aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp716
1 files changed, 358 insertions, 358 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp
index 4913340c4c..c8a599b0a9 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021, 2023 Arm Limited.
+ * Copyright (c) 2021, 2023-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -98,478 +98,478 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
activation_min, activation_max);
__asm__ __volatile__(
- "ldr x21, [%x[params_struct], %[offsetof_args_outptrs]]\n"
- "mov x17, #0x10\n" // cntb _, ALL, #1
- "lsr x9, %x[n_channels], #0x3\n"
+ "ldr x22, [%x[params_struct], %[offsetof_args_outptrs]]\n"
+ "mov x8, #0x10\n" // cntb _, ALL, #1
+ "lsr x17, %x[n_channels], #0x3\n"
"ldr x16, [%x[params_struct], %[offsetof_args_params]]\n"
- "add x20, %x[params_struct], %[offsetof_args_min]\n"
- "ld1r { v27.8h }, [x20]\n"
+ "add x21, %x[params_struct], %[offsetof_args_min]\n"
"add x20, %x[params_struct], %[offsetof_args_max]\n"
+ "ld1r { v27.8h }, [x21]\n"
"ld1r { v15.8h }, [x20]\n"
"add x15, %x[params_struct], %[offsetof_Args_inptrs]\n"
- "ldp x14, x13, [x21, #0x0]\n"
- "ldp x12, x11, [x21, #0x10]\n"
- "mov x10, #0x0\n"
- "sub x28, XZR, x17\n"
- "cbz x9, 3f\n"
+ "mov x14, #0x0\n"
+ "ldp x13, x12, [x22, #0x0]\n"
+ "ldp x11, x10, [x22, #0x10]\n"
+ "sub x9, XZR, x8\n"
+ "cbz x17, 3f\n"
"ldr q26, [x16, #0x0]\n"
"ldr q0, [x16, #0x10]\n"
- "cmp x17, x9, LSL #4\n"
+ "cmp x8, x17, LSL #4\n"
"ldr q1, [x16, #0x20]\n"
"ldr q2, [x16, #0x30]\n"
"ldr q3, [x16, #0x40]\n"
"ldr q4, [x16, #0x50]\n"
"add x16, x16, #0x60\n"
- "ldp x21, x20, [x15, #0x0]\n"
- "ldr q5, [x21, x10]\n"
- "ldr q6, [x20, x10]\n"
- "ldp x21, x20, [x15, #0x10]\n"
- "ldr q7, [x21, x10]\n"
- "ldr q8, [x20, x10]\n"
- "ldp x21, x20, [x15, #0x20]\n"
- "ldr q9, [x21, x10]\n"
- "ldr q13, [x20, x10]\n"
+ "ldp x27, x26, [x15, #0x0]\n"
+ "ldp x25, x24, [x15, #0x10]\n"
+ "ldp x23, x22, [x15, #0x20]\n"
"ldp x21, x20, [x15, #0x30]\n"
- "ldr q11, [x21, x10]\n"
- "ldr q12, [x20, x10]\n"
+ "ldr q5, [x27, x14]\n"
+ "ldr q6, [x26, x14]\n"
+ "ldr q7, [x25, x14]\n"
+ "ldr q8, [x24, x14]\n"
+ "ldr q9, [x23, x14]\n"
+ "ldr q13, [x22, x14]\n"
+ "ldr q11, [x21, x14]\n"
+ "ldr q12, [x20, x14]\n"
"ldp x21, x20, [x15, #0x40]\n"
- "ldr q10, [x21, x10]\n"
- "ldr q14, [x20, x10]\n"
+ "ldr q10, [x21, x14]\n"
+ "ldr q14, [x20, x14]\n"
"bge 2f\n"
"1:" // Channel loop
"mov v30.16b, v26.16b\n fmla v30.8h, v0.8h, v5.8h\n"
"mov v31.16b, v26.16b\n fmla v31.8h, v0.8h, v6.8h\n"
- "ldr x20, [x15, #0x50]\n"
- "ldr q24, [x20, x10]\n"
- "mov v28.16b, v26.16b\n fmla v28.8h, v0.8h, v7.8h\n"
- "mov v29.16b, v26.16b\n fmla v29.8h, v0.8h, v8.8h\n"
+ "ldr x21, [x15, #0x50]\n"
+ "ldr x20, [x15, #0x58]\n"
+ "mov v29.16b, v26.16b\n fmla v29.8h, v0.8h, v7.8h\n"
+ "mov v28.16b, v26.16b\n fmla v28.8h, v0.8h, v8.8h\n"
"ldr q23, [x16, #0x0]\n"
"ldr q26, [x16, #0x140]\n"
+ "ldr x22, [x15, #0x60]\n"
+ "ldr x25, [x15, #0x68]\n"
+ "add x9, x9, #0x10\n"
+ "ldr q22, [x21, x14]\n"
+ "ldr x24, [x15, #0x70]\n"
"fmla v30.8h, v1.8h, v6.8h\n"
+ "ldr q21, [x20, x14]\n"
"fmla v31.8h, v1.8h, v9.8h\n"
- "ldr x20, [x15, #0x58]\n"
- "ldr q22, [x20, x10]\n"
- "fmla v28.8h, v1.8h, v8.8h\n"
- "fmla v29.8h, v1.8h, v13.8h\n"
- "ldr q21, [x16, #0x10]\n"
- "ldr x20, [x15, #0x60]\n"
+ "ldr x21, [x15, #0x78]\n"
+ "fmla v29.8h, v1.8h, v8.8h\n"
+ "fmla v28.8h, v1.8h, v13.8h\n"
+ "ldr q0, [x16, #0x10]\n"
+ "ldr x27, [x15, #0x80]\n"
+ "ldr x20, [x15, #0x88]\n"
+ "ldr x23, [x15, #0x90]\n"
"fmla v30.8h, v2.8h, v9.8h\n"
- "ldr q17, [x20, x10]\n"
+ "ldr q18, [x22, x14]\n"
+ "ldr x26, [x15, #0x98]\n"
"fmla v31.8h, v2.8h, v11.8h\n"
- "ldr x20, [x15, #0x68]\n"
- "fmla v28.8h, v2.8h, v13.8h\n"
- "fmla v29.8h, v2.8h, v24.8h\n"
+ "fmla v29.8h, v2.8h, v13.8h\n"
+ "ldr x22, [x15, #0xa0]\n"
+ "fmla v28.8h, v2.8h, v22.8h\n"
"ldr q16, [x16, #0x20]\n"
- "ldr x22, [x15, #0x70]\n"
"fmla v30.8h, v3.8h, v11.8h\n"
- "ldr q5, [x20, x10]\n"
+ "ldr q20, [x25, x14]\n"
+ "ldr x25, [x15, #0xa8]\n"
"fmla v31.8h, v3.8h, v12.8h\n"
- "ldr x20, [x15, #0x78]\n"
- "fmla v28.8h, v3.8h, v24.8h\n"
"fmla v29.8h, v3.8h, v22.8h\n"
- "ldr q20, [x16, #0x30]\n"
- "ldr x21, [x15, #0x80]\n"
+ "fmla v28.8h, v3.8h, v21.8h\n"
+ "ldr q17, [x16, #0x30]\n"
"fmla v30.8h, v4.8h, v12.8h\n"
- "ldr q19, [x22, x10]\n"
- "fmla v31.8h, v4.8h, v17.8h\n"
- "ldr q2, [x20, x10]\n"
- "fmla v28.8h, v4.8h, v22.8h\n"
- "fmla v29.8h, v4.8h, v10.8h\n"
- "ldr q18, [x16, #0x40]\n"
- "ldr x20, [x15, #0x88]\n"
+ "ldr q3, [x24, x14]\n"
+ "ldr x24, [x15, #0xb0]\n"
+ "fmla v31.8h, v4.8h, v18.8h\n"
+ "ldr q2, [x21, x14]\n"
+ "ldr x21, [x15, #0xb8]\n"
+ "fmla v29.8h, v4.8h, v21.8h\n"
+ "fmla v28.8h, v4.8h, v10.8h\n"
+ "ldr q19, [x16, #0x40]\n"
"fmla v30.8h, v23.8h, v7.8h\n"
"fmla v31.8h, v23.8h, v8.8h\n"
- "ldr x23, [x15, #0x90]\n"
- "ldr x26, [x15, #0x98]\n"
- "fmla v28.8h, v23.8h, v14.8h\n"
- "fmla v29.8h, v23.8h, v5.8h\n"
- "ldr q1, [x16, #0x50]\n"
- "ldr x22, [x15, #0xa0]\n"
- "fmla v30.8h, v21.8h, v8.8h\n"
- "ldr q25, [x20, x10]\n"
- "fmla v31.8h, v21.8h, v13.8h\n"
- "ldr x25, [x15, #0xa8]\n"
- "fmla v28.8h, v21.8h, v5.8h\n"
- "fmla v29.8h, v21.8h, v19.8h\n"
- "ldr q17, [x16, #0x60]\n"
- "ldr x24, [x15, #0xb0]\n"
+ "fmla v29.8h, v23.8h, v14.8h\n"
+ "fmla v28.8h, v23.8h, v20.8h\n"
+ "ldr q18, [x16, #0x50]\n"
+ "fmla v30.8h, v0.8h, v8.8h\n"
+ "ldr q25, [x20, x14]\n"
+ "ldr x28, [x15, #0xc8]\n"
+ "fmla v31.8h, v0.8h, v13.8h\n"
+ "fmla v29.8h, v0.8h, v20.8h\n"
+ "fmla v28.8h, v0.8h, v3.8h\n"
+ "ldr q11, [x16, #0x60]\n"
"fmla v30.8h, v16.8h, v13.8h\n"
- "ldr q8, [x21, x10]\n"
- "fmla v31.8h, v16.8h, v24.8h\n"
- "ldr x20, [x15, #0xb8]\n"
- "fmla v28.8h, v16.8h, v19.8h\n"
- "fmla v29.8h, v16.8h, v2.8h\n"
+ "ldr q24, [x27, x14]\n"
+ "ldr x20, [x15, #0xc0]\n"
+ "fmla v31.8h, v16.8h, v22.8h\n"
+ "fmla v29.8h, v16.8h, v3.8h\n"
+ "fmla v28.8h, v16.8h, v2.8h\n"
"ldr q16, [x16, #0x70]\n"
- "ldr x21, [x15, #0xc0]\n"
- "fmla v30.8h, v20.8h, v24.8h\n"
- "ldr q24, [x23, x10]\n"
- "fmla v31.8h, v20.8h, v22.8h\n"
- "ldr x27, [x15, #0xc8]\n"
- "fmla v28.8h, v20.8h, v2.8h\n"
- "fmla v29.8h, v20.8h, v8.8h\n"
- "ldr q23, [x16, #0x80]\n"
+ "fmla v30.8h, v17.8h, v22.8h\n"
+ "ldr q5, [x23, x14]\n"
"ldr x23, [x15, #0xd0]\n"
- "fmla v30.8h, v18.8h, v22.8h\n"
- "ldr q22, [x26, x10]\n"
- "fmla v31.8h, v18.8h, v10.8h\n"
- "ldr q21, [x22, x10]\n"
- "fmla v28.8h, v18.8h, v8.8h\n"
- "fmla v29.8h, v18.8h, v25.8h\n"
- "ldr q20, [x16, #0x90]\n"
- "ldr x22, [x15, #0xd8]\n"
- "fmla v30.8h, v1.8h, v14.8h\n"
- "ldr q0, [x20, x10]\n"
- "fmla v31.8h, v1.8h, v5.8h\n"
- "ldr x20, [x15, #0xe0]\n"
- "fmla v28.8h, v1.8h, v24.8h\n"
- "fmla v29.8h, v1.8h, v22.8h\n"
- "ldr q6, [x16, #0xa0]\n"
+ "fmla v31.8h, v17.8h, v21.8h\n"
+ "fmla v29.8h, v17.8h, v2.8h\n"
+ "fmla v28.8h, v17.8h, v24.8h\n"
+ "ldr q17, [x16, #0x80]\n"
+ "fmla v30.8h, v19.8h, v21.8h\n"
+ "ldr q23, [x26, x14]\n"
+ "ldr x27, [x15, #0xd8]\n"
+ "fmla v31.8h, v19.8h, v10.8h\n"
+ "ldr q22, [x22, x14]\n"
+ "ldr x22, [x15, #0xe0]\n"
+ "fmla v29.8h, v19.8h, v24.8h\n"
+ "fmla v28.8h, v19.8h, v25.8h\n"
+ "ldr q21, [x16, #0x90]\n"
+ "fmla v30.8h, v18.8h, v14.8h\n"
+ "ldr q1, [x21, x14]\n"
"ldr x26, [x15, #0xf8]\n"
- "fmla v30.8h, v17.8h, v5.8h\n"
- "ldr q1, [x25, x10]\n"
- "fmla v31.8h, v17.8h, v19.8h\n"
+ "fmla v31.8h, v18.8h, v20.8h\n"
+ "fmla v29.8h, v18.8h, v5.8h\n"
+ "fmla v28.8h, v18.8h, v23.8h\n"
+ "ldr q12, [x16, #0xa0]\n"
+ "fmla v30.8h, v11.8h, v20.8h\n"
+ "ldr q0, [x25, x14]\n"
"ldr x25, [x15, #0xe8]\n"
- "fmla v28.8h, v17.8h, v22.8h\n"
- "fmla v29.8h, v17.8h, v21.8h\n"
- "ldr q18, [x16, #0xb0]\n"
- "add x28, x28, #0x10\n"
- "fmla v30.8h, v16.8h, v19.8h\n"
- "ldr q19, [x24, x10]\n"
- "fmla v31.8h, v16.8h, v2.8h\n"
+ "fmla v31.8h, v11.8h, v3.8h\n"
+ "fmla v29.8h, v11.8h, v23.8h\n"
+ "fmla v28.8h, v11.8h, v22.8h\n"
+ "ldr q20, [x16, #0xb0]\n"
+ "fmla v30.8h, v16.8h, v3.8h\n"
+ "ldr q19, [x24, x14]\n"
"ldr x24, [x15, #0xf0]\n"
- "fmla v28.8h, v16.8h, v21.8h\n"
- "fmla v29.8h, v16.8h, v1.8h\n"
- "ldr q17, [x16, #0xc0]\n"
- "fmla v30.8h, v23.8h, v2.8h\n"
- "ldr q16, [x21, x10]\n"
- "fmla v31.8h, v23.8h, v8.8h\n"
+ "fmla v31.8h, v16.8h, v2.8h\n"
+ "fmla v29.8h, v16.8h, v22.8h\n"
+ "fmla v28.8h, v16.8h, v0.8h\n"
+ "ldr q18, [x16, #0xc0]\n"
+ "fmla v30.8h, v17.8h, v2.8h\n"
+ "ldr q16, [x20, x14]\n"
"ldr x21, [x15, #0x100]\n"
- "fmla v28.8h, v23.8h, v1.8h\n"
- "fmla v29.8h, v23.8h, v19.8h\n"
- "ldr q13, [x16, #0xd0]\n"
- "fmla v30.8h, v20.8h, v8.8h\n"
- "ldr q2, [x27, x10]\n"
- "fmla v31.8h, v20.8h, v25.8h\n"
- "ldr q10, [x20, x10]\n"
- "fmla v28.8h, v20.8h, v19.8h\n"
- "fmla v29.8h, v20.8h, v0.8h\n"
- "ldr q9, [x16, #0xe0]\n"
+ "fmla v31.8h, v17.8h, v24.8h\n"
+ "fmla v29.8h, v17.8h, v0.8h\n"
+ "fmla v28.8h, v17.8h, v19.8h\n"
+ "ldr q17, [x16, #0xd0]\n"
+ "fmla v30.8h, v21.8h, v24.8h\n"
+ "ldr q14, [x28, x14]\n"
"ldr x20, [x15, #0x108]\n"
- "fmla v30.8h, v6.8h, v24.8h\n"
- "ldr q5, [x23, x10]\n"
- "fmla v31.8h, v6.8h, v22.8h\n"
+ "fmla v31.8h, v21.8h, v25.8h\n"
+ "ldr q4, [x22, x14]\n"
+ "fmla v29.8h, v21.8h, v19.8h\n"
+ "fmla v28.8h, v21.8h, v1.8h\n"
+ "ldr q7, [x16, #0xe0]\n"
+ "fmla v30.8h, v12.8h, v5.8h\n"
+ "ldr q25, [x23, x14]\n"
"ldr x23, [x15, #0x110]\n"
- "fmla v28.8h, v6.8h, v16.8h\n"
- "fmla v29.8h, v6.8h, v2.8h\n"
- "ldr q24, [x16, #0xf0]\n"
- "fmla v30.8h, v18.8h, v22.8h\n"
- "ldr q25, [x22, x10]\n"
- "fmla v31.8h, v18.8h, v21.8h\n"
+ "fmla v31.8h, v12.8h, v23.8h\n"
+ "fmla v29.8h, v12.8h, v16.8h\n"
+ "fmla v28.8h, v12.8h, v14.8h\n"
+ "ldr q11, [x16, #0xf0]\n"
+ "fmla v30.8h, v20.8h, v23.8h\n"
+ "ldr q24, [x27, x14]\n"
"ldr x22, [x15, #0x118]\n"
- "fmla v28.8h, v18.8h, v2.8h\n"
- "fmla v29.8h, v18.8h, v5.8h\n"
+ "fmla v31.8h, v20.8h, v22.8h\n"
+ "fmla v29.8h, v20.8h, v14.8h\n"
+ "fmla v28.8h, v20.8h, v25.8h\n"
"ldr q23, [x16, #0x100]\n"
- "fmla v30.8h, v17.8h, v21.8h\n"
- "ldr q22, [x25, x10]\n"
- "fmla v31.8h, v17.8h, v1.8h\n"
- "fmla v28.8h, v17.8h, v5.8h\n"
- "fmla v29.8h, v17.8h, v25.8h\n"
+ "fmla v30.8h, v18.8h, v22.8h\n"
+ "ldr q22, [x25, x14]\n"
+ "fmla v31.8h, v18.8h, v0.8h\n"
+ "fmla v29.8h, v18.8h, v25.8h\n"
+ "fmla v28.8h, v18.8h, v24.8h\n"
"ldr q21, [x16, #0x110]\n"
- "fmla v30.8h, v13.8h, v1.8h\n"
- "ldr q18, [x24, x10]\n"
- "fmla v31.8h, v13.8h, v19.8h\n"
- "fmla v28.8h, v13.8h, v25.8h\n"
- "fmla v29.8h, v13.8h, v10.8h\n"
+ "fmla v30.8h, v17.8h, v0.8h\n"
+ "ldr q18, [x24, x14]\n"
+ "fmla v31.8h, v17.8h, v19.8h\n"
+ "fmla v29.8h, v17.8h, v24.8h\n"
+ "fmla v28.8h, v17.8h, v4.8h\n"
"ldr q20, [x16, #0x120]\n"
- "fmla v30.8h, v9.8h, v19.8h\n"
- "ldr q17, [x26, x10]\n"
- "fmla v31.8h, v9.8h, v0.8h\n"
- "fmla v28.8h, v9.8h, v10.8h\n"
- "fmla v29.8h, v9.8h, v22.8h\n"
+ "fmla v30.8h, v7.8h, v19.8h\n"
+ "ldr q17, [x26, x14]\n"
+ "fmla v31.8h, v7.8h, v1.8h\n"
+ "fmla v29.8h, v7.8h, v4.8h\n"
+ "fmla v28.8h, v7.8h, v22.8h\n"
"ldr q19, [x16, #0x130]\n"
- "fmla v30.8h, v24.8h, v16.8h\n"
- "ldr q16, [x21, x10]\n"
- "fmla v31.8h, v24.8h, v2.8h\n"
- "fmla v28.8h, v24.8h, v18.8h\n"
- "ldr q18, [x20, x10]\n"
- "fmla v29.8h, v24.8h, v17.8h\n"
+ "fmla v30.8h, v11.8h, v16.8h\n"
+ "ldr q16, [x21, x14]\n"
+ "fmla v31.8h, v11.8h, v14.8h\n"
+ "fmla v29.8h, v11.8h, v18.8h\n"
+ "ldr q18, [x20, x14]\n"
+ "ldp x20, x21, [x15, #0x0]\n"
+ "fmla v28.8h, v11.8h, v17.8h\n"
"ldr q0, [x16, #0x150]\n"
- "fmla v30.8h, v23.8h, v2.8h\n"
- "fmla v31.8h, v23.8h, v5.8h\n"
- "ldp x21, x20, [x15, #0x0]\n"
- "fmla v28.8h, v23.8h, v17.8h\n"
- "ldr q17, [x23, x10]\n"
- "fmla v29.8h, v23.8h, v16.8h\n"
+ "fmla v30.8h, v23.8h, v14.8h\n"
+ "fmla v31.8h, v23.8h, v25.8h\n"
+ "fmla v29.8h, v23.8h, v17.8h\n"
+ "ldr q17, [x23, x14]\n"
+ "fmla v28.8h, v23.8h, v16.8h\n"
"ldr q1, [x16, #0x160]\n"
- "fmla v30.8h, v21.8h, v5.8h\n"
- "ldr q5, [x21, x17]\n"
- "fmla v31.8h, v21.8h, v25.8h\n"
- "fmla v28.8h, v21.8h, v16.8h\n"
- "ldr q16, [x22, x10]\n"
- "fmla v29.8h, v21.8h, v18.8h\n"
+ "fmla v30.8h, v21.8h, v25.8h\n"
+ "ldr q5, [x20, x8]\n"
+ "fmla v31.8h, v21.8h, v24.8h\n"
+ "fmla v29.8h, v21.8h, v16.8h\n"
+ "ldr q16, [x22, x14]\n"
+ "ldp x20, x26, [x15, #0x10]\n"
+ "ldp x25, x24, [x15, #0x20]\n"
+ "ldp x23, x22, [x15, #0x30]\n"
+ "add x14, x14, #0x10\n"
+ "ldr q7, [x20, x8]\n"
+ "fmla v28.8h, v21.8h, v18.8h\n"
"ldr q2, [x16, #0x170]\n"
- "fmla v30.8h, v20.8h, v25.8h\n"
- "ldr q6, [x20, x17]\n"
- "fmla v31.8h, v20.8h, v10.8h\n"
- "ldp x21, x20, [x15, #0x10]\n"
- "ldr q7, [x21, x17]\n"
- "fmla v28.8h, v20.8h, v18.8h\n"
- "fmla v29.8h, v20.8h, v17.8h\n"
+ "fmla v30.8h, v20.8h, v24.8h\n"
+ "ldr q6, [x21, x8]\n"
+ "ldp x21, x20, [x15, #0x40]\n"
+ "ldr q13, [x24, x8]\n"
+ "fmla v31.8h, v20.8h, v4.8h\n"
+ "fmla v29.8h, v20.8h, v18.8h\n"
+ "ldr q11, [x23, x8]\n"
+ "ldr q14, [x20, x8]\n"
+ "fmla v28.8h, v20.8h, v17.8h\n"
"ldr q3, [x16, #0x180]\n"
- "fmla v30.8h, v19.8h, v10.8h\n"
- "ldr q8, [x20, x17]\n"
+ "fmla v30.8h, v19.8h, v4.8h\n"
+ "ldr q8, [x26, x8]\n"
"fmla v31.8h, v19.8h, v22.8h\n"
- "ldp x21, x20, [x15, #0x20]\n"
- "ldr q13, [x20, x17]\n"
- "fmla v28.8h, v19.8h, v17.8h\n"
- "fmla v29.8h, v19.8h, v16.8h\n"
- "ldr q9, [x21, x17]\n"
+ "ldr q10, [x21, x8]\n"
+ "fmla v29.8h, v19.8h, v17.8h\n"
+ "ldr q12, [x22, x8]\n"
+ "fmla v28.8h, v19.8h, v16.8h\n"
+ "ldr q9, [x25, x8]\n"
+ "add x8, x8, #0x10\n"
"ldr q4, [x16, #0x190]\n"
- "ldp x21, x20, [x15, #0x30]\n"
+ "cmp x8, x17, LSL #4\n"
+ "add x16, x16, #0x1a0\n"
"fmax v30.8h, v30.8h, v27.8h\n"
"fmax v31.8h, v31.8h, v27.8h\n"
- "ldr q11, [x21, x17]\n"
- "ldr q12, [x20, x17]\n"
- "fmax v28.8h, v28.8h, v27.8h\n"
"fmax v29.8h, v29.8h, v27.8h\n"
- "ldp x21, x20, [x15, #0x40]\n"
- "ldr q10, [x21, x17]\n"
+ "fmax v28.8h, v28.8h, v27.8h\n"
"fmin v30.8h, v30.8h, v15.8h\n"
"fmin v31.8h, v31.8h, v15.8h\n"
- "ldr q14, [x20, x17]\n"
- "add x17, x17, #0x10\n"
- "cmp x17, x9, LSL #4\n"
- "fmin v28.8h, v28.8h, v15.8h\n"
"fmin v29.8h, v29.8h, v15.8h\n"
- "add x10, x10, #0x10\n"
- "str q30, [x14, x28]\n"
- "add x16, x16, #0x1a0\n"
- "str q31, [x13, x28]\n"
- "str q28, [x12, x28]\n"
- "str q29, [x11, x28]\n"
+ "fmin v28.8h, v28.8h, v15.8h\n"
+ "str q30, [x13, x9]\n"
+ "str q31, [x12, x9]\n"
+ "str q29, [x11, x9]\n"
+ "str q28, [x10, x9]\n"
"blt 1b\n"
"2:" // Channel tail
"mov v31.16b, v26.16b\n fmla v31.8h, v0.8h, v5.8h\n"
"mov v5.16b, v26.16b\n fmla v5.8h, v0.8h, v6.8h\n"
- "ldr x20, [x15, #0x50]\n"
- "ldr q22, [x20, x10]\n"
+ "ldr x22, [x15, #0x50]\n"
+ "ldr x21, [x15, #0x58]\n"
"mov v30.16b, v26.16b\n fmla v30.8h, v0.8h, v7.8h\n"
"mov v29.16b, v26.16b\n fmla v29.8h, v0.8h, v8.8h\n"
"ldr q19, [x16, #0x0]\n"
- "ldr x20, [x15, #0x58]\n"
+ "ldr x20, [x15, #0x60]\n"
+ "ldr x25, [x15, #0x68]\n"
+ "ldr x24, [x15, #0x70]\n"
+ "add x9, x9, #0x10\n"
+ "ldr q22, [x22, x14]\n"
+ "ldr x23, [x15, #0x78]\n"
"fmla v31.8h, v1.8h, v6.8h\n"
- "ldr q21, [x20, x10]\n"
+ "ldr q21, [x21, x14]\n"
"fmla v5.8h, v1.8h, v9.8h\n"
- "ldr x21, [x15, #0x60]\n"
+ "ldr x27, [x15, #0x80]\n"
"fmla v30.8h, v1.8h, v8.8h\n"
"fmla v29.8h, v1.8h, v13.8h\n"
"ldr q18, [x16, #0x10]\n"
- "ldr x20, [x15, #0x68]\n"
+ "ldr x22, [x15, #0x88]\n"
+ "ldr x21, [x15, #0x90]\n"
+ "ldr x26, [x15, #0x98]\n"
"fmla v31.8h, v2.8h, v9.8h\n"
- "ldr q16, [x21, x10]\n"
+ "ldr q16, [x20, x14]\n"
+ "ldr x20, [x15, #0xa0]\n"
"fmla v5.8h, v2.8h, v11.8h\n"
- "ldr x23, [x15, #0x70]\n"
"fmla v30.8h, v2.8h, v13.8h\n"
"fmla v29.8h, v2.8h, v22.8h\n"
"ldr q17, [x16, #0x20]\n"
- "ldr x21, [x15, #0x78]\n"
"fmla v31.8h, v3.8h, v11.8h\n"
- "ldr q6, [x20, x10]\n"
+ "ldr q6, [x25, x14]\n"
+ "ldr x25, [x15, #0xa8]\n"
"fmla v5.8h, v3.8h, v12.8h\n"
- "ldr x22, [x15, #0x80]\n"
"fmla v30.8h, v3.8h, v22.8h\n"
"fmla v29.8h, v3.8h, v21.8h\n"
"ldr q20, [x16, #0x30]\n"
- "ldr x20, [x15, #0x88]\n"
"fmla v31.8h, v4.8h, v12.8h\n"
- "ldr q2, [x23, x10]\n"
+ "ldr q2, [x24, x14]\n"
+ "ldr x24, [x15, #0xb0]\n"
"fmla v5.8h, v4.8h, v16.8h\n"
- "ldr q28, [x21, x10]\n"
+ "ldr q28, [x23, x14]\n"
+ "ldr x23, [x15, #0xb8]\n"
"fmla v30.8h, v4.8h, v21.8h\n"
"fmla v29.8h, v4.8h, v10.8h\n"
"ldr q16, [x16, #0x40]\n"
- "ldr x21, [x15, #0x90]\n"
"fmla v31.8h, v19.8h, v7.8h\n"
"fmla v5.8h, v19.8h, v8.8h\n"
- "ldr x27, [x15, #0x98]\n"
- "ldr x26, [x15, #0xa0]\n"
"fmla v30.8h, v19.8h, v14.8h\n"
"fmla v29.8h, v19.8h, v6.8h\n"
"ldr q19, [x16, #0x50]\n"
- "ldr x25, [x15, #0xa8]\n"
"fmla v31.8h, v18.8h, v8.8h\n"
- "ldr q1, [x20, x10]\n"
+ "ldr q1, [x22, x14]\n"
+ "ldr x28, [x15, #0xc8]\n"
"fmla v5.8h, v18.8h, v13.8h\n"
- "ldr x24, [x15, #0xb0]\n"
"fmla v30.8h, v18.8h, v6.8h\n"
"fmla v29.8h, v18.8h, v2.8h\n"
"ldr q18, [x16, #0x60]\n"
- "ldr x20, [x15, #0xb8]\n"
"fmla v31.8h, v17.8h, v13.8h\n"
- "ldr q26, [x22, x10]\n"
+ "ldr q26, [x27, x14]\n"
+ "ldr x22, [x15, #0xc0]\n"
"fmla v5.8h, v17.8h, v22.8h\n"
- "ldr x23, [x15, #0xc0]\n"
"fmla v30.8h, v17.8h, v2.8h\n"
"fmla v29.8h, v17.8h, v28.8h\n"
"ldr q17, [x16, #0x70]\n"
- "ldr x22, [x15, #0xc8]\n"
"fmla v31.8h, v20.8h, v22.8h\n"
- "ldr q25, [x21, x10]\n"
- "fmla v5.8h, v20.8h, v21.8h\n"
+ "ldr q25, [x21, x14]\n"
"ldr x21, [x15, #0xd0]\n"
+ "fmla v5.8h, v20.8h, v21.8h\n"
"fmla v30.8h, v20.8h, v28.8h\n"
"fmla v29.8h, v20.8h, v26.8h\n"
"ldr q24, [x16, #0x80]\n"
- "add x28, x28, #0x10\n"
"fmla v31.8h, v16.8h, v21.8h\n"
- "ldr q23, [x27, x10]\n"
+ "ldr q23, [x26, x14]\n"
+ "ldr x27, [x15, #0xd8]\n"
"fmla v5.8h, v16.8h, v10.8h\n"
- "ldr q0, [x26, x10]\n"
+ "ldr q0, [x20, x14]\n"
+ "ldr x20, [x15, #0xe0]\n"
"fmla v30.8h, v16.8h, v26.8h\n"
"fmla v29.8h, v16.8h, v1.8h\n"
"ldr q22, [x16, #0x90]\n"
- "ldr x27, [x15, #0xd8]\n"
"fmla v31.8h, v19.8h, v14.8h\n"
- "ldr q16, [x20, x10]\n"
+ "ldr q16, [x23, x14]\n"
+ "ldr x26, [x15, #0xf8]\n"
"fmla v5.8h, v19.8h, v6.8h\n"
- "ldr x20, [x15, #0xe0]\n"
"fmla v30.8h, v19.8h, v25.8h\n"
"fmla v29.8h, v19.8h, v23.8h\n"
"ldr q21, [x16, #0xa0]\n"
- "ldr x26, [x15, #0xf8]\n"
"fmla v31.8h, v18.8h, v6.8h\n"
- "ldr q20, [x25, x10]\n"
- "fmla v5.8h, v18.8h, v2.8h\n"
+ "ldr q20, [x25, x14]\n"
"ldr x25, [x15, #0xe8]\n"
+ "fmla v5.8h, v18.8h, v2.8h\n"
"fmla v30.8h, v18.8h, v23.8h\n"
"fmla v29.8h, v18.8h, v0.8h\n"
"ldr q18, [x16, #0xb0]\n"
"fmla v31.8h, v17.8h, v2.8h\n"
- "ldr q19, [x24, x10]\n"
- "fmla v5.8h, v17.8h, v28.8h\n"
+ "ldr q19, [x24, x14]\n"
"ldr x24, [x15, #0xf0]\n"
+ "fmla v5.8h, v17.8h, v28.8h\n"
"fmla v30.8h, v17.8h, v0.8h\n"
"fmla v29.8h, v17.8h, v20.8h\n"
"ldr q17, [x16, #0xc0]\n"
"fmla v31.8h, v24.8h, v28.8h\n"
- "ldr q7, [x23, x10]\n"
- "fmla v5.8h, v24.8h, v26.8h\n"
+ "ldr q10, [x22, x14]\n"
"ldr x23, [x15, #0x100]\n"
+ "fmla v5.8h, v24.8h, v26.8h\n"
"fmla v30.8h, v24.8h, v20.8h\n"
"fmla v29.8h, v24.8h, v19.8h\n"
- "ldr q3, [x16, #0xd0]\n"
+ "ldr q13, [x16, #0xd0]\n"
"fmla v31.8h, v22.8h, v26.8h\n"
- "ldr q28, [x22, x10]\n"
+ "ldr q28, [x28, x14]\n"
+ "ldr x22, [x15, #0x108]\n"
"fmla v5.8h, v22.8h, v1.8h\n"
- "ldr q13, [x20, x10]\n"
+ "ldr q14, [x20, x14]\n"
"fmla v30.8h, v22.8h, v19.8h\n"
"fmla v29.8h, v22.8h, v16.8h\n"
- "ldr q11, [x16, #0xe0]\n"
- "ldr x22, [x15, #0x108]\n"
+ "ldr q12, [x16, #0xe0]\n"
"fmla v31.8h, v21.8h, v25.8h\n"
- "ldr q26, [x21, x10]\n"
- "fmla v5.8h, v21.8h, v23.8h\n"
+ "ldr q26, [x21, x14]\n"
"ldr x21, [x15, #0x110]\n"
- "fmla v30.8h, v21.8h, v7.8h\n"
+ "fmla v5.8h, v21.8h, v23.8h\n"
+ "fmla v30.8h, v21.8h, v10.8h\n"
"fmla v29.8h, v21.8h, v28.8h\n"
"ldr q25, [x16, #0xf0]\n"
"fmla v31.8h, v18.8h, v23.8h\n"
- "ldr q24, [x27, x10]\n"
- "fmla v5.8h, v18.8h, v0.8h\n"
+ "ldr q24, [x27, x14]\n"
"ldr x20, [x15, #0x118]\n"
+ "fmla v5.8h, v18.8h, v0.8h\n"
"fmla v30.8h, v18.8h, v28.8h\n"
"fmla v29.8h, v18.8h, v26.8h\n"
"ldr q23, [x16, #0x100]\n"
"fmla v31.8h, v17.8h, v0.8h\n"
- "ldr q22, [x25, x10]\n"
+ "ldr q22, [x25, x14]\n"
"fmla v5.8h, v17.8h, v20.8h\n"
"fmla v30.8h, v17.8h, v26.8h\n"
"fmla v29.8h, v17.8h, v24.8h\n"
"ldr q21, [x16, #0x110]\n"
- "fmla v31.8h, v3.8h, v20.8h\n"
- "ldr q18, [x24, x10]\n"
- "fmla v5.8h, v3.8h, v19.8h\n"
- "fmla v30.8h, v3.8h, v24.8h\n"
- "fmla v29.8h, v3.8h, v13.8h\n"
+ "fmla v31.8h, v13.8h, v20.8h\n"
+ "ldr q18, [x24, x14]\n"
+ "fmla v5.8h, v13.8h, v19.8h\n"
+ "fmla v30.8h, v13.8h, v24.8h\n"
+ "fmla v29.8h, v13.8h, v14.8h\n"
"ldr q20, [x16, #0x120]\n"
- "fmla v31.8h, v11.8h, v19.8h\n"
- "ldr q17, [x26, x10]\n"
- "fmla v5.8h, v11.8h, v16.8h\n"
- "fmla v30.8h, v11.8h, v13.8h\n"
- "fmla v29.8h, v11.8h, v22.8h\n"
+ "fmla v31.8h, v12.8h, v19.8h\n"
+ "ldr q17, [x26, x14]\n"
+ "fmla v5.8h, v12.8h, v16.8h\n"
+ "fmla v30.8h, v12.8h, v14.8h\n"
+ "fmla v29.8h, v12.8h, v22.8h\n"
"ldr q19, [x16, #0x130]\n"
"add x16, x16, #0x140\n"
- "fmla v31.8h, v25.8h, v7.8h\n"
- "ldr q16, [x23, x10]\n"
+ "fmla v31.8h, v25.8h, v10.8h\n"
+ "ldr q16, [x23, x14]\n"
"fmla v5.8h, v25.8h, v28.8h\n"
"fmla v30.8h, v25.8h, v18.8h\n"
- "ldr q18, [x22, x10]\n"
+ "ldr q18, [x22, x14]\n"
"fmla v29.8h, v25.8h, v17.8h\n"
"fmla v31.8h, v23.8h, v28.8h\n"
"fmla v5.8h, v23.8h, v26.8h\n"
"fmla v30.8h, v23.8h, v17.8h\n"
- "ldr q17, [x21, x10]\n"
+ "ldr q17, [x21, x14]\n"
"fmla v29.8h, v23.8h, v16.8h\n"
"fmla v31.8h, v21.8h, v26.8h\n"
"fmla v5.8h, v21.8h, v24.8h\n"
"fmla v30.8h, v21.8h, v16.8h\n"
- "ldr q16, [x20, x10]\n"
+ "ldr q16, [x20, x14]\n"
+ "add x14, x14, #0x10\n"
"fmla v29.8h, v21.8h, v18.8h\n"
- "add x10, x10, #0x10\n"
"fmla v31.8h, v20.8h, v24.8h\n"
- "fmla v5.8h, v20.8h, v13.8h\n"
+ "fmla v5.8h, v20.8h, v14.8h\n"
"fmla v30.8h, v20.8h, v18.8h\n"
"fmla v29.8h, v20.8h, v17.8h\n"
- "fmla v31.8h, v19.8h, v13.8h\n"
+ "fmla v31.8h, v19.8h, v14.8h\n"
"fmla v5.8h, v19.8h, v22.8h\n"
- "fmax v31.8h, v31.8h, v27.8h\n"
"fmla v30.8h, v19.8h, v17.8h\n"
"fmla v29.8h, v19.8h, v16.8h\n"
+ "fmax v31.8h, v31.8h, v27.8h\n"
"fmax v5.8h, v5.8h, v27.8h\n"
"fmax v30.8h, v30.8h, v27.8h\n"
- "fmax v29.8h, v29.8h, v27.8h\n"
"fmin v31.8h, v31.8h, v15.8h\n"
+ "fmax v29.8h, v29.8h, v27.8h\n"
"fmin v5.8h, v5.8h, v15.8h\n"
- "str q31, [x14, x28]\n"
"fmin v30.8h, v30.8h, v15.8h\n"
"fmin v29.8h, v29.8h, v15.8h\n"
- "str q5, [x13, x28]\n"
- "str q30, [x12, x28]\n"
- "str q29, [x11, x28]\n"
+ "str q31, [x13, x9]\n"
+ "str q5, [x12, x9]\n"
+ "str q30, [x11, x9]\n"
+ "str q29, [x10, x9]\n"
"3:" // Oddments
"tst %x[n_channels], #0x7\n"
"beq 116f\n"
"ldr q26, [x16, #0x0]\n"
"ldr q0, [x16, #0x10]\n"
- "mov x20, x10\n"
- "add x14, x14, x20\n"
+ "mov x20, x14\n"
"ldr q1, [x16, #0x20]\n"
"ldr q2, [x16, #0x30]\n"
- "add x13, x13, x20\n"
- "add x12, x12, x20\n"
"ldr q3, [x16, #0x40]\n"
"ldr q4, [x16, #0x50]\n"
- "add x11, x11, x20\n"
"ldr x9, [x15, #0x0]\n"
"ldr x28, [x15, #0x8]\n"
- "add x9, x9, x10\n"
- "add x28, x28, x10\n"
+ "add x13, x13, x20\n"
+ "add x12, x12, x20\n"
"ldr x27, [x15, #0x10]\n"
"ldr x26, [x15, #0x18]\n"
- "add x27, x27, x10\n"
- "add x26, x26, x10\n"
+ "add x11, x11, x20\n"
+ "add x10, x10, x20\n"
"ldr x25, [x15, #0x20]\n"
"ldr x24, [x15, #0x28]\n"
- "add x25, x25, x10\n"
- "add x24, x24, x10\n"
"ldr x23, [x15, #0x30]\n"
"ldr x22, [x15, #0x38]\n"
- "add x23, x23, x10\n"
- "add x22, x22, x10\n"
+ "add x9, x9, x14\n"
+ "add x28, x28, x14\n"
"ldr x21, [x15, #0x40]\n"
"ldr x20, [x15, #0x48]\n"
- "add x21, x21, x10\n"
- "add x20, x20, x10\n"
+ "add x27, x27, x14\n"
+ "add x26, x26, x14\n"
+ "add x25, x25, x14\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"
"add x16, x16, #0x60\n"
"tbz %x[n_channels], #2, 5f\n"
"ld1 { v5.d }[0], [x9], #0x8\n"
@@ -657,9 +657,9 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"mov v28.16b, v26.16b\n fmla v28.8h, v0.8h, v5.8h\n"
"mov v29.16b, v26.16b\n fmla v29.8h, v0.8h, v6.8h\n"
"ldr x20, [x15, #0x50]\n"
- "add x20, x20, x10\n"
"mov v30.16b, v26.16b\n fmla v30.8h, v0.8h, v7.8h\n"
"mov v31.16b, v26.16b\n fmla v31.8h, v0.8h, v8.8h\n"
+ "add x20, x20, x14\n"
"fmla v28.8h, v1.8h, v6.8h\n"
"fmla v29.8h, v1.8h, v9.8h\n"
"fmla v30.8h, v1.8h, v8.8h\n"
@@ -690,9 +690,9 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr x20, [x15, #0x58]\n"
"fmla v31.8h, v2.8h, v5.8h\n"
"fmla v28.8h, v3.8h, v11.8h\n"
- "add x20, x20, x10\n"
"fmla v29.8h, v3.8h, v12.8h\n"
"fmla v30.8h, v3.8h, v5.8h\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 13f\n"
"ld1 { v6.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 12f\n"
@@ -716,7 +716,7 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr x20, [x15, #0x60]\n"
"fmla v31.8h, v3.8h, v6.8h\n"
"fmla v28.8h, v4.8h, v12.8h\n"
- "add x20, x20, x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 17f\n"
"ld1 { v9.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 16f\n"
@@ -742,11 +742,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"fmla v30.8h, v4.8h, v6.8h\n"
"ldr x20, [x15, #0x68]\n"
"fmla v31.8h, v4.8h, v10.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v0.8h, v7.8h\n"
- "add x20, x20, x10\n"
+ "add x20, x20, x14\n"
"fmla v29.8h, v0.8h, v8.8h\n"
"fmla v30.8h, v0.8h, v14.8h\n"
- "add x16, x16, #0x10\n"
"tbz %x[n_channels], #2, 21f\n"
"ld1 { v11.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 20f\n"
@@ -770,11 +770,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q1, [x16, #0x0]\n"
"ldr x20, [x15, #0x70]\n"
"fmla v31.8h, v0.8h, v11.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v1.8h, v8.8h\n"
"fmla v29.8h, v1.8h, v13.8h\n"
"fmla v30.8h, v1.8h, v11.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 25f\n"
"ld1 { v12.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 24f\n"
@@ -798,11 +798,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q2, [x16, #0x0]\n"
"ldr x20, [x15, #0x78]\n"
"fmla v31.8h, v1.8h, v12.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v2.8h, v13.8h\n"
"fmla v29.8h, v2.8h, v5.8h\n"
"fmla v30.8h, v2.8h, v12.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 29f\n"
"ld1 { v9.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 28f\n"
@@ -826,11 +826,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q3, [x16, #0x0]\n"
"ldr x20, [x15, #0x80]\n"
"fmla v31.8h, v2.8h, v9.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v3.8h, v5.8h\n"
"fmla v29.8h, v3.8h, v6.8h\n"
"fmla v30.8h, v3.8h, v9.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 33f\n"
"ld1 { v13.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 32f\n"
@@ -854,11 +854,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q4, [x16, #0x0]\n"
"ldr x20, [x15, #0x88]\n"
"fmla v31.8h, v3.8h, v13.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v4.8h, v6.8h\n"
"fmla v29.8h, v4.8h, v10.8h\n"
"fmla v30.8h, v4.8h, v13.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 37f\n"
"ld1 { v8.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 36f\n"
@@ -882,10 +882,10 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q0, [x16, #0x0]\n"
"ldr x20, [x15, #0x90]\n"
"fmla v31.8h, v4.8h, v8.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v0.8h, v14.8h\n"
"fmla v29.8h, v0.8h, v11.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 41f\n"
"ld1 { v5.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 40f\n"
@@ -908,7 +908,7 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"43:" // Oddments: Load input (3, 0): Bit 2: End
"ldr x20, [x15, #0x98]\n"
"fmla v30.8h, v0.8h, v5.8h\n"
- "add x20, x20, x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 45f\n"
"ld1 { v6.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 44f\n"
@@ -932,11 +932,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q1, [x16, #0x0]\n"
"ldr x20, [x15, #0xa0]\n"
"fmla v31.8h, v0.8h, v6.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v1.8h, v11.8h\n"
"fmla v29.8h, v1.8h, v12.8h\n"
"fmla v30.8h, v1.8h, v6.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 49f\n"
"ld1 { v10.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 48f\n"
@@ -960,11 +960,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q2, [x16, #0x0]\n"
"ldr x20, [x15, #0xa8]\n"
"fmla v31.8h, v1.8h, v10.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v2.8h, v12.8h\n"
"fmla v29.8h, v2.8h, v9.8h\n"
"fmla v30.8h, v2.8h, v10.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 53f\n"
"ld1 { v11.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 52f\n"
@@ -988,11 +988,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q3, [x16, #0x0]\n"
"ldr x20, [x15, #0xb0]\n"
"fmla v31.8h, v2.8h, v11.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v3.8h, v9.8h\n"
"fmla v29.8h, v3.8h, v13.8h\n"
"fmla v30.8h, v3.8h, v11.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 57f\n"
"ld1 { v12.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 56f\n"
@@ -1016,11 +1016,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q4, [x16, #0x0]\n"
"ldr x20, [x15, #0xb8]\n"
"fmla v31.8h, v3.8h, v12.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v4.8h, v13.8h\n"
"fmla v29.8h, v4.8h, v8.8h\n"
"fmla v30.8h, v4.8h, v12.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 61f\n"
"ld1 { v14.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 60f\n"
@@ -1044,10 +1044,10 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q0, [x16, #0x0]\n"
"ldr x20, [x15, #0xc0]\n"
"fmla v31.8h, v4.8h, v14.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v0.8h, v5.8h\n"
"fmla v29.8h, v0.8h, v6.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 65f\n"
"ld1 { v9.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 64f\n"
@@ -1070,7 +1070,7 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"67:" // Oddments: Load input (4, 0): Bit 2: End
"ldr x20, [x15, #0xc8]\n"
"fmla v30.8h, v0.8h, v9.8h\n"
- "add x20, x20, x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 69f\n"
"ld1 { v13.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 68f\n"
@@ -1094,11 +1094,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q1, [x16, #0x0]\n"
"ldr x20, [x15, #0xd0]\n"
"fmla v31.8h, v0.8h, v13.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v1.8h, v6.8h\n"
"fmla v29.8h, v1.8h, v10.8h\n"
"fmla v30.8h, v1.8h, v13.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 73f\n"
"ld1 { v5.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 72f\n"
@@ -1122,11 +1122,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q2, [x16, #0x0]\n"
"ldr x20, [x15, #0xd8]\n"
"fmla v31.8h, v1.8h, v5.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v2.8h, v10.8h\n"
"fmla v29.8h, v2.8h, v11.8h\n"
"fmla v30.8h, v2.8h, v5.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 77f\n"
"ld1 { v6.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 76f\n"
@@ -1150,11 +1150,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q3, [x16, #0x0]\n"
"ldr x20, [x15, #0xe0]\n"
"fmla v31.8h, v2.8h, v6.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v3.8h, v11.8h\n"
"fmla v29.8h, v3.8h, v12.8h\n"
"fmla v30.8h, v3.8h, v6.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 81f\n"
"ld1 { v8.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 80f\n"
@@ -1178,11 +1178,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q4, [x16, #0x0]\n"
"ldr x20, [x15, #0xe8]\n"
"fmla v31.8h, v3.8h, v8.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v4.8h, v12.8h\n"
"fmla v29.8h, v4.8h, v14.8h\n"
"fmla v30.8h, v4.8h, v8.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 85f\n"
"ld1 { v10.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 84f\n"
@@ -1206,10 +1206,10 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q0, [x16, #0x0]\n"
"ldr x20, [x15, #0xf0]\n"
"fmla v31.8h, v4.8h, v10.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v0.8h, v9.8h\n"
"fmla v29.8h, v0.8h, v13.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 89f\n"
"ld1 { v11.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 88f\n"
@@ -1232,7 +1232,7 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"91:" // Oddments: Load input (5, 0): Bit 2: End
"ldr x20, [x15, #0xf8]\n"
"fmla v30.8h, v0.8h, v11.8h\n"
- "add x20, x20, x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 93f\n"
"ld1 { v12.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 92f\n"
@@ -1256,11 +1256,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q1, [x16, #0x0]\n"
"ldr x20, [x15, #0x100]\n"
"fmla v31.8h, v0.8h, v12.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v1.8h, v13.8h\n"
"fmla v29.8h, v1.8h, v5.8h\n"
"fmla v30.8h, v1.8h, v12.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 97f\n"
"ld1 { v9.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 96f\n"
@@ -1284,11 +1284,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q2, [x16, #0x0]\n"
"ldr x20, [x15, #0x108]\n"
"fmla v31.8h, v1.8h, v9.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v2.8h, v5.8h\n"
"fmla v29.8h, v2.8h, v6.8h\n"
"fmla v30.8h, v2.8h, v9.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 101f\n"
"ld1 { v11.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 100f\n"
@@ -1312,11 +1312,11 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q3, [x16, #0x0]\n"
"ldr x20, [x15, #0x110]\n"
"fmla v31.8h, v2.8h, v11.8h\n"
+ "add x16, x16, #0x10\n"
"fmla v28.8h, v3.8h, v6.8h\n"
"fmla v29.8h, v3.8h, v8.8h\n"
"fmla v30.8h, v3.8h, v11.8h\n"
- "add x20, x20, x10\n"
- "add x16, x16, #0x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 105f\n"
"ld1 { v12.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 104f\n"
@@ -1343,7 +1343,7 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"fmla v28.8h, v4.8h, v8.8h\n"
"fmla v29.8h, v4.8h, v10.8h\n"
"fmla v30.8h, v4.8h, v12.8h\n"
- "add x20, x20, x10\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 109f\n"
"ld1 { v9.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 108f\n"
@@ -1368,56 +1368,56 @@ void a64_fp16_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"fmax v28.8h, v28.8h, v27.8h\n"
"fmax v29.8h, v29.8h, v27.8h\n"
"fmax v30.8h, v30.8h, v27.8h\n"
- "fmax v31.8h, v31.8h, v27.8h\n"
"fmin v28.8h, v28.8h, v15.8h\n"
+ "fmax v31.8h, v31.8h, v27.8h\n"
"fmin v29.8h, v29.8h, v15.8h\n"
"fmin v30.8h, v30.8h, v15.8h\n"
"fmin v31.8h, v31.8h, v15.8h\n"
"tbz %x[n_channels], #2, 113f\n"
- "st1 { v28.d }[0], [x14], #0x8\n"
- "st1 { v29.d }[0], [x13], #0x8\n"
- "st1 { v30.d }[0], [x12], #0x8\n"
- "st1 { v31.d }[0], [x11], #0x8\n"
+ "st1 { v28.d }[0], [x13], #0x8\n"
+ "st1 { v29.d }[0], [x12], #0x8\n"
+ "st1 { v30.d }[0], [x11], #0x8\n"
+ "st1 { v31.d }[0], [x10], #0x8\n"
"tbz %x[n_channels], #1, 112f\n"
- "st1 { v28.s }[2], [x14], #0x4\n"
- "st1 { v29.s }[2], [x13], #0x4\n"
- "st1 { v30.s }[2], [x12], #0x4\n"
- "st1 { v31.s }[2], [x11], #0x4\n"
+ "st1 { v28.s }[2], [x13], #0x4\n"
+ "st1 { v29.s }[2], [x12], #0x4\n"
+ "st1 { v30.s }[2], [x11], #0x4\n"
+ "st1 { v31.s }[2], [x10], #0x4\n"
"tbz %x[n_channels], #0, 115f\n"
- "st1 { v28.h }[6], [x14], #0x2\n"
- "st1 { v29.h }[6], [x13], #0x2\n"
- "st1 { v30.h }[6], [x12], #0x2\n"
- "st1 { v31.h }[6], [x11], #0x2\n"
+ "st1 { v28.h }[6], [x13], #0x2\n"
+ "st1 { v29.h }[6], [x12], #0x2\n"
+ "st1 { v30.h }[6], [x11], #0x2\n"
+ "st1 { v31.h }[6], [x10], #0x2\n"
"b 115f\n"
"112:" // Oddments: Store: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 115f\n"
- "st1 { v28.h }[4], [x14], #0x2\n"
- "st1 { v29.h }[4], [x13], #0x2\n"
- "st1 { v30.h }[4], [x12], #0x2\n"
- "st1 { v31.h }[4], [x11], #0x2\n"
+ "st1 { v28.h }[4], [x13], #0x2\n"
+ "st1 { v29.h }[4], [x12], #0x2\n"
+ "st1 { v30.h }[4], [x11], #0x2\n"
+ "st1 { v31.h }[4], [x10], #0x2\n"
"b 115f\n"
"113:" // Oddments: Store: Bit 2: Unset
"tbz %x[n_channels], #1, 114f\n"
- "st1 { v28.s }[0], [x14], #0x4\n"
- "st1 { v29.s }[0], [x13], #0x4\n"
- "st1 { v30.s }[0], [x12], #0x4\n"
- "st1 { v31.s }[0], [x11], #0x4\n"
+ "st1 { v28.s }[0], [x13], #0x4\n"
+ "st1 { v29.s }[0], [x12], #0x4\n"
+ "st1 { v30.s }[0], [x11], #0x4\n"
+ "st1 { v31.s }[0], [x10], #0x4\n"
"tbz %x[n_channels], #0, 115f\n"
- "st1 { v28.h }[2], [x14], #0x2\n"
- "st1 { v29.h }[2], [x13], #0x2\n"
- "st1 { v30.h }[2], [x12], #0x2\n"
- "st1 { v31.h }[2], [x11], #0x2\n"
+ "st1 { v28.h }[2], [x13], #0x2\n"
+ "st1 { v29.h }[2], [x12], #0x2\n"
+ "st1 { v30.h }[2], [x11], #0x2\n"
+ "st1 { v31.h }[2], [x10], #0x2\n"
"b 115f\n"
"114:" // Oddments: Store: Bit 2: Unset: Bit 1: Unset
- "st1 { v28.h }[0], [x14], #0x2\n"
- "st1 { v29.h }[0], [x13], #0x2\n"
- "st1 { v30.h }[0], [x12], #0x2\n"
- "st1 { v31.h }[0], [x11], #0x2\n"
+ "st1 { v28.h }[0], [x13], #0x2\n"
+ "st1 { v29.h }[0], [x12], #0x2\n"
+ "st1 { v30.h }[0], [x11], #0x2\n"
+ "st1 { v31.h }[0], [x10], #0x2\n"
"115:" // Oddments: Store: Bit 2: End
"116:" // End
:
: [n_channels] "r" ((unsigned long) n_channels), [offsetof_Args_inptrs] "I" (offsetof(Args, inptrs)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_outptrs] "I" (offsetof(Args, outptrs)), [offsetof_args_params] "I" (offsetof(Args, params)), [params_struct] "r" (&params_struct)
- : "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"
+ : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
);
}