8h, v11.8h\n"
- "sqadd v9.8h, v9.8h, v11.8h\n"
- "sqadd v21.8h, v21.8h, v11.8h\n"
- "smax v15.8h, v15.8h, v16.8h\n"
- "smax v10.8h, v10.8h, v16.8h\n"
- "smax v9.8h, v9.8h, v16.8h\n"
- "smax v21.8h, v21.8h, v16.8h\n"
- "smin v15.8h, v15.8h, v14.8h\n"
- "smin v10.8h, v10.8h, v14.8h\n"
- "smin v9.8h, v9.8h, v14.8h\n"
- "smin v21.8h, v21.8h, v14.8h\n"
- "uzp1 v15.16b, v15.16b, v15.16b\n"
- "str d15, [x11, x16]\n"
- "uzp1 v10.16b, v10.16b, v10.16b\n"
- "uzp1 v9.16b, v9.16b, v9.16b\n"
- "str d10, [x10, x16]\n"
+ "srshl v0.4s, v0.4s, v25.4s\n"
+ "sqxtn v20.4h, v20.4s\n"
+ "srshl v31.4s, v31.4s, v25.4s\n"
+ "sqxtn v19.4h, v19.4s\n"
+ "sqxtn2 v5.8h, v3.4s\n"
+ "sqxtn2 v21.8h, v8.4s\n"
+ "sqxtn2 v20.8h, v0.4s\n"
+ "sqxtn2 v19.8h, v31.4s\n"
+ "sqadd v5.8h, v5.8h, v13.8h\n"
+ "sqadd v21.8h, v21.8h, v13.8h\n"
+ "sqadd v20.8h, v20.8h, v13.8h\n"
+ "sqadd v19.8h, v19.8h, v13.8h\n"
+ "smax v5.8h, v5.8h, v17.8h\n"
+ "smax v21.8h, v21.8h, v17.8h\n"
+ "smax v20.8h, v20.8h, v17.8h\n"
+ "smax v19.8h, v19.8h, v17.8h\n"
+ "smin v5.8h, v5.8h, v24.8h\n"
+ "smin v21.8h, v21.8h, v24.8h\n"
+ "smin v20.8h, v20.8h, v24.8h\n"
+ "smin v19.8h, v19.8h, v24.8h\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "str d5, [x11, x16]\n"
"uzp1 v21.16b, v21.16b, v21.16b\n"
- "str d9, [x9, x16]\n"
- "str d21, [x28, x16]\n"
- "ldr q15, [x24, #0x0]\n"
- "ldr q17, [x24, #0x10]\n"
- "add x24, x24, #0x20\n"
- "ldr d0, [x14, #0x0]\n"
- "ldr d1, [x14, #0x8]\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "str d21, [x10, x16]\n"
+ "uzp1 v19.16b, v19.16b, v19.16b\n"
+ "str d20, [x9, x16]\n"
+ "str d19, [x28, x16]\n"
+ "ldr q5, [x20, #0x0]\n"
+ "ldr q3, [x20, #0x10]\n"
+ "add x20, x20, #0x20\n"
+ "ldr d11, [x14, #0x0]\n"
+ "ldr d22, [x14, #0x8]\n"
"add x16, x16, #0x8\n"
- "str x24, [%x[params], %[offsetof_Params_bias]]\n"
- "ldr d2, [x14, #0x10]\n"
- "ldr d3, [x14, #0x18]\n"
- "mov v10.16b, v15.16b\n"
- "mov v20.16b, v17.16b\n"
- "ldr d4, [x14, #0x20]\n"
- "ldr d5, [x14, #0x28]\n"
- "mov v9.16b, v15.16b\n"
- "mov v23.16b, v17.16b\n"
- "ldr d6, [x14, #0x30]\n"
+ "str x20, [%x[params], %[offsetof_Params_bias]]\n"
+ "ldr d14, [x14, #0x10]\n"
+ "ldr d28, [x14, #0x18]\n"
+ "mov v21.16b, v5.16b\n"
+ "mov v8.16b, v3.16b\n"
+ "ldr d18, [x14, #0x20]\n"
+ "ldr d9, [x14, #0x28]\n"
+ "mov v20.16b, v5.16b\n"
+ "mov v0.16b, v3.16b\n"
+ "ldr d26, [x14, #0x30]\n"
"ldr d7, [x14, #0x38]\n"
- "mov v21.16b, v15.16b\n"
- "mov v22.16b, v17.16b\n"
- "ldr d8, [x14, #0x40]\n"
+ "mov v19.16b, v5.16b\n"
+ "mov v31.16b, v3.16b\n"
+ "ldr d4, [x14, #0x40]\n"
"ldp x27, x26, [x15, #0x0]\n"
- "usubl v0.8h, v0.8b, v13.8b\n"
- "usubl v1.8h, v1.8b, v13.8b\n"
+ "usubl v11.8h, v11.8b, v15.8b\n"
+ "usubl v22.8h, v22.8b, v15.8b\n"
"ldp x25, x24, [x15, #0x10]\n"
"ldp x23, x22, [x15, #0x20]\n"
- "usubl v2.8h, v2.8b, v13.8b\n"
- "usubl v3.8h, v3.8b, v13.8b\n"
+ "usubl v14.8h, v14.8b, v15.8b\n"
+ "usubl v28.8h, v28.8b, v15.8b\n"
"ldp x21, x20, [x15, #0x30]\n"
- "ldr d31, [x27, x17]\n"
- "usubl v4.8h, v4.8b, v13.8b\n"
- "usubl v5.8h, v5.8b, v13.8b\n"
- "ldr d30, [x26, x17]\n"
- "ldr d29, [x25, x17]\n"
- "usubl v6.8h, v6.8b, v13.8b\n"
- "usubl v7.8h, v7.8b, v13.8b\n"
- "ldr d28, [x24, x17]\n"
- "ldr d27, [x23, x17]\n"
- "usubl v8.8h, v8.8b, v13.8b\n"
- "usubl v31.8h, v31.8b, v12.8b\n"
- "ldr d26, [x22, x17]\n"
- "ldr d25, [x21, x17]\n"
- "usubl v30.8h, v30.8b, v12.8b\n"
- "usubl v29.8h, v29.8b, v12.8b\n"
- "ldr d24, [x20, x17]\n"
- "usubl v28.8h, v28.8b, v12.8b\n"
- "usubl v27.8h, v27.8b, v12.8b\n"
- "usubl v26.8h, v26.8b, v12.8b\n"
- "usubl v25.8h, v25.8b, v12.8b\n"
- "usubl v24.8h, v24.8b, v12.8b\n"
+ "ldr d25, [x27, x17]\n"
+ "usubl v18.8h, v18.8b, v15.8b\n"
+ "usubl v9.8h, v9.8b, v15.8b\n"
+ "ldr d27, [x26, x17]\n"
+ "ldr d1, [x25, x17]\n"
+ "usubl v26.8h, v26.8b, v15.8b\n"
+ "usubl v7.8h, v7.8b, v15.8b\n"
+ "ldr d2, [x24, x17]\n"
+ "ldr d12, [x23, x17]\n"
+ "usubl v4.8h, v4.8b, v15.8b\n"
+ "usubl v25.8h, v25.8b, v6.8b\n"
+ "ldr d16, [x22, x17]\n"
+ "ldr d23, [x21, x17]\n"
+ "usubl v27.8h, v27.8b, v6.8b\n"
+ "usubl v1.8h, v1.8b, v6.8b\n"
+ "ldr d10, [x20, x17]\n"
+ "usubl v2.8h, v2.8b, v6.8b\n"
+ "usubl v12.8h, v12.8b, v6.8b\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "usubl v23.8h, v23.8b, v6.8b\n"
+ "usubl v10.8h, v10.8b, v6.8b\n"
"bgt 1b\n"
"2:" // Tail
- "smlal v15.4s, v31.4h, v8.4h\n"
- "smlal2 v17.4s, v31.8h, v8.8h\n"
- "ldr x24, [x15, #0x40]\n"
- "ldr x22, [x15, #0x48]\n"
- "smlal v10.4s, v31.4h, v6.4h\n"
- "smlal2 v20.4s, v31.8h, v6.8h\n"
- "ldr x21, [x15, #0x50]\n"
- "ldr x20, [x15, #0x58]\n"
- "smlal v15.4s, v30.4h, v0.4h\n"
- "smlal2 v17.4s, v30.8h, v0.8h\n"
- "ldr q19, [x13, #0x0]\n"
- "ldr x23, [x15, #0x78]\n"
- "smlal v10.4s, v28.4h, v1.4h\n"
- "smlal2 v20.4s, v28.8h, v1.8h\n"
- "ldr d28, [x22, x17]\n"
- "usubl v28.8h, v28.8b, v12.8b\n"
- "smlal v15.4s, v29.4h, v1.4h\n"
- "smlal2 v17.4s, v29.8h, v1.8h\n"
- "ldr d29, [x24, x17]\n"
- "usubl v29.8h, v29.8b, v12.8b\n"
- "smlal v10.4s, v27.4h, v2.4h\n"
- "smlal2 v20.4s, v27.8h, v2.8h\n"
+ "ldr q29, [x13, #0x0]\n"
+ "ldr q30, [x12, #0x0]\n"
+ "smlal v5.4s, v25.4h, v4.4h\n"
+ "smlal2 v3.4s, v25.8h, v4.8h\n"
+ "ldr x21, [x15, #0x58]\n"
+ "ldr x20, [x15, #0x78]\n"
+ "smlal v5.4s, v27.4h, v11.4h\n"
+ "smlal v21.4s, v25.4h, v26.4h\n"
+ "ldr x25, [x15, #0x60]\n"
+ "ldr x24, [x15, #0x80]\n"
+ "smlal v20.4s, v25.4h, v14.4h\n"
+ "smlal v19.4s, v25.4h, v11.4h\n"
+ "smlal2 v3.4s, v27.8h, v11.8h\n"
"ldr d27, [x21, x17]\n"
- "usubl v27.8h, v27.8b, v12.8b\n"
- "smlal v15.4s, v26.4h, v3.4h\n"
- "smlal2 v17.4s, v26.8h, v3.8h\n"
- "ldr d26, [x20, x17]\n"
- "ldr x20, [x15, #0x60]\n"
- "smlal v10.4s, v24.4h, v0.4h\n"
- "smlal2 v20.4s, v24.8h, v0.8h\n"
- "usubl v26.8h, v26.8b, v12.8b\n"
- "ldr x21, [x15, #0x80]\n"
- "smlal v15.4s, v25.4h, v4.4h\n"
- "smlal2 v17.4s, v25.8h, v4.8h\n"
- "ldr d25, [x20, x17]\n"
- "ldr x20, [x15, #0x68]\n"
- "smlal v10.4s, v29.4h, v4.4h\n"
- "smlal2 v20.4s, v29.8h, v4.8h\n"
- "ldr d29, [x20, x17]\n"
- "usubl v25.8h, v25.8b, v12.8b\n"
- "smlal v15.4s, v24.4h, v2.4h\n"
- "smlal2 v17.4s, v24.8h, v2.8h\n"
- "ldr q18, [x12, #0x0]\n"
+ "usubl v27.8h, v27.8b, v6.8b\n"
+ "smlal v5.4s, v1.4h, v22.4h\n"
+ "smlal2 v8.4s, v25.8h, v26.8h\n"
+ "smlal2 v0.4s, v25.8h, v14.8h\n"
+ "ldr x23, [x15, #0x68]\n"
"ldr x22, [x15, #0x88]\n"
- "smlal v10.4s, v28.4h, v5.4h\n"
- "smlal2 v20.4s, v28.8h, v5.8h\n"
- "ldr d28, [x21, x17]\n"
- "ldr x21, [x15, #0x70]\n"
- "smlal v9.4s, v31.4h, v2.4h\n"
- "smlal2 v23.4s, v31.8h, v2.8h\n"
- "usubl v28.8h, v28.8b, v12.8b\n"
+ "smlal2 v31.4s, v25.8h, v11.8h\n"
+ "ldr d25, [x20, x17]\n"
+ "usubl v25.8h, v25.8b, v6.8b\n"
+ "smlal v21.4s, v2.4h, v22.4h\n"
+ "smlal v20.4s, v27.4h, v28.4h\n"
+ "smlal v19.4s, v25.4h, v18.4h\n"
+ "ldr x21, [x15, #0x40]\n"
+ "ldr x20, [x15, #0x70]\n"
+ "smlal2 v3.4s, v1.8h, v22.8h\n"
+ "ldr d1, [x25, x17]\n"
+ "usubl v1.8h, v1.8b, v6.8b\n"
+ "smlal v5.4s, v16.4h, v28.4h\n"
+ "smlal2 v8.4s, v2.8h, v22.8h\n"
+ "ldr d2, [x24, x17]\n"
+ "usubl v2.8h, v2.8b, v6.8b\n"
+ "smlal2 v0.4s, v27.8h, v28.8h\n"
+ "ldr d27, [x23, x17]\n"
+ "smlal2 v31.4s, v25.8h, v18.8h\n"
+ "ldr d25, [x22, x17]\n"
+ "smlal v21.4s, v12.4h, v14.4h\n"
"ldr x25, [x15, #0x98]\n"
- "smlal v15.4s, v27.4h, v5.4h\n"
- "smlal2 v17.4s, v27.8h, v5.8h\n"
- "usubl v29.8h, v29.8b, v12.8b\n"
- "ldr x24, [x15, #0x90]\n"
- "smlal v10.4s, v27.4h, v3.4h\n"
- "smlal2 v20.4s, v27.8h, v3.8h\n"
+ "smlal v20.4s, v1.4h, v11.4h\n"
+ "smlal v19.4s, v2.4h, v22.4h\n"
+ "ldr x24, [x15, #0x50]\n"
+ "smlal2 v3.4s, v16.8h, v28.8h\n"
+ "ldr d16, [x21, x17]\n"
+ "usubl v27.8h, v27.8b, v6.8b\n"
+ "smlal v5.4s, v23.4h, v18.4h\n"
+ "usubl v25.8h, v25.8b, v6.8b\n"
+ "smlal2 v8.4s, v12.8h, v14.8h\n"
+ "ldr d12, [x20, x17]\n"
+ "ldr x23, [x15, #0x48]\n"
+ "smlal2 v0.4s, v1.8h, v11.8h\n"
+ "smlal2 v31.4s, v2.8h, v22.8h\n"
+ "ldr x21, [x15, #0x90]\n"
+ "ldr x20, [x15, #0xa8]\n"
+ "smlal v21.4s, v10.4h, v11.4h\n"
+ "smlal v20.4s, v27.4h, v18.4h\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x22, [x15, #0xa0]\n"
+ "smlal v19.4s, v25.4h, v9.4h\n"
+ "smlal2 v3.4s, v23.8h, v18.8h\n"
+ "ldr d23, [x25, x17]\n"
+ "usubl v12.8h, v12.8b, v6.8b\n"
+ "usubl v23.8h, v23.8b, v6.8b\n"
+ "smlal v5.4s, v10.4h, v14.4h\n"
+ "smlal2 v8.4s, v10.8h, v11.8h\n"
+ "ldr d11, [x24, x17]\n"
+ "usubl v11.8h, v11.8b, v6.8b\n"
+ "smlal2 v0.4s, v27.8h, v18.8h\n"
"ldr d27, [x23, x17]\n"
- "usubl v27.8h, v27.8b, v12.8b\n"
- "smlal v21.4s, v31.4h, v0.4h\n"
- "smlal v9.4s, v26.4h, v3.4h\n"
- "ldr x23, [x15, #0xa8]\n"
- "ldr x20, [x15, #0xa0]\n"
- "smlal2 v23.4s, v26.8h, v3.8h\n"
- "ldr d26, [x22, x17]\n"
- "smlal2 v22.4s, v31.8h, v0.8h\n"
- "ldr d24, [x21, x17]\n"
- "smlal v21.4s, v27.4h, v4.4h\n"
- "smlal v9.4s, v25.4h, v0.4h\n"
- "usubl v26.8h, v26.8b, v12.8b\n"
- "ldr x22, [x15, #0xb0]\n"
- "smlal2 v23.4s, v25.8h, v0.8h\n"
- "ldr q30, [x13, #0x10]\n"
- "smlal2 v22.4s, v27.8h, v4.8h\n"
- "ldr d27, [x20, x17]\n"
- "smlal v21.4s, v28.4h, v1.4h\n"
- "smlal v15.4s, v25.4h, v6.4h\n"
- "usubl v24.8h, v24.8b, v12.8b\n"
- "ldr x21, [x15, #0xb8]\n"
- "smlal2 v17.4s, v25.8h, v6.8h\n"
- "ldr d25, [x24, x17]\n"
- "smlal v9.4s, v29.4h, v4.4h\n"
- "usubl v25.8h, v25.8b, v12.8b\n"
- "smlal2 v23.4s, v29.8h, v4.8h\n"
- "ldr d29, [x25, x17]\n"
- "ldr q31, [x12, #0x10]\n"
- "smlal2 v22.4s, v28.8h, v1.8h\n"
- "smlal v21.4s, v26.4h, v5.4h\n"
- "usubl v29.8h, v29.8b, v12.8b\n"
- "smlal v15.4s, v24.4h, v7.4h\n"
+ "smlal2 v31.4s, v25.8h, v9.8h\n"
+ "ldr d25, [x21, x17]\n"
+ "ldr x21, [x15, #0xb0]\n"
+ "smlal v21.4s, v16.4h, v18.4h\n"
+ "smlal v20.4s, v12.4h, v22.4h\n"
+ "smlal v19.4s, v23.4h, v14.4h\n"
+ "smlal2 v3.4s, v10.8h, v14.8h\n"
+ "ldr d10, [x20, x17]\n"
+ "usubl v27.8h, v27.8b, v6.8b\n"
+ "usubl v25.8h, v25.8b, v6.8b\n"
+ "usubl v10.8h, v10.8b, v6.8b\n"
+ "smlal v5.4s, v11.4h, v9.4h\n"
+ "ldr x20, [x15, #0xb8]\n"
+ "smlal2 v8.4s, v16.8h, v18.8h\n"
+ "ldr d16, [x22, x17]\n"
+ "ldr d18, [x21, x17]\n"
+ "smlal2 v0.4s, v12.8h, v22.8h\n"
+ "ldr d22, [x20, x17]\n"
+ "smlal2 v31.4s, v23.8h, v14.8h\n"
+ "ldr q14, [x13, #0x10]\n"
+ "smlal v21.4s, v27.4h, v9.4h\n"
+ "smlal v20.4s, v25.4h, v26.4h\n"
+ "smlal v19.4s, v10.4h, v28.4h\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
"ldr x20, [x15, #0xc0]\n"
- "smlal2 v17.4s, v24.8h, v7.8h\n"
- "smlal v9.4s, v24.4h, v1.4h\n"
- "usubl v27.8h, v27.8b, v12.8b\n"
+ "smlal2 v3.4s, v11.8h, v9.8h\n"
+ "usubl v18.8h, v18.8b, v6.8b\n"
+ "smlal v5.4s, v1.4h, v26.4h\n"
"tst x7, #0x7\n"
- "smlal2 v23.4s, v24.8h, v1.8h\n"
- "ldr d24, [x23, x17]\n"
- "smlal2 v22.4s, v26.8h, v5.8h\n"
- "ldr d26, [x22, x17]\n"
- "smlal v21.4s, v29.4h, v2.4h\n"
- "usubl v24.8h, v24.8b, v12.8b\n"
- "smlal2 v22.4s, v29.8h, v2.8h\n"
+ "smlal2 v8.4s, v27.8h, v9.8h\n"
+ "ldr d27, [x20, x17]\n"
+ "smlal2 v0.4s, v25.8h, v26.8h\n"
+ "ldr q25, [x12, #0x10]\n"
+ "smlal2 v31.4s, v10.8h, v28.8h\n"
+ "smlal v21.4s, v11.4h, v28.4h\n"
+ "usubl v22.8h, v22.8b, v6.8b\n"
+ "add x17, x17, #0x8\n"
+ "smlal v20.4s, v16.4h, v7.4h\n"
+ "smlal v19.4s, v18.4h, v7.4h\n"
+ "usubl v27.8h, v27.8b, v6.8b\n"
"add x13, x13, #0x20\n"
- "smlal v9.4s, v25.4h, v6.4h\n"
- "smlal v21.4s, v24.4h, v3.4h\n"
- "usubl v26.8h, v26.8b, v12.8b\n"
+ "smlal2 v3.4s, v1.8h, v26.8h\n"
+ "smlal v5.4s, v12.4h, v7.4h\n"
+ "sqrdmulh v5.4s, v5.4s, v29.4s\n"
"add x12, x12, #0x20\n"
- "smlal v10.4s, v28.4h, v7.4h\n"
- "smlal2 v20.4s, v28.8h, v7.8h\n"
- "sqrdmulh v15.4s, v15.4s, v19.4s\n"
- "smlal2 v23.4s, v25.8h, v6.8h\n"
- "ldr d25, [x21, x17]\n"
- "smlal2 v22.4s, v24.8h, v3.8h\n"
- "usubl v25.8h, v25.8b, v12.8b\n"
- "smlal v9.4s, v27.4h, v7.4h\n"
- "smlal v21.4s, v26.4h, v7.4h\n"
- "and v0.16b, v15.16b, v18.16b\n"
- "smlal v10.4s, v29.4h, v8.4h\n"
- "smlal2 v20.4s, v29.8h, v8.8h\n"
- "ldr d29, [x20, x17]\n"
- "usubl v29.8h, v29.8b, v12.8b\n"
- "smlal2 v23.4s, v27.8h, v7.8h\n"
- "smlal2 v22.4s, v26.8h, v7.8h\n"
- "sqrdmulh v17.4s, v17.4s, v30.4s\n"
- "add x17, x17, #0x8\n"
- "smlal v9.4s, v24.4h, v5.4h\n"
- "smlal v21.4s, v25.4h, v6.4h\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "smlal2 v23.4s, v24.8h, v5.8h\n"
- "smlal2 v22.4s, v25.8h, v6.8h\n"
- "and v7.16b, v17.16b, v31.16b\n"
- "smlal v9.4s, v25.4h, v8.4h\n"
- "smlal v21.4s, v29.4h, v8.4h\n"
- "sqrdmulh v10.4s, v10.4s, v19.4s\n"
- "smlal2 v23.4s, v25.8h, v8.8h\n"
- "smlal2 v22.4s, v29.8h, v8.8h\n"
- "sqrdmulh v9.4s, v9.4s, v19.4s\n"
- "sqrdmulh v21.4s, v21.4s, v19.4s\n"
- "sqadd v15.4s, v15.4s, v0.4s\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "and v19.16b, v10.16b, v18.16b\n"
- "sqrdmulh v20.4s, v20.4s, v30.4s\n"
- "and v27.16b, v9.16b, v18.16b\n"
- "sqrdmulh v23.4s, v23.4s, v30.4s\n"
- "and v0.16b, v21.16b, v18.16b\n"
- "sqrdmulh v22.4s, v22.4s, v30.4s\n"
- "sqadd v17.4s, v17.4s, v7.4s\n"
- "sshr v19.4s, v19.4s, #0x1f\n"
- "and v5.16b, v20.16b, v31.16b\n"
+ "smlal2 v8.4s, v11.8h, v28.8h\n"
+ "smlal2 v0.4s, v16.8h, v7.8h\n"
+ "and v16.16b, v5.16b, v30.16b\n"
+ "smlal2 v31.4s, v18.8h, v7.8h\n"
+ "smlal v21.4s, v2.4h, v7.4h\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "smlal v20.4s, v10.4h, v9.4h\n"
+ "smlal v19.4s, v22.4h, v26.4h\n"
+ "sqadd v5.4s, v5.4s, v16.4s\n"
+ "smlal2 v3.4s, v12.8h, v7.8h\n"
+ "smlal2 v8.4s, v2.8h, v7.8h\n"
+ "sqrdmulh v3.4s, v3.4s, v14.4s\n"
+ "smlal2 v0.4s, v10.8h, v9.8h\n"
+ "smlal2 v31.4s, v22.8h, v26.8h\n"
+ "and v16.16b, v3.16b, v25.16b\n"
+ "smlal v21.4s, v23.4h, v4.4h\n"
+ "smlal v20.4s, v22.4h, v4.4h\n"
+ "sqrdmulh v21.4s, v21.4s, v29.4s\n"
+ "smlal v19.4s, v27.4h, v4.4h\n"
+ "smlal2 v8.4s, v23.8h, v4.8h\n"
+ "sqrdmulh v20.4s, v20.4s, v29.4s\n"
+ "smlal2 v0.4s, v22.8h, v4.8h\n"
+ "smlal2 v31.4s, v27.8h, v4.8h\n"
+ "sqrdmulh v19.4s, v19.4s, v29.4s\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "and v23.16b, v21.16b, v30.16b\n"
+ "sqrdmulh v8.4s, v8.4s, v14.4s\n"
+ "and v27.16b, v20.16b, v30.16b\n"
+ "sqrdmulh v0.4s, v0.4s, v14.4s\n"
+ "and v22.16b, v19.16b, v30.16b\n"
+ "sqrdmulh v31.4s, v31.4s, v14.4s\n"
+ "sqadd v3.4s, v3.4s, v16.4s\n"
+ "sshr v23.4s, v23.4s, #0x1f\n"
+ "and v14.16b, v8.16b, v25.16b\n"
"sshr v27.4s, v27.4s, #0x1f\n"
- "and v4.16b, v23.16b, v31.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "and v7.16b, v22.16b, v31.16b\n"
- "sqadd v10.4s, v10.4s, v19.4s\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sqadd v9.4s, v9.4s, v27.4s\n"
- "sshr v4.4s, v4.4s, #0x1f\n"
- "sqadd v21.4s, v21.4s, v0.4s\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "srshl v15.4s, v15.4s, v18.4s\n"
- "srshl v10.4s, v10.4s, v18.4s\n"
- "sqadd v20.4s, v20.4s, v5.4s\n"
- "srshl v9.4s, v9.4s, v18.4s\n"
- "sqadd v23.4s, v23.4s, v4.4s\n"
- "srshl v21.4s, v21.4s, v18.4s\n"
- "sqadd v22.4s, v22.4s, v7.4s\n"
- "srshl v17.4s, v17.4s, v31.4s\n"
- "sqxtn v15.4h, v15.4s\n"
- "srshl v20.4s, v20.4s, v31.4s\n"
- "sqxtn v10.4h, v10.4s\n"
- "srshl v23.4s, v23.4s, v31.4s\n"
- "sqxtn v9.4h, v9.4s\n"
- "srshl v22.4s, v22.4s, v31.4s\n"
+ "and v18.16b, v0.16b, v25.16b\n"
+ "sshr v22.4s, v22.4s, #0x1f\n"
+ "and v16.16b, v31.16b, v25.16b\n"
+ "sqadd v21.4s, v21.4s, v23.4s\n"
+ "sshr v14.4s, v14.4s, #0x1f\n"
+ "sqadd v20.4s, v20.4s, v27.4s\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sqadd v19.4s, v19.4s, v22.4s\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "srshl v5.4s, v5.4s, v30.4s\n"
+ "srshl v21.4s, v21.4s, v30.4s\n"
+ "sqadd v8.4s, v8.4s, v14.4s\n"
+ "srshl v20.4s, v20.4s, v30.4s\n"
+ "sqadd v0.4s, v0.4s, v18.4s\n"
+ "srshl v19.4s, v19.4s, v30.4s\n"
+ "sqadd v31.4s, v31.4s, v16.4s\n"
+ "srshl v3.4s, v3.4s, v25.4s\n"
+ "sqxtn v5.4h, v5.4s\n"
+ "srshl v8.4s, v8.4s, v25.4s\n"
"sqxtn v21.4h, v21.4s\n"
- "sqxtn2 v15.8h, v17.4s\n"
- "sqxtn2 v10.8h, v20.4s\n"
- "sqxtn2 v9.8h, v23.4s\n"
- "sqxtn2 v21.8h, v22.4s\n"
- "sqadd v15.8h, v15.8h, v11.8h\n"
- "sqadd v10.8h, v10.8h, v11.8h\n"
- "sqadd v9.8h, v9.8h, v11.8h\n"
- "sqadd v21.8h, v21.8h, v11.8h\n"
- "smax v15.8h, v15.8h, v16.8h\n"
- "smax v10.8h, v10.8h, v16.8h\n"
- "smax v9.8h, v9.8h, v16.8h\n"
- "smax v21.8h, v21.8h, v16.8h\n"
- "smin v15.8h, v15.8h, v14.8h\n"
- "smin v10.8h, v10.8h, v14.8h\n"
- "smin v9.8h, v9.8h, v14.8h\n"
- "smin v21.8h, v21.8h, v14.8h\n"
- "uzp1 v15.16b, v15.16b, v15.16b\n"
- "str d15, [x11, x16]\n"
- "uzp1 v10.16b, v10.16b, v10.16b\n"
- "uzp1 v9.16b, v9.16b, v9.16b\n"
- "str d10, [x10, x16]\n"
+ "srshl v0.4s, v0.4s, v25.4s\n"
+ "sqxtn v20.4h, v20.4s\n"
+ "srshl v31.4s, v31.4s, v25.4s\n"
+ "sqxtn v19.4h, v19.4s\n"
+ "sqxtn2 v5.8h, v3.4s\n"
+ "sqxtn2 v21.8h, v8.4s\n"
+ "sqxtn2 v20.8h, v0.4s\n"
+ "sqxtn2 v19.8h, v31.4s\n"
+ "sqadd v5.8h, v5.8h, v13.8h\n"
+ "sqadd v21.8h, v21.8h, v13.8h\n"
+ "sqadd v20.8h, v20.8h, v13.8h\n"
+ "sqadd v19.8h, v19.8h, v13.8h\n"
+ "smax v5.8h, v5.8h, v17.8h\n"
+ "smax v21.8h, v21.8h, v17.8h\n"
+ "smax v20.8h, v20.8h, v17.8h\n"
+ "smax v19.8h, v19.8h, v17.8h\n"
+ "smin v5.8h, v5.8h, v24.8h\n"
+ "smin v21.8h, v21.8h, v24.8h\n"
+ "smin v20.8h, v20.8h, v24.8h\n"
+ "smin v19.8h, v19.8h, v24.8h\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
+ "str d5, [x11, x16]\n"
"uzp1 v21.16b, v21.16b, v21.16b\n"
- "str d9, [x9, x16]\n"
- "str d21, [x28, x16]\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "str d21, [x10, x16]\n"
+ "uzp1 v19.16b, v19.16b, v19.16b\n"
+ "str d20, [x9, x16]\n"
+ "str d19, [x28, x16]\n"
"add x16, x16, #0x8\n"
"beq 88f\n"
"add x14, x14, #0x48\n"
"3:" // Oddments
- "ldr x24, [%x[params], %[offsetof_Params_bias]]\n"
+ "ldr x20, [%x[params], %[offsetof_Params_bias]]\n"
"tbz x7, #2, 5f\n"
- "ld1 { v15.4s }, [x24], #0x10\n"
+ "ld1 { v5.4s }, [x20], #0x10\n"
"tbz x7, #1, 4f\n"
- "ld1 { v17.d }[0], [x24], #0x8\n"
+ "ld1 { v3.d }[0], [x20], #0x8\n"
"tbz x7, #0, 7f\n"
- "ld1 { v17.s }[2], [x24]\n"
+ "ld1 { v3.s }[2], [x20]\n"
"b 7f\n"
"4:" // Oddments: Load bias: Bit 2: Bit 1: Unset
"tbz x7, #0, 7f\n"
- "ld1 { v17.s }[0], [x24]\n"
+ "ld1 { v3.s }[0], [x20]\n"
"b 7f\n"
"5:" // Oddments: Load bias: Bit 2: Unset
"tbz x7, #1, 6f\n"
- "ld1 { v15.d }[0], [x24], #0x8\n"
+ "ld1 { v5.d }[0], [x20], #0x8\n"
"tbz x7, #0, 7f\n"
- "ld1 { v15.s }[2], [x24]\n"
+ "ld1 { v5.s }[2], [x20]\n"
"b 7f\n"
"6:" // Oddments: Load bias: Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 7f\n"
- "ld1 { v15.s }[0], [x24]\n"
+ "ld1 { v5.s }[0], [x20]\n"
"7:" // Oddments: Load bias: Bit 2: End
- "ldr d0, [x14, #0x0]\n"
- "ldr d1, [x14, #0x8]\n"
- "mov v10.16b, v15.16b\n"
- "mov v20.16b, v17.16b\n"
- "ldr d2, [x14, #0x10]\n"
- "ldr d3, [x14, #0x18]\n"
- "mov v9.16b, v15.16b\n"
- "mov v23.16b, v17.16b\n"
- "ldr d4, [x14, #0x20]\n"
- "ldr d5, [x14, #0x28]\n"
- "mov v21.16b, v15.16b\n"
- "mov v22.16b, v17.16b\n"
- "ldr d6, [x14, #0x30]\n"
+ "ldr d11, [x14, #0x0]\n"
+ "ldr d22, [x14, #0x8]\n"
+ "mov v21.16b, v5.16b\n"
+ "mov v8.16b, v3.16b\n"
+ "ldr d14, [x14, #0x10]\n"
+ "ldr d28, [x14, #0x18]\n"
+ "mov v20.16b, v5.16b\n"
+ "mov v0.16b, v3.16b\n"
+ "ldr d18, [x14, #0x20]\n"
+ "ldr d9, [x14, #0x28]\n"
+ "mov v19.16b, v5.16b\n"
+ "mov v31.16b, v3.16b\n"
+ "ldr d26, [x14, #0x30]\n"
"ldr d7, [x14, #0x38]\n"
- "usubl v0.8h, v0.8b, v13.8b\n"
- "usubl v1.8h, v1.8b, v13.8b\n"
- "ldr d8, [x14, #0x40]\n"
+ "usubl v11.8h, v11.8b, v15.8b\n"
+ "usubl v22.8h, v22.8b, v15.8b\n"
+ "ldr d4, [x14, #0x40]\n"
"ldp x27, x26, [x15, #0x0]\n"
- "usubl v2.8h, v2.8b, v13.8b\n"
- "usubl v3.8h, v3.8b, v13.8b\n"
+ "usubl v14.8h, v14.8b, v15.8b\n"
+ "usubl v28.8h, v28.8b, v15.8b\n"
"ldp x25, x24, [x15, #0x10]\n"
"ldp x23, x22, [x15, #0x20]\n"
- "usubl v4.8h, v4.8b, v13.8b\n"
- "usubl v5.8h, v5.8b, v13.8b\n"
+ "usubl v18.8h, v18.8b, v15.8b\n"
+ "usubl v9.8h, v9.8b, v15.8b\n"
"ldp x21, x20, [x15, #0x30]\n"
- "usubl v6.8h, v6.8b, v13.8b\n"
- "usubl v7.8h, v7.8b, v13.8b\n"
- "usubl v8.8h, v8.8b, v13.8b\n"
+ "usubl v26.8h, v26.8b, v15.8b\n"
+ "usubl v7.8h, v7.8b, v15.8b\n"
+ "usubl v4.8h, v4.8b, v15.8b\n"
"add x27, x27, x17\n"
"add x26, x26, x17\n"
"add x25, x25, x17\n"
@@ -689,700 +689,700 @@ void a64_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
"add x21, x21, x17\n"
"add x20, x20, x17\n"
"tbz x7, #2, 9f\n"
- "ld1 { v31.s }[0], [x27], #0x4\n"
- "ld1 { v30.s }[0], [x26], #0x4\n"
- "ld1 { v29.s }[0], [x25], #0x4\n"
- "ld1 { v28.s }[0], [x24], #0x4\n"
- "ld1 { v27.s }[0], [x23], #0x4\n"
- "ld1 { v26.s }[0], [x22], #0x4\n"
- "ld1 { v25.s }[0], [x21], #0x4\n"
- "ld1 { v24.s }[0], [x20], #0x4\n"
+ "ld1 { v25.s }[0], [x27], #0x4\n"
+ "ld1 { v27.s }[0], [x26], #0x4\n"
+ "ld1 { v1.s }[0], [x25], #0x4\n"
+ "ld1 { v2.s }[0], [x24], #0x4\n"
+ "ld1 { v12.s }[0], [x23], #0x4\n"
+ "ld1 { v16.s }[0], [x22], #0x4\n"
+ "ld1 { v23.s }[0], [x21], #0x4\n"
+ "ld1 { v10.s }[0], [x20], #0x4\n"
"tbz x7, #1, 8f\n"
- "ld1 { v31.h }[2], [x27], #0x2\n"
- "ld1 { v30.h }[2], [x26], #0x2\n"
- "ld1 { v29.h }[2], [x25], #0x2\n"
- "ld1 { v28.h }[2], [x24], #0x2\n"
- "ld1 { v27.h }[2], [x23], #0x2\n"
- "ld1 { v26.h }[2], [x22], #0x2\n"
- "ld1 { v25.h }[2], [x21], #0x2\n"
- "ld1 { v24.h }[2], [x20], #0x2\n"
+ "ld1 { v25.h }[2], [x27], #0x2\n"
+ "ld1 { v27.h }[2], [x26], #0x2\n"
+ "ld1 { v1.h }[2], [x25], #0x2\n"
+ "ld1 { v2.h }[2], [x24], #0x2\n"
+ "ld1 { v12.h }[2], [x23], #0x2\n"
+ "ld1 { v16.h }[2], [x22], #0x2\n"
+ "ld1 { v23.h }[2], [x21], #0x2\n"
+ "ld1 { v10.h }[2], [x20], #0x2\n"
"tbz x7, #0, 11f\n"
- "ld1 { v31.b }[6], [x27]\n"
- "ld1 { v30.b }[6], [x26]\n"
- "ld1 { v29.b }[6], [x25]\n"
- "ld1 { v28.b }[6], [x24]\n"
- "ld1 { v27.b }[6], [x23]\n"
- "ld1 { v26.b }[6], [x22]\n"
- "ld1 { v25.b }[6], [x21]\n"
- "ld1 { v24.b }[6], [x20]\n"
+ "ld1 { v25.b }[6], [x27]\n"
+ "ld1 { v27.b }[6], [x26]\n"
+ "ld1 { v1.b }[6], [x25]\n"
+ "ld1 { v2.b }[6], [x24]\n"
+ "ld1 { v12.b }[6], [x23]\n"
+ "ld1 { v16.b }[6], [x22]\n"
+ "ld1 { v23.b }[6], [x21]\n"
+ "ld1 { v10.b }[6], [x20]\n"
"b 11f\n"
"8:" // Oddments: Initial loads: Bit 2: Bit 1: Unset
"tbz x7, #0, 11f\n"
- "ld1 { v31.b }[4], [x27]\n"
- "ld1 { v30.b }[4], [x26]\n"
- "ld1 { v29.b }[4], [x25]\n"
- "ld1 { v28.b }[4], [x24]\n"
- "ld1 { v27.b }[4], [x23]\n"
- "ld1 { v26.b }[4], [x22]\n"
- "ld1 { v25.b }[4], [x21]\n"
- "ld1 { v24.b }[4], [x20]\n"
+ "ld1 { v25.b }[4], [x27]\n"
+ "ld1 { v27.b }[4], [x26]\n"
+ "ld1 { v1.b }[4], [x25]\n"
+ "ld1 { v2.b }[4], [x24]\n"
+ "ld1 { v12.b }[4], [x23]\n"
+ "ld1 { v16.b }[4], [x22]\n"
+ "ld1 { v23.b }[4], [x21]\n"
+ "ld1 { v10.b }[4], [x20]\n"
"b 11f\n"
"9:" // Oddments: Initial loads: Bit 2: Unset
"tbz x7, #1, 10f\n"
- "ld1 { v31.h }[0], [x27], #0x2\n"
- "ld1 { v30.h }[0], [x26], #0x2\n"
- "ld1 { v29.h }[0], [x25], #0x2\n"
- "ld1 { v28.h }[0], [x24], #0x2\n"
- "ld1 { v27.h }[0], [x23], #0x2\n"
- "ld1 { v26.h }[0], [x22], #0x2\n"
- "ld1 { v25.h }[0], [x21], #0x2\n"
- "ld1 { v24.h }[0], [x20], #0x2\n"
+ "ld1 { v25.h }[0], [x27], #0x2\n"
+ "ld1 { v27.h }[0], [x26], #0x2\n"
+ "ld1 { v1.h }[0], [x25], #0x2\n"
+ "ld1 { v2.h }[0], [x24], #0x2\n"
+ "ld1 { v12.h }[0], [x23], #0x2\n"
+ "ld1 { v16.h }[0], [x22], #0x2\n"
+ "ld1 { v23.h }[0], [x21], #0x2\n"
+ "ld1 { v10.h }[0], [x20], #0x2\n"
"tbz x7, #0, 11f\n"
- "ld1 { v31.b }[2], [x27]\n"
- "ld1 { v30.b }[2], [x26]\n"
- "ld1 { v29.b }[2], [x25]\n"
- "ld1 { v28.b }[2], [x24]\n"
- "ld1 { v27.b }[2], [x23]\n"
- "ld1 { v26.b }[2], [x22]\n"
- "ld1 { v25.b }[2], [x21]\n"
- "ld1 { v24.b }[2], [x20]\n"
+ "ld1 { v25.b }[2], [x27]\n"
+ "ld1 { v27.b }[2], [x26]\n"
+ "ld1 { v1.b }[2], [x25]\n"
+ "ld1 { v2.b }[2], [x24]\n"
+ "ld1 { v12.b }[2], [x23]\n"
+ "ld1 { v16.b }[2], [x22]\n"
+ "ld1 { v23.b }[2], [x21]\n"
+ "ld1 { v10.b }[2], [x20]\n"
"b 11f\n"
"10:" // Oddments: Initial loads: Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 11f\n"
- "ld1 { v31.b }[0], [x27]\n"
- "ld1 { v30.b }[0], [x26]\n"
- "ld1 { v29.b }[0], [x25]\n"
- "ld1 { v28.b }[0], [x24]\n"
- "ld1 { v27.b }[0], [x23]\n"
- "ld1 { v26.b }[0], [x22]\n"
- "ld1 { v25.b }[0], [x21]\n"
- "ld1 { v24.b }[0], [x20]\n"
+ "ld1 { v25.b }[0], [x27]\n"
+ "ld1 { v27.b }[0], [x26]\n"
+ "ld1 { v1.b }[0], [x25]\n"
+ "ld1 { v2.b }[0], [x24]\n"
+ "ld1 { v12.b }[0], [x23]\n"
+ "ld1 { v16.b }[0], [x22]\n"
+ "ld1 { v23.b }[0], [x21]\n"
+ "ld1 { v10.b }[0], [x20]\n"
"11:" // Oddments: Initial loads: Bit 2: End
- "usubl v31.8h, v31.8b, v12.8b\n"
- "smlal v15.4s, v31.4h, v8.4h\n"
- "smlal2 v17.4s, v31.8h, v8.8h\n"
- "ldr x24, [x15, #0x40]\n"
- "usubl v30.8h, v30.8b, v12.8b\n"
- "smlal v15.4s, v30.4h, v0.4h\n"
- "smlal2 v17.4s, v30.8h, v0.8h\n"
- "add x24, x24, x17\n"
- "usubl v29.8h, v29.8b, v12.8b\n"
- "smlal v10.4s, v31.4h, v6.4h\n"
- "smlal2 v20.4s, v31.8h, v6.8h\n"
- "smlal v15.4s, v29.4h, v1.4h\n"
- "smlal2 v17.4s, v29.8h, v1.8h\n"
- "usubl v28.8h, v28.8b, v12.8b\n"
- "usubl v26.8h, v26.8b, v12.8b\n"
- "smlal v10.4s, v28.4h, v1.4h\n"
- "smlal2 v20.4s, v28.8h, v1.8h\n"
- "smlal v15.4s, v26.4h, v3.4h\n"
- "smlal2 v17.4s, v26.8h, v3.8h\n"
- "usubl v27.8h, v27.8b, v12.8b\n"
- "usubl v25.8h, v25.8b, v12.8b\n"
- "smlal v10.4s, v27.4h, v2.4h\n"
- "smlal2 v20.4s, v27.8h, v2.8h\n"
- "smlal v15.4s, v25.4h, v4.4h\n"
- "smlal2 v17.4s, v25.8h, v4.8h\n"
- "usubl v24.8h, v24.8b, v12.8b\n"
- "smlal v9.4s, v31.4h, v2.4h\n"
- "smlal2 v23.4s, v31.8h, v2.8h\n"
- "smlal v21.4s, v31.4h, v0.4h\n"
- "smlal2 v22.4s, v31.8h, v0.8h\n"
- "smlal v15.4s, v24.4h, v2.4h\n"
- "smlal2 v17.4s, v24.8h, v2.8h\n"
- "smlal v10.4s, v24.4h, v0.4h\n"
- "smlal2 v20.4s, v24.8h, v0.8h\n"
+ "usubl v25.8h, v25.8b, v6.8b\n"
+ "smlal v5.4s, v25.4h, v4.4h\n"
+ "smlal2 v3.4s, v25.8h, v4.8h\n"
+ "ldr x20, [x15, #0x40]\n"
+ "usubl v27.8h, v27.8b, v6.8b\n"
+ "smlal v5.4s, v27.4h, v11.4h\n"
+ "smlal2 v3.4s, v27.8h, v11.8h\n"
+ "usubl v1.8h, v1.8b, v6.8b\n"
+ "smlal v21.4s, v25.4h, v26.4h\n"
+ "smlal2 v8.4s, v25.8h, v26.8h\n"
+ "add x20, x20, x17\n"
+ "smlal v5.4s, v1.4h, v22.4h\n"
+ "smlal2 v3.4s, v1.8h, v22.8h\n"
+ "usubl v2.8h, v2.8b, v6.8b\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "smlal v21.4s, v2.4h, v22.4h\n"
+ "smlal2 v8.4s, v2.8h, v22.8h\n"
+ "smlal v5.4s, v16.4h, v28.4h\n"
+ "smlal2 v3.4s, v16.8h, v28.8h\n"
+ "usubl v12.8h, v12.8b, v6.8b\n"
+ "usubl v23.8h, v23.8b, v6.8b\n"
+ "smlal v21.4s, v12.4h, v14.4h\n"
+ "smlal2 v8.4s, v12.8h, v14.8h\n"
+ "smlal v5.4s, v23.4h, v18.4h\n"
+ "smlal2 v3.4s, v23.8h, v18.8h\n"
+ "usubl v10.8h, v10.8b, v6.8b\n"
+ "smlal v20.4s, v25.4h, v14.4h\n"
+ "smlal2 v0.4s, v25.8h, v14.8h\n"
+ "smlal v19.4s, v25.4h, v11.4h\n"
+ "smlal2 v31.4s, v25.8h, v11.8h\n"
+ "smlal v5.4s, v10.4h, v14.4h\n"
+ "smlal2 v3.4s, v10.8h, v14.8h\n"
+ "smlal v21.4s, v10.4h, v11.4h\n"
+ "smlal2 v8.4s, v10.8h, v11.8h\n"
"tbz x7, #2, 13f\n"
- "ld1 { v29.s }[0], [x24], #0x4\n"
+ "ld1 { v15.s }[0], [x20], #0x4\n"
"tbz x7, #1, 12f\n"
- "ld1 { v29.h }[2], [x24], #0x2\n"
+ "ld1 { v15.h }[2], [x20], #0x2\n"
"tbz x7, #0, 15f\n"
- "ld1 { v29.b }[6], [x24]\n"
+ "ld1 { v15.b }[6], [x20]\n"
"b 15f\n"
"12:" // Oddments: Load (1, 3): Bit 2: Bit 1: Unset
"tbz x7, #0, 15f\n"
- "ld1 { v29.b }[4], [x24]\n"
+ "ld1 { v15.b }[4], [x20]\n"
"b 15f\n"
"13:" // Oddments: Load (1, 3): Bit 2: Unset
"tbz x7, #1, 14f\n"
- "ld1 { v29.h }[0], [x24], #0x2\n"
+ "ld1 { v15.h }[0], [x20], #0x2\n"
"tbz x7, #0, 15f\n"
- "ld1 { v29.b }[2], [x24]\n"
+ "ld1 { v15.b }[2], [x20]\n"
"b 15f\n"
"14:" // Oddments: Load (1, 3): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 15f\n"
- "ld1 { v29.b }[0], [x24]\n"
+ "ld1 { v15.b }[0], [x20]\n"
"15:" // Oddments: Load (1, 3): Bit 2: End
- "usubl v29.8h, v29.8b, v12.8b\n"
- "ldr x22, [x15, #0x48]\n"
- "smlal v10.4s, v29.4h, v4.4h\n"
- "smlal2 v20.4s, v29.8h, v4.8h\n"
- "add x22, x22, x17\n"
+ "usubl v15.8h, v15.8b, v6.8b\n"
+ "ldr x20, [x15, #0x48]\n"
+ "smlal v21.4s, v15.4h, v18.4h\n"
+ "smlal2 v8.4s, v15.8h, v18.8h\n"
+ "add x20, x20, x17\n"
"tbz x7, #2, 17f\n"
- "ld1 { v28.s }[0], [x22], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 16f\n"
- "ld1 { v28.h }[2], [x22], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 19f\n"
- "ld1 { v28.b }[6], [x22]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 19f\n"
"16:" // Oddments: Load (1, 4): Bit 2: Bit 1: Unset
"tbz x7, #0, 19f\n"
- "ld1 { v28.b }[4], [x22]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 19f\n"
"17:" // Oddments: Load (1, 4): Bit 2: Unset
"tbz x7, #1, 18f\n"
- "ld1 { v28.h }[0], [x22], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 19f\n"
- "ld1 { v28.b }[2], [x22]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 19f\n"
"18:" // Oddments: Load (1, 4): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 19f\n"
- "ld1 { v28.b }[0], [x22]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"19:" // Oddments: Load (1, 4): Bit 2: End
- "usubl v28.8h, v28.8b, v12.8b\n"
- "ldr x21, [x15, #0x50]\n"
- "smlal v10.4s, v28.4h, v5.4h\n"
- "smlal2 v20.4s, v28.8h, v5.8h\n"
- "add x21, x21, x17\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x20, [x15, #0x50]\n"
+ "smlal v21.4s, v16.4h, v9.4h\n"
+ "smlal2 v8.4s, v16.8h, v9.8h\n"
+ "add x20, x20, x17\n"
"tbz x7, #2, 21f\n"
- "ld1 { v27.s }[0], [x21], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 20f\n"
- "ld1 { v27.h }[2], [x21], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 23f\n"
- "ld1 { v27.b }[6], [x21]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 23f\n"
"20:" // Oddments: Load (1, 2): Bit 2: Bit 1: Unset
"tbz x7, #0, 23f\n"
- "ld1 { v27.b }[4], [x21]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 23f\n"
"21:" // Oddments: Load (1, 2): Bit 2: Unset
"tbz x7, #1, 22f\n"
- "ld1 { v27.h }[0], [x21], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 23f\n"
- "ld1 { v27.b }[2], [x21]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 23f\n"
"22:" // Oddments: Load (1, 2): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 23f\n"
- "ld1 { v27.b }[0], [x21]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"23:" // Oddments: Load (1, 2): Bit 2: End
- "usubl v27.8h, v27.8b, v12.8b\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
"ldr x20, [x15, #0x58]\n"
- "smlal v15.4s, v27.4h, v5.4h\n"
- "smlal2 v17.4s, v27.8h, v5.8h\n"
- "smlal v10.4s, v27.4h, v3.4h\n"
- "smlal2 v20.4s, v27.8h, v3.8h\n"
+ "smlal v5.4s, v16.4h, v9.4h\n"
+ "smlal2 v3.4s, v16.8h, v9.8h\n"
+ "smlal v21.4s, v16.4h, v28.4h\n"
+ "smlal2 v8.4s, v16.8h, v28.8h\n"
"add x20, x20, x17\n"
"tbz x7, #2, 25f\n"
- "ld1 { v26.s }[0], [x20], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 24f\n"
- "ld1 { v26.h }[2], [x20], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 27f\n"
- "ld1 { v26.b }[6], [x20]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 27f\n"
"24:" // Oddments: Load (3, 0): Bit 2: Bit 1: Unset
"tbz x7, #0, 27f\n"
- "ld1 { v26.b }[4], [x20]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 27f\n"
"25:" // Oddments: Load (3, 0): Bit 2: Unset
"tbz x7, #1, 26f\n"
- "ld1 { v26.h }[0], [x20], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 27f\n"
- "ld1 { v26.b }[2], [x20]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 27f\n"
"26:" // Oddments: Load (3, 0): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 27f\n"
- "ld1 { v26.b }[0], [x20]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"27:" // Oddments: Load (3, 0): Bit 2: End
- "usubl v26.8h, v26.8b, v12.8b\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
"ldr x20, [x15, #0x60]\n"
- "smlal v9.4s, v26.4h, v3.4h\n"
- "smlal2 v23.4s, v26.8h, v3.8h\n"
+ "smlal v20.4s, v16.4h, v28.4h\n"
+ "smlal2 v0.4s, v16.8h, v28.8h\n"
"add x20, x20, x17\n"
"tbz x7, #2, 29f\n"
- "ld1 { v25.s }[0], [x20], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 28f\n"
- "ld1 { v25.h }[2], [x20], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 31f\n"
- "ld1 { v25.b }[6], [x20]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 31f\n"
"28:" // Oddments: Load (2, 0): Bit 2: Bit 1: Unset
"tbz x7, #0, 31f\n"
- "ld1 { v25.b }[4], [x20]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 31f\n"
"29:" // Oddments: Load (2, 0): Bit 2: Unset
"tbz x7, #1, 30f\n"
- "ld1 { v25.h }[0], [x20], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 31f\n"
- "ld1 { v25.b }[2], [x20]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 31f\n"
"30:" // Oddments: Load (2, 0): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 31f\n"
- "ld1 { v25.b }[0], [x20]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"31:" // Oddments: Load (2, 0): Bit 2: End
- "usubl v25.8h, v25.8b, v12.8b\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
"ldr x20, [x15, #0x68]\n"
- "smlal v15.4s, v25.4h, v6.4h\n"
- "smlal2 v17.4s, v25.8h, v6.8h\n"
- "smlal v9.4s, v25.4h, v0.4h\n"
- "smlal2 v23.4s, v25.8h, v0.8h\n"
+ "smlal v5.4s, v16.4h, v26.4h\n"
+ "smlal2 v3.4s, v16.8h, v26.8h\n"
+ "smlal v20.4s, v16.4h, v11.4h\n"
+ "smlal2 v0.4s, v16.8h, v11.8h\n"
"add x20, x20, x17\n"
"tbz x7, #2, 33f\n"
- "ld1 { v29.s }[0], [x20], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 32f\n"
- "ld1 { v29.h }[2], [x20], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 35f\n"
- "ld1 { v29.b }[6], [x20]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 35f\n"
"32:" // Oddments: Load (3, 1): Bit 2: Bit 1: Unset
"tbz x7, #0, 35f\n"
- "ld1 { v29.b }[4], [x20]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 35f\n"
"33:" // Oddments: Load (3, 1): Bit 2: Unset
"tbz x7, #1, 34f\n"
- "ld1 { v29.h }[0], [x20], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 35f\n"
- "ld1 { v29.b }[2], [x20]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 35f\n"
"34:" // Oddments: Load (3, 1): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 35f\n"
- "ld1 { v29.b }[0], [x20]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"35:" // Oddments: Load (3, 1): Bit 2: End
- "usubl v29.8h, v29.8b, v12.8b\n"
- "ldr x21, [x15, #0x70]\n"
- "smlal v9.4s, v29.4h, v4.4h\n"
- "smlal2 v23.4s, v29.8h, v4.8h\n"
- "add x21, x21, x17\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x20, [x15, #0x70]\n"
+ "smlal v20.4s, v16.4h, v18.4h\n"
+ "smlal2 v0.4s, v16.8h, v18.8h\n"
+ "add x20, x20, x17\n"
"tbz x7, #2, 37f\n"
- "ld1 { v24.s }[0], [x21], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 36f\n"
- "ld1 { v24.h }[2], [x21], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 39f\n"
- "ld1 { v24.b }[6], [x21]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 39f\n"
"36:" // Oddments: Load (2, 1): Bit 2: Bit 1: Unset
"tbz x7, #0, 39f\n"
- "ld1 { v24.b }[4], [x21]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 39f\n"
"37:" // Oddments: Load (2, 1): Bit 2: Unset
"tbz x7, #1, 38f\n"
- "ld1 { v24.h }[0], [x21], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 39f\n"
- "ld1 { v24.b }[2], [x21]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 39f\n"
"38:" // Oddments: Load (2, 1): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 39f\n"
- "ld1 { v24.b }[0], [x21]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"39:" // Oddments: Load (2, 1): Bit 2: End
- "usubl v24.8h, v24.8b, v12.8b\n"
- "ldr x23, [x15, #0x78]\n"
- "smlal v15.4s, v24.4h, v7.4h\n"
- "smlal2 v17.4s, v24.8h, v7.8h\n"
- "smlal v9.4s, v24.4h, v1.4h\n"
- "smlal2 v23.4s, v24.8h, v1.8h\n"
- "add x23, x23, x17\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x20, [x15, #0x78]\n"
+ "smlal v5.4s, v16.4h, v7.4h\n"
+ "smlal2 v3.4s, v16.8h, v7.8h\n"
+ "smlal v20.4s, v16.4h, v22.4h\n"
+ "smlal2 v0.4s, v16.8h, v22.8h\n"
+ "add x20, x20, x17\n"
"tbz x7, #2, 41f\n"
- "ld1 { v27.s }[0], [x23], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 40f\n"
- "ld1 { v27.h }[2], [x23], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 43f\n"
- "ld1 { v27.b }[6], [x23]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 43f\n"
"40:" // Oddments: Load (3, 3): Bit 2: Bit 1: Unset
"tbz x7, #0, 43f\n"
- "ld1 { v27.b }[4], [x23]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 43f\n"
"41:" // Oddments: Load (3, 3): Bit 2: Unset
"tbz x7, #1, 42f\n"
- "ld1 { v27.h }[0], [x23], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 43f\n"
- "ld1 { v27.b }[2], [x23]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 43f\n"
"42:" // Oddments: Load (3, 3): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 43f\n"
- "ld1 { v27.b }[0], [x23]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"43:" // Oddments: Load (3, 3): Bit 2: End
- "usubl v27.8h, v27.8b, v12.8b\n"
- "ldr x21, [x15, #0x80]\n"
- "smlal v21.4s, v27.4h, v4.4h\n"
- "smlal2 v22.4s, v27.8h, v4.8h\n"
- "add x21, x21, x17\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x20, [x15, #0x80]\n"
+ "smlal v19.4s, v16.4h, v18.4h\n"
+ "smlal2 v31.4s, v16.8h, v18.8h\n"
+ "add x20, x20, x17\n"
"tbz x7, #2, 45f\n"
- "ld1 { v28.s }[0], [x21], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 44f\n"
- "ld1 { v28.h }[2], [x21], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 47f\n"
- "ld1 { v28.b }[6], [x21]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 47f\n"
"44:" // Oddments: Load (2, 3): Bit 2: Bit 1: Unset
"tbz x7, #0, 47f\n"
- "ld1 { v28.b }[4], [x21]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 47f\n"
"45:" // Oddments: Load (2, 3): Bit 2: Unset
"tbz x7, #1, 46f\n"
- "ld1 { v28.h }[0], [x21], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 47f\n"
- "ld1 { v28.b }[2], [x21]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 47f\n"
"46:" // Oddments: Load (2, 3): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 47f\n"
- "ld1 { v28.b }[0], [x21]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"47:" // Oddments: Load (2, 3): Bit 2: End
- "usubl v28.8h, v28.8b, v12.8b\n"
- "ldr x22, [x15, #0x88]\n"
- "smlal v10.4s, v28.4h, v7.4h\n"
- "smlal2 v20.4s, v28.8h, v7.8h\n"
- "smlal v21.4s, v28.4h, v1.4h\n"
- "smlal2 v22.4s, v28.8h, v1.8h\n"
- "add x22, x22, x17\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x20, [x15, #0x88]\n"
+ "smlal v21.4s, v16.4h, v7.4h\n"
+ "smlal2 v8.4s, v16.8h, v7.8h\n"
+ "smlal v19.4s, v16.4h, v22.4h\n"
+ "smlal2 v31.4s, v16.8h, v22.8h\n"
+ "add x20, x20, x17\n"
"tbz x7, #2, 49f\n"
- "ld1 { v26.s }[0], [x22], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 48f\n"
- "ld1 { v26.h }[2], [x22], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 51f\n"
- "ld1 { v26.b }[6], [x22]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 51f\n"
"48:" // Oddments: Load (3, 4): Bit 2: Bit 1: Unset
"tbz x7, #0, 51f\n"
- "ld1 { v26.b }[4], [x22]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 51f\n"
"49:" // Oddments: Load (3, 4): Bit 2: Unset
"tbz x7, #1, 50f\n"
- "ld1 { v26.h }[0], [x22], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 51f\n"
- "ld1 { v26.b }[2], [x22]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 51f\n"
"50:" // Oddments: Load (3, 4): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 51f\n"
- "ld1 { v26.b }[0], [x22]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"51:" // Oddments: Load (3, 4): Bit 2: End
- "usubl v26.8h, v26.8b, v12.8b\n"
- "ldr x24, [x15, #0x90]\n"
- "smlal v21.4s, v26.4h, v5.4h\n"
- "smlal2 v22.4s, v26.8h, v5.8h\n"
- "add x24, x24, x17\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x20, [x15, #0x90]\n"
+ "smlal v19.4s, v16.4h, v9.4h\n"
+ "smlal2 v31.4s, v16.8h, v9.8h\n"
+ "add x20, x20, x17\n"
"tbz x7, #2, 53f\n"
- "ld1 { v25.s }[0], [x24], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 52f\n"
- "ld1 { v25.h }[2], [x24], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 55f\n"
- "ld1 { v25.b }[6], [x24]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 55f\n"
"52:" // Oddments: Load (4, 0): Bit 2: Bit 1: Unset
"tbz x7, #0, 55f\n"
- "ld1 { v25.b }[4], [x24]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 55f\n"
"53:" // Oddments: Load (4, 0): Bit 2: Unset
"tbz x7, #1, 54f\n"
- "ld1 { v25.h }[0], [x24], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 55f\n"
- "ld1 { v25.b }[2], [x24]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 55f\n"
"54:" // Oddments: Load (4, 0): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 55f\n"
- "ld1 { v25.b }[0], [x24]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"55:" // Oddments: Load (4, 0): Bit 2: End
- "usubl v25.8h, v25.8b, v12.8b\n"
- "ldr x25, [x15, #0x98]\n"
- "smlal v9.4s, v25.4h, v6.4h\n"
- "smlal2 v23.4s, v25.8h, v6.8h\n"
- "add x25, x25, x17\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x20, [x15, #0x98]\n"
+ "smlal v20.4s, v16.4h, v26.4h\n"
+ "smlal2 v0.4s, v16.8h, v26.8h\n"
+ "add x20, x20, x17\n"
"tbz x7, #2, 57f\n"
- "ld1 { v29.s }[0], [x25], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 56f\n"
- "ld1 { v29.h }[2], [x25], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 59f\n"
- "ld1 { v29.b }[6], [x25]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 59f\n"
"56:" // Oddments: Load (2, 4): Bit 2: Bit 1: Unset
"tbz x7, #0, 59f\n"
- "ld1 { v29.b }[4], [x25]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 59f\n"
"57:" // Oddments: Load (2, 4): Bit 2: Unset
"tbz x7, #1, 58f\n"
- "ld1 { v29.h }[0], [x25], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 59f\n"
- "ld1 { v29.b }[2], [x25]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 59f\n"
"58:" // Oddments: Load (2, 4): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 59f\n"
- "ld1 { v29.b }[0], [x25]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"59:" // Oddments: Load (2, 4): Bit 2: End
- "usubl v29.8h, v29.8b, v12.8b\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
"ldr x20, [x15, #0xa0]\n"
- "smlal v10.4s, v29.4h, v8.4h\n"
- "smlal2 v20.4s, v29.8h, v8.8h\n"
- "smlal v21.4s, v29.4h, v2.4h\n"
- "smlal2 v22.4s, v29.8h, v2.8h\n"
+ "smlal v21.4s, v16.4h, v4.4h\n"
+ "smlal2 v8.4s, v16.8h, v4.8h\n"
+ "smlal v19.4s, v16.4h, v14.4h\n"
+ "smlal2 v31.4s, v16.8h, v14.8h\n"
"add x20, x20, x17\n"
"tbz x7, #2, 61f\n"
- "ld1 { v27.s }[0], [x20], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 60f\n"
- "ld1 { v27.h }[2], [x20], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 63f\n"
- "ld1 { v27.b }[6], [x20]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 63f\n"
"60:" // Oddments: Load (4, 1): Bit 2: Bit 1: Unset
"tbz x7, #0, 63f\n"
- "ld1 { v27.b }[4], [x20]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 63f\n"
"61:" // Oddments: Load (4, 1): Bit 2: Unset
"tbz x7, #1, 62f\n"
- "ld1 { v27.h }[0], [x20], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 63f\n"
- "ld1 { v27.b }[2], [x20]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 63f\n"
"62:" // Oddments: Load (4, 1): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 63f\n"
- "ld1 { v27.b }[0], [x20]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"63:" // Oddments: Load (4, 1): Bit 2: End
- "usubl v27.8h, v27.8b, v12.8b\n"
- "ldr x23, [x15, #0xa8]\n"
- "smlal v9.4s, v27.4h, v7.4h\n"
- "smlal2 v23.4s, v27.8h, v7.8h\n"
- "add x23, x23, x17\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x20, [x15, #0xa8]\n"
+ "smlal v20.4s, v16.4h, v7.4h\n"
+ "smlal2 v0.4s, v16.8h, v7.8h\n"
+ "add x20, x20, x17\n"
"tbz x7, #2, 65f\n"
- "ld1 { v24.s }[0], [x23], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 64f\n"
- "ld1 { v24.h }[2], [x23], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 67f\n"
- "ld1 { v24.b }[6], [x23]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 67f\n"
"64:" // Oddments: Load (3, 2): Bit 2: Bit 1: Unset
"tbz x7, #0, 67f\n"
- "ld1 { v24.b }[4], [x23]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 67f\n"
"65:" // Oddments: Load (3, 2): Bit 2: Unset
"tbz x7, #1, 66f\n"
- "ld1 { v24.h }[0], [x23], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 67f\n"
- "ld1 { v24.b }[2], [x23]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 67f\n"
"66:" // Oddments: Load (3, 2): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 67f\n"
- "ld1 { v24.b }[0], [x23]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"67:" // Oddments: Load (3, 2): Bit 2: End
- "usubl v24.8h, v24.8b, v12.8b\n"
- "ldr x22, [x15, #0xb0]\n"
- "smlal v9.4s, v24.4h, v5.4h\n"
- "smlal2 v23.4s, v24.8h, v5.8h\n"
- "smlal v21.4s, v24.4h, v3.4h\n"
- "smlal2 v22.4s, v24.8h, v3.8h\n"
- "add x22, x22, x17\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x20, [x15, #0xb0]\n"
+ "smlal v20.4s, v16.4h, v9.4h\n"
+ "smlal2 v0.4s, v16.8h, v9.8h\n"
+ "smlal v19.4s, v16.4h, v28.4h\n"
+ "smlal2 v31.4s, v16.8h, v28.8h\n"
+ "add x20, x20, x17\n"
"tbz x7, #2, 69f\n"
- "ld1 { v26.s }[0], [x22], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 68f\n"
- "ld1 { v26.h }[2], [x22], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 71f\n"
- "ld1 { v26.b }[6], [x22]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 71f\n"
"68:" // Oddments: Load (4, 3): Bit 2: Bit 1: Unset
"tbz x7, #0, 71f\n"
- "ld1 { v26.b }[4], [x22]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 71f\n"
"69:" // Oddments: Load (4, 3): Bit 2: Unset
"tbz x7, #1, 70f\n"
- "ld1 { v26.h }[0], [x22], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 71f\n"
- "ld1 { v26.b }[2], [x22]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 71f\n"
"70:" // Oddments: Load (4, 3): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 71f\n"
- "ld1 { v26.b }[0], [x22]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"71:" // Oddments: Load (4, 3): Bit 2: End
- "usubl v26.8h, v26.8b, v12.8b\n"
- "ldr x21, [x15, #0xb8]\n"
- "smlal v21.4s, v26.4h, v7.4h\n"
- "smlal2 v22.4s, v26.8h, v7.8h\n"
- "add x21, x21, x17\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "ldr x20, [x15, #0xb8]\n"
+ "smlal v19.4s, v16.4h, v7.4h\n"
+ "smlal2 v31.4s, v16.8h, v7.8h\n"
+ "add x20, x20, x17\n"
"tbz x7, #2, 73f\n"
- "ld1 { v25.s }[0], [x21], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 72f\n"
- "ld1 { v25.h }[2], [x21], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 75f\n"
- "ld1 { v25.b }[6], [x21]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 75f\n"
"72:" // Oddments: Load (4, 2): Bit 2: Bit 1: Unset
"tbz x7, #0, 75f\n"
- "ld1 { v25.b }[4], [x21]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 75f\n"
"73:" // Oddments: Load (4, 2): Bit 2: Unset
"tbz x7, #1, 74f\n"
- "ld1 { v25.h }[0], [x21], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 75f\n"
- "ld1 { v25.b }[2], [x21]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 75f\n"
"74:" // Oddments: Load (4, 2): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 75f\n"
- "ld1 { v25.b }[0], [x21]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"75:" // Oddments: Load (4, 2): Bit 2: End
- "usubl v25.8h, v25.8b, v12.8b\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
"ldr x20, [x15, #0xc0]\n"
- "smlal v9.4s, v25.4h, v8.4h\n"
- "smlal2 v23.4s, v25.8h, v8.8h\n"
- "smlal v21.4s, v25.4h, v6.4h\n"
- "smlal2 v22.4s, v25.8h, v6.8h\n"
+ "smlal v20.4s, v16.4h, v4.4h\n"
+ "smlal2 v0.4s, v16.8h, v4.8h\n"
+ "smlal v19.4s, v16.4h, v26.4h\n"
+ "smlal2 v31.4s, v16.8h, v26.8h\n"
"add x20, x20, x17\n"
"tbz x7, #2, 77f\n"
- "ld1 { v29.s }[0], [x20], #0x4\n"
+ "ld1 { v16.s }[0], [x20], #0x4\n"
"tbz x7, #1, 76f\n"
- "ld1 { v29.h }[2], [x20], #0x2\n"
+ "ld1 { v16.h }[2], [x20], #0x2\n"
"tbz x7, #0, 79f\n"
- "ld1 { v29.b }[6], [x20]\n"
+ "ld1 { v16.b }[6], [x20]\n"
"b 79f\n"
"76:" // Oddments: Load (4, 4): Bit 2: Bit 1: Unset
"tbz x7, #0, 79f\n"
- "ld1 { v29.b }[4], [x20]\n"
+ "ld1 { v16.b }[4], [x20]\n"
"b 79f\n"
"77:" // Oddments: Load (4, 4): Bit 2: Unset
"tbz x7, #1, 78f\n"
- "ld1 { v29.h }[0], [x20], #0x2\n"
+ "ld1 { v16.h }[0], [x20], #0x2\n"
"tbz x7, #0, 79f\n"
- "ld1 { v29.b }[2], [x20]\n"
+ "ld1 { v16.b }[2], [x20]\n"
"b 79f\n"
"78:" // Oddments: Load (4, 4): Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 79f\n"
- "ld1 { v29.b }[0], [x20]\n"
+ "ld1 { v16.b }[0], [x20]\n"
"79:" // Oddments: Load (4, 4): Bit 2: End
- "usubl v29.8h, v29.8b, v12.8b\n"
- "smlal v21.4s, v29.4h, v8.4h\n"
- "smlal2 v22.4s, v29.8h, v8.8h\n"
+ "usubl v16.8h, v16.8b, v6.8b\n"
+ "smlal v19.4s, v16.4h, v4.4h\n"
+ "smlal2 v31.4s, v16.8h, v4.8h\n"
"tbz x7, #2, 81f\n"
- "ld1 { v19.4s }, [x13], #0x10\n"
- "ld1 { v18.4s }, [x12], #0x10\n"
+ "ld1 { v14.4s }, [x13], #0x10\n"
+ "ld1 { v25.4s }, [x12], #0x10\n"
"tbz x7, #1, 80f\n"
- "ld1 { v30.d }[0], [x13], #0x8\n"
- "ld1 { v31.d }[0], [x12], #0x8\n"
+ "ld1 { v18.d }[0], [x13], #0x8\n"
+ "ld1 { v12.d }[0], [x12], #0x8\n"
"tbz x7, #0, 83f\n"
- "ld1 { v30.s }[2], [x13]\n"
- "ld1 { v31.s }[2], [x12]\n"
+ "ld1 { v18.s }[2], [x13]\n"
+ "ld1 { v12.s }[2], [x12]\n"
"b 83f\n"
"80:" // Oddments: Load requant params: Bit 2: Bit 1: Unset
"tbz x7, #0, 83f\n"
- "ld1 { v30.s }[0], [x13]\n"
- "ld1 { v31.s }[0], [x12]\n"
+ "ld1 { v18.s }[0], [x13]\n"
+ "ld1 { v12.s }[0], [x12]\n"
"b 83f\n"
"81:" // Oddments: Load requant params: Bit 2: Unset
"tbz x7, #1, 82f\n"
- "ld1 { v19.d }[0], [x13], #0x8\n"
- "ld1 { v18.d }[0], [x12], #0x8\n"
+ "ld1 { v14.d }[0], [x13], #0x8\n"
+ "ld1 { v25.d }[0], [x12], #0x8\n"
"tbz x7, #0, 83f\n"
- "ld1 { v19.s }[2], [x13]\n"
- "ld1 { v18.s }[2], [x12]\n"
+ "ld1 { v14.s }[2], [x13]\n"
+ "ld1 { v25.s }[2], [x12]\n"
"b 83f\n"
"82:" // Oddments: Load requant params: Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 83f\n"
- "ld1 { v19.s }[0], [x13]\n"
- "ld1 { v18.s }[0], [x12]\n"
+ "ld1 { v14.s }[0], [x13]\n"
+ "ld1 { v25.s }[0], [x12]\n"
"83:" // Oddments: Load requant params: Bit 2: End
- "sqrdmulh v15.4s, v15.4s, v19.4s\n"
- "and v0.16b, v15.16b, v18.16b\n"
+ "sqrdmulh v5.4s, v5.4s, v14.4s\n"
+ "and v28.16b, v5.16b, v25.16b\n"
"add x11, x11, x16\n"
"add x10, x10, x16\n"
- "sqrdmulh v17.4s, v17.4s, v30.4s\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
+ "sqrdmulh v3.4s, v3.4s, v18.4s\n"
+ "sshr v28.4s, v28.4s, #0x1f\n"
"add x9, x9, x16\n"
"add x28, x28, x16\n"
- "and v7.16b, v17.16b, v31.16b\n"
- "sqrdmulh v10.4s, v10.4s, v19.4s\n"
- "sqrdmulh v9.4s, v9.4s, v19.4s\n"
- "sqrdmulh v21.4s, v21.4s, v19.4s\n"
- "sqadd v15.4s, v15.4s, v0.4s\n"
- "sshr v7.4s, v7.4s, #0x1f\n"
- "and v19.16b, v10.16b, v18.16b\n"
- "sqrdmulh v20.4s, v20.4s, v30.4s\n"
- "and v27.16b, v9.16b, v18.16b\n"
- "sqrdmulh v23.4s, v23.4s, v30.4s\n"
- "and v0.16b, v21.16b, v18.16b\n"
- "sqrdmulh v22.4s, v22.4s, v30.4s\n"
- "sqadd v17.4s, v17.4s, v7.4s\n"
- "sshr v19.4s, v19.4s, #0x1f\n"
- "and v5.16b, v20.16b, v31.16b\n"
- "sshr v27.4s, v27.4s, #0x1f\n"
- "and v4.16b, v23.16b, v31.16b\n"
- "sshr v0.4s, v0.4s, #0x1f\n"
- "and v7.16b, v22.16b, v31.16b\n"
- "sqadd v10.4s, v10.4s, v19.4s\n"
- "sshr v5.4s, v5.4s, #0x1f\n"
- "sqadd v9.4s, v9.4s, v27.4s\n"
+ "and v16.16b, v3.16b, v12.16b\n"
+ "sqrdmulh v21.4s, v21.4s, v14.4s\n"
+ "sqrdmulh v20.4s, v20.4s, v14.4s\n"
+ "sqrdmulh v19.4s, v19.4s, v14.4s\n"
+ "sqadd v5.4s, v5.4s, v28.4s\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "and v14.16b, v21.16b, v25.16b\n"
+ "sqrdmulh v8.4s, v8.4s, v18.4s\n"
+ "and v6.16b, v20.16b, v25.16b\n"
+ "sqrdmulh v0.4s, v0.4s, v18.4s\n"
+ "and v4.16b, v19.16b, v25.16b\n"
+ "sqrdmulh v31.4s, v31.4s, v18.4s\n"
+ "sqadd v3.4s, v3.4s, v16.4s\n"
+ "sshr v14.4s, v14.4s, #0x1f\n"
+ "and v18.16b, v8.16b, v12.16b\n"
+ "sshr v6.4s, v6.4s, #0x1f\n"
+ "and v7.16b, v0.16b, v12.16b\n"
"sshr v4.4s, v4.4s, #0x1f\n"
- "sqadd v21.4s, v21.4s, v0.4s\n"
+ "and v16.16b, v31.16b, v12.16b\n"
+ "sqadd v21.4s, v21.4s, v14.4s\n"
+ "sshr v18.4s, v18.4s, #0x1f\n"
+ "sqadd v20.4s, v20.4s, v6.4s\n"
"sshr v7.4s, v7.4s, #0x1f\n"
- "srshl v15.4s, v15.4s, v18.4s\n"
- "srshl v10.4s, v10.4s, v18.4s\n"
- "sqadd v20.4s, v20.4s, v5.4s\n"
- "srshl v9.4s, v9.4s, v18.4s\n"
- "sqadd v23.4s, v23.4s, v4.4s\n"
- "srshl v21.4s, v21.4s, v18.4s\n"
- "sqadd v22.4s, v22.4s, v7.4s\n"
- "srshl v17.4s, v17.4s, v31.4s\n"
- "sqxtn v15.4h, v15.4s\n"
- "srshl v20.4s, v20.4s, v31.4s\n"
- "sqxtn v10.4h, v10.4s\n"
- "srshl v23.4s, v23.4s, v31.4s\n"
- "sqxtn v9.4h, v9.4s\n"
- "srshl v22.4s, v22.4s, v31.4s\n"
+ "sqadd v19.4s, v19.4s, v4.4s\n"
+ "sshr v16.4s, v16.4s, #0x1f\n"
+ "srshl v5.4s, v5.4s, v25.4s\n"
+ "srshl v21.4s, v21.4s, v25.4s\n"
+ "sqadd v8.4s, v8.4s, v18.4s\n"
+ "srshl v20.4s, v20.4s, v25.4s\n"
+ "sqadd v0.4s, v0.4s, v7.4s\n"
+ "srshl v19.4s, v19.4s, v25.4s\n"
+ "sqadd v31.4s, v31.4s, v16.4s\n"
+ "srshl v3.4s, v3.4s, v12.4s\n"
+ "sqxtn v5.4h, v5.4s\n"
+ "srshl v8.4s, v8.4s, v12.4s\n"
"sqxtn v21.4h, v21.4s\n"
- "sqxtn2 v15.8h, v17.4s\n"
- "sqxtn2 v10.8h, v20.4s\n"
- "sqxtn2 v9.8h, v23.4s\n"
- "sqxtn2 v21.8h, v22.4s\n"
- "sqadd v15.8h, v15.8h, v11.8h\n"
- "sqadd v10.8h, v10.8h, v11.8h\n"
- "sqadd v9.8h, v9.8h, v11.8h\n"
- "sqadd v21.8h, v21.8h, v11.8h\n"
- "smax v15.8h, v15.8h, v16.8h\n"
- "smax v10.8h, v10.8h, v16.8h\n"
- "smax v9.8h, v9.8h, v16.8h\n"
- "smax v21.8h, v21.8h, v16.8h\n"
- "smin v15.8h, v15.8h, v14.8h\n"
- "smin v10.8h, v10.8h, v14.8h\n"
- "smin v9.8h, v9.8h, v14.8h\n"
- "smin v21.8h, v21.8h, v14.8h\n"
- "uzp1 v15.16b, v15.16b, v15.16b\n"
- "uzp1 v10.16b, v10.16b, v10.16b\n"
- "uzp1 v9.16b, v9.16b, v9.16b\n"
+ "srshl v0.4s, v0.4s, v12.4s\n"
+ "sqxtn v20.4h, v20.4s\n"
+ "srshl v31.4s, v31.4s, v12.4s\n"
+ "sqxtn v19.4h, v19.4s\n"
+ "sqxtn2 v5.8h, v3.4s\n"
+ "sqxtn2 v21.8h, v8.4s\n"
+ "sqxtn2 v20.8h, v0.4s\n"
+ "sqxtn2 v19.8h, v31.4s\n"
+ "sqadd v5.8h, v5.8h, v13.8h\n"
+ "sqadd v21.8h, v21.8h, v13.8h\n"
+ "sqadd v20.8h, v20.8h, v13.8h\n"
+ "sqadd v19.8h, v19.8h, v13.8h\n"
+ "smax v5.8h, v5.8h, v17.8h\n"
+ "smax v21.8h, v21.8h, v17.8h\n"
+ "smax v20.8h, v20.8h, v17.8h\n"
+ "smax v19.8h, v19.8h, v17.8h\n"
+ "smin v5.8h, v5.8h, v24.8h\n"
+ "smin v21.8h, v21.8h, v24.8h\n"
+ "smin v20.8h, v20.8h, v24.8h\n"
+ "smin v19.8h, v19.8h, v24.8h\n"
+ "uzp1 v5.16b, v5.16b, v5.16b\n"
"uzp1 v21.16b, v21.16b, v21.16b\n"
+ "uzp1 v20.16b, v20.16b, v20.16b\n"
+ "uzp1 v19.16b, v19.16b, v19.16b\n"
"tbz x7, #2, 85f\n"
- "st1 { v15.s }[0], [x11], #0x4\n"
- "st1 { v10.s }[0], [x10], #0x4\n"
- "st1 { v9.s }[0], [x9], #0x4\n"
- "st1 { v21.s }[0], [x28], #0x4\n"
+ "st1 { v5.s }[0], [x11], #0x4\n"
+ "st1 { v21.s }[0], [x10], #0x4\n"
+ "st1 { v20.s }[0], [x9], #0x4\n"
+ "st1 { v19.s }[0], [x28], #0x4\n"
"tbz x7, #1, 84f\n"
- "st1 { v15.h }[2], [x11], #0x2\n"
- "st1 { v10.h }[2], [x10], #0x2\n"
- "st1 { v9.h }[2], [x9], #0x2\n"
- "st1 { v21.h }[2], [x28], #0x2\n"
+ "st1 { v5.h }[2], [x11], #0x2\n"
+ "st1 { v21.h }[2], [x10], #0x2\n"
+ "st1 { v20.h }[2], [x9], #0x2\n"
+ "st1 { v19.h }[2], [x28], #0x2\n"
"tbz x7, #0, 87f\n"
- "st1 { v15.b }[6], [x11], #0x1\n"
- "st1 { v10.b }[6], [x10], #0x1\n"
- "st1 { v9.b }[6], [x9], #0x1\n"
- "st1 { v21.b }[6], [x28], #0x1\n"
+ "st1 { v5.b }[6], [x11], #0x1\n"
+ "st1 { v21.b }[6], [x10], #0x1\n"
+ "st1 { v20.b }[6], [x9], #0x1\n"
+ "st1 { v19.b }[6], [x28], #0x1\n"
"b 87f\n"
"84:" // Oddments: Bit 2: Bit 1: Unset
"tbz x7, #0, 87f\n"
- "st1 { v15.b }[4], [x11], #0x1\n"
- "st1 { v10.b }[4], [x10], #0x1\n"
- "st1 { v9.b }[4], [x9], #0x1\n"
- "st1 { v21.b }[4], [x28], #0x1\n"
+ "st1 { v5.b }[4], [x11], #0x1\n"
+ "st1 { v21.b }[4], [x10], #0x1\n"
+ "st1 { v20.b }[4], [x9], #0x1\n"
+ "st1 { v19.b }[4], [x28], #0x1\n"
"b 87f\n"
"85:" // Oddments: Bit 2: Unset
"tbz x7, #1, 86f\n"
- "st1 { v15.h }[0], [x11], #0x2\n"
- "st1 { v10.h }[0], [x10], #0x2\n"
- "st1 { v9.h }[0], [x9], #0x2\n"
- "st1 { v21.h }[0], [x28], #0x2\n"
+ "st1 { v5.h }[0], [x11], #0x2\n"
+ "st1 { v21.h }[0], [x10], #0x2\n"
+ "st1 { v20.h }[0], [x9], #0x2\n"
+ "st1 { v19.h }[0], [x28], #0x2\n"
"tbz x7, #0, 87f\n"
- "st1 { v15.b }[2], [x11], #0x1\n"
- "st1 { v10.b }[2], [x10], #0x1\n"
- "st1 { v9.b }[2], [x9], #0x1\n"
- "st1 { v21.b }[2], [x28], #0x1\n"
+ "st1 { v5.b }[2], [x11], #0x1\n"
+ "st1 { v21.b }[2], [x10], #0x1\n"
+ "st1 { v20.b }[2], [x9], #0x1\n"
+ "st1 { v19.b }[2], [x28], #0x1\n"
"b 87f\n"
"86:" // Oddments: Bit 2: Unset: Bit 1: Unset
"tbz x7, #0, 87f\n"
- "st1 { v15.b }[0], [x11], #0x1\n"
- "st1 { v10.b }[0], [x10], #0x1\n"
- "st1 { v9.b }[0], [x9], #0x1\n"
- "st1 { v21.b }[0], [x28], #0x1\n"
+ "st1 { v5.b }[0], [x11], #0x1\n"
+ "st1 { v21.b }[0], [x10], #0x1\n"
+ "st1 { v20.b }[0], [x9], #0x1\n"
+ "st1 { v19.b }[0], [x28], #0x1\n"
"87:" // Oddments: Bit 2: End
"88:" // End
: