aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp1290
1 files changed, 644 insertions, 646 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp
index 878aa29bcf..72e68482c6 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst/generic_indirect.cpp
@@ -87,405 +87,405 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
activation_min, activation_max);
__asm__ __volatile__(
- "mov x8, #0x10\n" // cntb _, ALL, #1
- "lsr x17, %x[n_channels], #0x3\n"
- "ldr x16, [%x[params_struct], %[offsetof_args_outptrs]]\n"
- "ldr x15, [%x[params_struct], %[offsetof_args_params]]\n"
+ "mov x7, #0x10\n" // cntb _, ALL, #1
+ "lsr x8, %x[n_channels], #0x3\n"
+ "ldr x17, [%x[params_struct], %[offsetof_args_outptrs]]\n"
+ "ldr x16, [%x[params_struct], %[offsetof_args_params]]\n"
"add x20, %x[params_struct], %[offsetof_args_min]\n"
- "ld1r { v18.8h }, [x20]\n"
+ "ld1r { v15.8h }, [x20]\n"
"add x20, %x[params_struct], %[offsetof_args_max]\n"
- "ld1r { v17.8h }, [x20]\n"
- "add x14, %x[params_struct], %[offsetof_Args_inptrs]\n"
- "mov x13, #0x0\n"
- "sub x12, XZR, x8\n"
- "cbz x17, 3f\n"
- "ldr q16, [x15, #0x0]\n"
- "ldr q0, [x15, #0x10]\n"
- "cmp x8, x17, LSL #4\n"
- "ldr q1, [x15, #0x20]\n"
- "ldr q2, [x15, #0x30]\n"
- "ldr q3, [x15, #0x40]\n"
- "ldr q4, [x15, #0x50]\n"
- "ldr q5, [x15, #0x60]\n"
- "ldr q6, [x15, #0x70]\n"
- "ldr q7, [x15, #0x80]\n"
- "ldr q8, [x15, #0x90]\n"
- "add x15, x15, #0xa0\n"
- "ldp x11, x10, [x14, #0x0]\n"
- "ldr q9, [x11, x13]\n"
- "ldr q10, [x10, x13]\n"
- "ldp x9, x28, [x14, #0x10]\n"
- "ldr q11, [x9, x13]\n"
- "ldr q12, [x28, x13]\n"
- "ldr x27, [x14, #0x20]\n"
- "ldr q13, [x27, x13]\n"
+ "ld1r { v14.8h }, [x20]\n"
+ "add x15, %x[params_struct], %[offsetof_Args_inptrs]\n"
+ "mov x14, #0x0\n"
+ "sub x13, XZR, x7\n"
+ "cbz x8, 3f\n"
+ "ldr q31, [x16, #0x0]\n"
+ "ldr q0, [x16, #0x10]\n"
+ "cmp x7, x8, LSL #4\n"
+ "ldr q1, [x16, #0x20]\n"
+ "ldr q2, [x16, #0x30]\n"
+ "ldr q3, [x16, #0x40]\n"
+ "ldr q4, [x16, #0x50]\n"
+ "ldr q5, [x16, #0x60]\n"
+ "ldr q6, [x16, #0x70]\n"
+ "ldr q7, [x16, #0x80]\n"
+ "ldr q8, [x16, #0x90]\n"
+ "add x16, x16, #0xa0\n"
+ "ldp x21, x20, [x15, #0x0]\n"
+ "ldr q9, [x21, x14]\n"
+ "ldr q10, [x20, x14]\n"
+ "ldp x21, x20, [x15, #0x10]\n"
+ "ldr q11, [x21, x14]\n"
+ "ldr q12, [x20, x14]\n"
+ "ldr x20, [x15, #0x20]\n"
+ "ldr q13, [x20, x14]\n"
"bge 2f\n"
"1:" // Channel loop
- "mov v23.16b, v16.16b\n fmla v23.8h, v8.8h, v9.8h\n"
- "mov v24.16b, v16.16b\n fmla v24.8h, v7.8h, v9.8h\n"
- "ldr x26, [x14, #0x30]\n"
- "ldr x25, [x14, #0x38]\n"
- "mov v25.16b, v16.16b\n fmla v25.8h, v6.8h, v9.8h\n"
- "fmla v23.8h, v0.8h, v10.8h\n"
- "ldr x24, [x14, #0x28]\n"
- "ldr x10, [x14, #0x48]\n"
- "ldr q10, [x10, x13]\n"
- "fmla v24.8h, v4.8h, v13.8h\n"
- "mov v26.16b, v16.16b\n fmla v26.8h, v5.8h, v9.8h\n"
- "ldr x11, [x14, #0x40]\n"
- "mov v27.16b, v16.16b\n fmla v27.8h, v4.8h, v9.8h\n"
- "mov v28.16b, v16.16b\n fmla v28.8h, v3.8h, v9.8h\n"
- "ldr x9, [x14, #0x50]\n"
- "ldr x28, [x14, #0x58]\n"
- "fmla v25.8h, v2.8h, v11.8h\n"
- "ldr q11, [x26, x13]\n"
- "mov v29.16b, v16.16b\n fmla v29.8h, v2.8h, v9.8h\n"
- "ldr x27, [x14, #0x60]\n"
- "fmla v23.8h, v5.8h, v13.8h\n"
- "fmla v24.8h, v6.8h, v11.8h\n"
- "ldr x26, [x14, #0x70]\n"
- "ldr x10, [x14, #0x88]\n"
- "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n"
- "fmla v25.8h, v3.8h, v13.8h\n"
- "ldr x23, [x16, #0x0]\n"
- "add x12, x12, #0x10\n"
+ "mov v29.16b, v31.16b\n fmla v29.8h, v8.8h, v9.8h\n"
+ "mov v28.16b, v31.16b\n fmla v28.8h, v7.8h, v9.8h\n"
+ "ldr x26, [x15, #0x30]\n"
+ "ldr x23, [x15, #0x38]\n"
+ "mov v27.16b, v31.16b\n fmla v27.8h, v6.8h, v9.8h\n"
+ "fmla v29.8h, v0.8h, v10.8h\n"
+ "ldr x22, [x15, #0x28]\n"
+ "ldr x20, [x15, #0x48]\n"
+ "ldr q19, [x20, x14]\n"
+ "fmla v28.8h, v4.8h, v13.8h\n"
+ "mov v26.16b, v31.16b\n fmla v26.8h, v5.8h, v9.8h\n"
+ "ldr x21, [x15, #0x40]\n"
+ "mov v25.16b, v31.16b\n fmla v25.8h, v4.8h, v9.8h\n"
+ "mov v24.16b, v31.16b\n fmla v24.8h, v3.8h, v9.8h\n"
+ "ldr x25, [x15, #0x50]\n"
+ "ldr x24, [x15, #0x58]\n"
+ "fmla v27.8h, v2.8h, v11.8h\n"
+ "ldr q17, [x26, x14]\n"
+ "mov v23.16b, v31.16b\n fmla v23.8h, v2.8h, v9.8h\n"
+ "ldr x20, [x15, #0x60]\n"
+ "fmla v29.8h, v5.8h, v13.8h\n"
+ "fmla v28.8h, v6.8h, v17.8h\n"
+ "ldr x12, [x15, #0x70]\n"
+ "ldr x11, [x15, #0x88]\n"
+ "mov v22.16b, v31.16b\n fmla v22.8h, v0.8h, v9.8h\n"
+ "fmla v27.8h, v3.8h, v13.8h\n"
+ "ldr x10, [x17, #0x0]\n"
+ "add x13, x13, #0x10\n"
"fmla v26.8h, v2.8h, v13.8h\n"
- "fmla v27.8h, v1.8h, v13.8h\n"
- "ldr x22, [x16, #0x8]\n"
- "ldr x21, [x16, #0x10]\n"
- "fmla v28.8h, v0.8h, v13.8h\n"
- "ldr q13, [x25, x13]\n"
- "fmla v29.8h, v6.8h, v12.8h\n"
- "ldr q12, [x24, x13]\n"
- "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n"
- "ldr q16, [x15, #0x0]\n"
- "fmla v23.8h, v7.8h, v11.8h\n"
- "ldr x24, [x14, #0x68]\n"
+ "fmla v25.8h, v1.8h, v13.8h\n"
+ "ldr x9, [x17, #0x8]\n"
+ "ldr x28, [x17, #0x10]\n"
"fmla v24.8h, v0.8h, v13.8h\n"
- "fmla v31.8h, v8.8h, v12.8h\n"
- "ldr q12, [x11, x13]\n"
- "ldr x25, [x14, #0x78]\n"
- "fmla v26.8h, v4.8h, v11.8h\n"
- "fmla v27.8h, v3.8h, v11.8h\n"
- "ldr x11, [x14, #0x80]\n"
- "ldr x20, [x16, #0x18]\n"
- "fmla v30.8h, v0.8h, v11.8h\n"
- "fmla v28.8h, v4.8h, v10.8h\n"
- "fmla v29.8h, v1.8h, v11.8h\n"
- "ldr q11, [x9, x13]\n"
- "fmla v23.8h, v1.8h, v13.8h\n"
- "ldr q13, [x28, x13]\n"
- "fmla v24.8h, v2.8h, v12.8h\n"
- "fmla v25.8h, v1.8h, v12.8h\n"
- "ldr q12, [x27, x13]\n"
- "ldr x9, [x14, #0x90]\n"
- "fmla v27.8h, v5.8h, v10.8h\n"
- "fmla v30.8h, v2.8h, v10.8h\n"
- "ldr x27, [x14, #0xa0]\n"
- "ldr x28, [x14, #0x98]\n"
- "fmla v26.8h, v0.8h, v11.8h\n"
- "fmla v28.8h, v2.8h, v13.8h\n"
- "fmla v24.8h, v8.8h, v10.8h\n"
- "fmla v25.8h, v7.8h, v10.8h\n"
- "fmla v31.8h, v1.8h, v10.8h\n"
- "ldr q10, [x24, x13]\n"
- "fmla v29.8h, v3.8h, v12.8h\n"
- "ldr x24, [x14, #0xa8]\n"
- "fmla v26.8h, v6.8h, v12.8h\n"
- "ldr q12, [x11, x13]\n"
- "fmla v27.8h, v7.8h, v10.8h\n"
- "ldr x11, [x14, #0xc0]\n"
- "fmla v28.8h, v6.8h, v10.8h\n"
- "fmla v30.8h, v4.8h, v10.8h\n"
- "fmla v23.8h, v3.8h, v11.8h\n"
- "ldr q11, [x26, x13]\n"
- "fmla v25.8h, v5.8h, v13.8h\n"
- "ldr q13, [x25, x13]\n"
- "fmla v29.8h, v5.8h, v10.8h\n"
- "fmla v31.8h, v3.8h, v10.8h\n"
- "ldr x26, [x14, #0xb0]\n"
- "ldr x25, [x14, #0xb8]\n"
- "fmla v26.8h, v8.8h, v10.8h\n"
- "fmla v28.8h, v8.8h, v11.8h\n"
- "fmla v30.8h, v6.8h, v13.8h\n"
- "fmla v24.8h, v3.8h, v12.8h\n"
- "fmla v27.8h, v0.8h, v12.8h\n"
- "fmla v31.8h, v5.8h, v11.8h\n"
- "ldr q11, [x10, x13]\n"
- "fmla v29.8h, v7.8h, v13.8h\n"
- "ldr q13, [x9, x13]\n"
- "fmla v23.8h, v4.8h, v12.8h\n"
- "fmla v26.8h, v1.8h, v12.8h\n"
- "ldr q12, [x28, x13]\n"
- "fmla v24.8h, v5.8h, v11.8h\n"
- "fmla v25.8h, v4.8h, v11.8h\n"
- "fmla v27.8h, v2.8h, v11.8h\n"
- "fmla v28.8h, v1.8h, v11.8h\n"
- "ldr q11, [x27, x13]\n"
- "fmla v30.8h, v8.8h, v13.8h\n"
- "ldr x27, [x14, #0x20]\n"
- "fmla v31.8h, v7.8h, v13.8h\n"
- "ldr q13, [x24, x13]\n"
- "fmla v23.8h, v2.8h, v11.8h\n"
- "fmla v26.8h, v7.8h, v12.8h\n"
- "fmla v27.8h, v6.8h, v12.8h\n"
- "fmla v29.8h, v4.8h, v12.8h\n"
- "fmla v30.8h, v3.8h, v12.8h\n"
- "ldr q12, [x26, x13]\n"
- "fmla v31.8h, v4.8h, v13.8h\n"
- "ldr q4, [x15, #0x50]\n"
- "fmla v24.8h, v1.8h, v11.8h\n"
- "ldr q1, [x15, #0x20]\n"
- "fmla v25.8h, v0.8h, v11.8h\n"
- "ldr q11, [x25, x13]\n"
+ "ldr q18, [x23, x14]\n"
"fmla v23.8h, v6.8h, v12.8h\n"
- "fmax v23.8h, v23.8h, v18.8h\n"
- "fmla v28.8h, v7.8h, v13.8h\n"
- "fmla v30.8h, v5.8h, v13.8h\n"
- "fmin v23.8h, v23.8h, v17.8h\n"
- "str q23, [x23, x12]\n"
- "fmla v29.8h, v0.8h, v12.8h\n"
- "ldr q0, [x15, #0x10]\n"
- "fmla v31.8h, v2.8h, v11.8h\n"
- "ldr q2, [x15, #0x30]\n"
- "fmla v27.8h, v8.8h, v13.8h\n"
- "ldr q13, [x11, x13]\n"
- "fmla v26.8h, v3.8h, v12.8h\n"
- "ldr q3, [x15, #0x40]\n"
- "fmla v25.8h, v8.8h, v11.8h\n"
- "fmla v28.8h, v5.8h, v11.8h\n"
- "ldr q5, [x15, #0x60]\n"
- "fmax v24.8h, v24.8h, v18.8h\n"
- "fmla v29.8h, v8.8h, v13.8h\n"
- "ldr q8, [x15, #0x90]\n"
- "fmla v30.8h, v7.8h, v13.8h\n"
- "ldr q7, [x15, #0x80]\n"
- "fmla v31.8h, v6.8h, v13.8h\n"
- "ldr q13, [x27, x8]\n"
- "ldr q6, [x15, #0x70]\n"
- "fmax v25.8h, v25.8h, v18.8h\n"
- "fmax v26.8h, v26.8h, v18.8h\n"
- "fmax v27.8h, v27.8h, v18.8h\n"
- "ldr x23, [x16, #0x20]\n"
- "ldp x11, x10, [x14, #0x0]\n"
- "ldr q9, [x11, x8]\n"
- "ldr q10, [x10, x8]\n"
- "fmin v24.8h, v24.8h, v17.8h\n"
- "fmin v25.8h, v25.8h, v17.8h\n"
- "ldp x9, x28, [x14, #0x10]\n"
- "ldr q11, [x9, x8]\n"
- "fmin v26.8h, v26.8h, v17.8h\n"
- "fmin v27.8h, v27.8h, v17.8h\n"
- "ldr q12, [x28, x8]\n"
- "fmax v28.8h, v28.8h, v18.8h\n"
- "fmax v29.8h, v29.8h, v18.8h\n"
- "str q24, [x22, x12]\n"
- "fmax v30.8h, v30.8h, v18.8h\n"
- "fmax v31.8h, v31.8h, v18.8h\n"
- "str q25, [x21, x12]\n"
- "ldr x22, [x16, #0x28]\n"
- "str q26, [x20, x12]\n"
- "ldr x21, [x16, #0x30]\n"
- "ldr x20, [x16, #0x38]\n"
- "add x8, x8, #0x10\n"
- "str q27, [x23, x12]\n"
- "ldr x23, [x16, #0x40]\n"
- "cmp x8, x17, LSL #4\n"
- "fmin v28.8h, v28.8h, v17.8h\n"
- "fmin v29.8h, v29.8h, v17.8h\n"
- "fmin v30.8h, v30.8h, v17.8h\n"
- "add x13, x13, #0x10\n"
- "str q28, [x22, x12]\n"
- "fmin v31.8h, v31.8h, v17.8h\n"
- "str q29, [x21, x12]\n"
- "add x15, x15, #0xa0\n"
- "str q30, [x20, x12]\n"
- "str q31, [x23, x12]\n"
+ "ldr q16, [x22, x14]\n"
+ "mov v21.16b, v31.16b\n fmla v21.8h, v1.8h, v9.8h\n"
+ "ldr q31, [x16, #0x0]\n"
+ "fmla v29.8h, v7.8h, v17.8h\n"
+ "ldr x23, [x15, #0x68]\n"
+ "fmla v28.8h, v0.8h, v18.8h\n"
+ "fmla v22.8h, v8.8h, v16.8h\n"
+ "ldr q16, [x21, x14]\n"
+ "ldr x22, [x15, #0x78]\n"
+ "fmla v26.8h, v4.8h, v17.8h\n"
+ "fmla v25.8h, v3.8h, v17.8h\n"
+ "ldr x21, [x15, #0x80]\n"
+ "ldr x27, [x17, #0x18]\n"
+ "fmla v21.8h, v0.8h, v17.8h\n"
+ "fmla v24.8h, v4.8h, v19.8h\n"
+ "fmla v23.8h, v1.8h, v17.8h\n"
+ "ldr q17, [x25, x14]\n"
+ "fmla v29.8h, v1.8h, v18.8h\n"
+ "ldr q20, [x24, x14]\n"
+ "fmla v28.8h, v2.8h, v16.8h\n"
+ "fmla v27.8h, v1.8h, v16.8h\n"
+ "ldr q16, [x20, x14]\n"
+ "ldr x26, [x15, #0x90]\n"
+ "fmla v25.8h, v5.8h, v19.8h\n"
+ "fmla v21.8h, v2.8h, v19.8h\n"
+ "ldr x25, [x15, #0xa0]\n"
+ "ldr x20, [x15, #0x98]\n"
+ "fmla v26.8h, v0.8h, v17.8h\n"
+ "fmla v24.8h, v2.8h, v20.8h\n"
+ "fmla v28.8h, v8.8h, v19.8h\n"
+ "fmla v27.8h, v7.8h, v19.8h\n"
+ "fmla v22.8h, v1.8h, v19.8h\n"
+ "ldr q19, [x23, x14]\n"
+ "fmla v23.8h, v3.8h, v16.8h\n"
+ "ldr x24, [x15, #0xa8]\n"
+ "fmla v26.8h, v6.8h, v16.8h\n"
+ "ldr q18, [x21, x14]\n"
+ "fmla v25.8h, v7.8h, v19.8h\n"
+ "ldr x23, [x15, #0xc0]\n"
+ "fmla v24.8h, v6.8h, v19.8h\n"
+ "fmla v21.8h, v4.8h, v19.8h\n"
+ "fmla v29.8h, v3.8h, v17.8h\n"
+ "ldr q17, [x12, x14]\n"
+ "fmla v27.8h, v5.8h, v20.8h\n"
+ "ldr q16, [x22, x14]\n"
+ "fmla v23.8h, v5.8h, v19.8h\n"
+ "fmla v22.8h, v3.8h, v19.8h\n"
+ "ldr x22, [x15, #0xb0]\n"
+ "ldr x21, [x15, #0xb8]\n"
+ "fmla v26.8h, v8.8h, v19.8h\n"
+ "fmla v24.8h, v8.8h, v17.8h\n"
+ "fmla v21.8h, v6.8h, v16.8h\n"
+ "fmla v28.8h, v3.8h, v18.8h\n"
+ "fmla v25.8h, v0.8h, v18.8h\n"
+ "fmla v22.8h, v5.8h, v17.8h\n"
+ "ldr q17, [x11, x14]\n"
+ "fmla v23.8h, v7.8h, v16.8h\n"
+ "ldr q16, [x26, x14]\n"
+ "fmla v29.8h, v4.8h, v18.8h\n"
+ "fmla v26.8h, v1.8h, v18.8h\n"
+ "ldr q18, [x20, x14]\n"
+ "fmla v28.8h, v5.8h, v17.8h\n"
+ "fmla v27.8h, v4.8h, v17.8h\n"
+ "fmla v25.8h, v2.8h, v17.8h\n"
+ "fmla v24.8h, v1.8h, v17.8h\n"
+ "ldr q17, [x25, x14]\n"
+ "fmla v21.8h, v8.8h, v16.8h\n"
+ "ldr x20, [x15, #0x20]\n"
+ "fmla v22.8h, v7.8h, v16.8h\n"
+ "ldr q16, [x24, x14]\n"
+ "fmla v29.8h, v2.8h, v17.8h\n"
+ "fmla v26.8h, v7.8h, v18.8h\n"
+ "fmla v25.8h, v6.8h, v18.8h\n"
+ "fmla v23.8h, v4.8h, v18.8h\n"
+ "fmla v21.8h, v3.8h, v18.8h\n"
+ "ldr q18, [x22, x14]\n"
+ "fmla v22.8h, v4.8h, v16.8h\n"
+ "ldr q4, [x16, #0x50]\n"
+ "fmla v28.8h, v1.8h, v17.8h\n"
+ "ldr q1, [x16, #0x20]\n"
+ "fmla v27.8h, v0.8h, v17.8h\n"
+ "ldr q17, [x21, x14]\n"
+ "fmla v29.8h, v6.8h, v18.8h\n"
+ "fmax v29.8h, v29.8h, v15.8h\n"
+ "fmla v24.8h, v7.8h, v16.8h\n"
+ "fmla v21.8h, v5.8h, v16.8h\n"
+ "fmin v29.8h, v29.8h, v14.8h\n"
+ "str q29, [x10, x13]\n"
+ "fmla v23.8h, v0.8h, v18.8h\n"
+ "ldr q0, [x16, #0x10]\n"
+ "fmla v22.8h, v2.8h, v17.8h\n"
+ "ldr q2, [x16, #0x30]\n"
+ "fmla v25.8h, v8.8h, v16.8h\n"
+ "ldr q16, [x23, x14]\n"
+ "fmla v26.8h, v3.8h, v18.8h\n"
+ "ldr q3, [x16, #0x40]\n"
+ "fmla v27.8h, v8.8h, v17.8h\n"
+ "fmla v24.8h, v5.8h, v17.8h\n"
+ "ldr q5, [x16, #0x60]\n"
+ "fmax v28.8h, v28.8h, v15.8h\n"
+ "fmla v23.8h, v8.8h, v16.8h\n"
+ "ldr q8, [x16, #0x90]\n"
+ "fmla v21.8h, v7.8h, v16.8h\n"
+ "ldr q7, [x16, #0x80]\n"
+ "fmla v22.8h, v6.8h, v16.8h\n"
+ "ldr q13, [x20, x7]\n"
+ "ldr q6, [x16, #0x70]\n"
+ "fmax v27.8h, v27.8h, v15.8h\n"
+ "fmax v26.8h, v26.8h, v15.8h\n"
+ "fmax v25.8h, v25.8h, v15.8h\n"
+ "ldr x24, [x17, #0x20]\n"
+ "ldp x21, x20, [x15, #0x0]\n"
+ "ldr q9, [x21, x7]\n"
+ "ldr q10, [x20, x7]\n"
+ "fmin v28.8h, v28.8h, v14.8h\n"
+ "fmin v27.8h, v27.8h, v14.8h\n"
+ "ldp x21, x20, [x15, #0x10]\n"
+ "ldr q11, [x21, x7]\n"
+ "fmin v26.8h, v26.8h, v14.8h\n"
+ "fmin v25.8h, v25.8h, v14.8h\n"
+ "ldr q12, [x20, x7]\n"
+ "fmax v24.8h, v24.8h, v15.8h\n"
+ "fmax v23.8h, v23.8h, v15.8h\n"
+ "str q28, [x9, x13]\n"
+ "fmax v21.8h, v21.8h, v15.8h\n"
+ "fmax v22.8h, v22.8h, v15.8h\n"
+ "str q27, [x28, x13]\n"
+ "ldr x23, [x17, #0x28]\n"
+ "str q26, [x27, x13]\n"
+ "ldr x22, [x17, #0x30]\n"
+ "ldr x21, [x17, #0x38]\n"
+ "add x7, x7, #0x10\n"
+ "str q25, [x24, x13]\n"
+ "ldr x20, [x17, #0x40]\n"
+ "cmp x7, x8, LSL #4\n"
+ "fmin v24.8h, v24.8h, v14.8h\n"
+ "fmin v23.8h, v23.8h, v14.8h\n"
+ "fmin v21.8h, v21.8h, v14.8h\n"
+ "add x14, x14, #0x10\n"
+ "str q24, [x23, x13]\n"
+ "fmin v22.8h, v22.8h, v14.8h\n"
+ "str q23, [x22, x13]\n"
+ "add x16, x16, #0xa0\n"
+ "str q21, [x21, x13]\n"
+ "str q22, [x20, x13]\n"
"blt 1b\n"
"2:" // Channel tail
- "mov v23.16b, v16.16b\n fmla v23.8h, v8.8h, v9.8h\n"
- "mov v24.16b, v16.16b\n fmla v24.8h, v7.8h, v9.8h\n"
- "ldr x26, [x14, #0x30]\n"
- "ldr x25, [x14, #0x38]\n"
- "mov v25.16b, v16.16b\n fmla v25.8h, v6.8h, v9.8h\n"
- "fmla v23.8h, v0.8h, v10.8h\n"
- "ldr x24, [x14, #0x28]\n"
- "ldr x10, [x14, #0x48]\n"
- "ldr q10, [x10, x13]\n"
- "fmla v24.8h, v4.8h, v13.8h\n"
- "mov v26.16b, v16.16b\n fmla v26.8h, v5.8h, v9.8h\n"
- "ldr x11, [x14, #0x40]\n"
- "mov v27.16b, v16.16b\n fmla v27.8h, v4.8h, v9.8h\n"
- "mov v28.16b, v16.16b\n fmla v28.8h, v3.8h, v9.8h\n"
- "ldr x9, [x14, #0x50]\n"
- "ldr x28, [x14, #0x58]\n"
- "fmla v25.8h, v2.8h, v11.8h\n"
- "ldr q11, [x26, x13]\n"
- "mov v29.16b, v16.16b\n fmla v29.8h, v2.8h, v9.8h\n"
- "ldr x27, [x14, #0x60]\n"
- "fmla v23.8h, v5.8h, v13.8h\n"
- "fmla v24.8h, v6.8h, v11.8h\n"
- "ldr x26, [x14, #0x70]\n"
- "ldr x10, [x14, #0x88]\n"
- "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n"
- "fmla v25.8h, v3.8h, v13.8h\n"
- "ldr x23, [x16, #0x0]\n"
- "add x12, x12, #0x10\n"
+ "mov v29.16b, v31.16b\n fmla v29.8h, v8.8h, v9.8h\n"
+ "mov v28.16b, v31.16b\n fmla v28.8h, v7.8h, v9.8h\n"
+ "ldr x23, [x15, #0x30]\n"
+ "ldr x22, [x15, #0x38]\n"
+ "mov v27.16b, v31.16b\n fmla v27.8h, v6.8h, v9.8h\n"
+ "fmla v29.8h, v0.8h, v10.8h\n"
+ "ldr x21, [x15, #0x28]\n"
+ "ldr x20, [x15, #0x48]\n"
+ "ldr q19, [x20, x14]\n"
+ "fmla v28.8h, v4.8h, v13.8h\n"
+ "mov v26.16b, v31.16b\n fmla v26.8h, v5.8h, v9.8h\n"
+ "ldr x20, [x15, #0x40]\n"
+ "mov v25.16b, v31.16b\n fmla v25.8h, v4.8h, v9.8h\n"
+ "mov v24.16b, v31.16b\n fmla v24.8h, v3.8h, v9.8h\n"
+ "ldr x25, [x15, #0x50]\n"
+ "ldr x24, [x15, #0x58]\n"
+ "fmla v27.8h, v2.8h, v11.8h\n"
+ "ldr q17, [x23, x14]\n"
+ "mov v23.16b, v31.16b\n fmla v23.8h, v2.8h, v9.8h\n"
+ "ldr x23, [x15, #0x60]\n"
+ "fmla v29.8h, v5.8h, v13.8h\n"
+ "fmla v28.8h, v6.8h, v17.8h\n"
+ "ldr x12, [x15, #0x70]\n"
+ "ldr x11, [x15, #0x88]\n"
+ "mov v22.16b, v31.16b\n fmla v22.8h, v0.8h, v9.8h\n"
+ "fmla v27.8h, v3.8h, v13.8h\n"
+ "ldr x10, [x17, #0x0]\n"
+ "add x13, x13, #0x10\n"
"fmla v26.8h, v2.8h, v13.8h\n"
- "fmla v27.8h, v1.8h, v13.8h\n"
- "ldr x22, [x16, #0x8]\n"
- "ldr x21, [x16, #0x10]\n"
- "fmla v28.8h, v0.8h, v13.8h\n"
- "ldr q13, [x25, x13]\n"
- "fmla v29.8h, v6.8h, v12.8h\n"
- "ldr q12, [x24, x13]\n"
- "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n"
- "fmla v23.8h, v7.8h, v11.8h\n"
- "ldr x24, [x14, #0x68]\n"
- "ldr x25, [x14, #0x78]\n"
+ "fmla v25.8h, v1.8h, v13.8h\n"
+ "ldr x9, [x17, #0x8]\n"
+ "ldr x28, [x17, #0x10]\n"
"fmla v24.8h, v0.8h, v13.8h\n"
- "fmla v31.8h, v8.8h, v12.8h\n"
- "ldr q12, [x11, x13]\n"
- "ldr x11, [x14, #0x80]\n"
- "fmla v26.8h, v4.8h, v11.8h\n"
- "fmla v27.8h, v3.8h, v11.8h\n"
- "ldr x20, [x16, #0x18]\n"
- "fmla v30.8h, v0.8h, v11.8h\n"
- "fmla v28.8h, v4.8h, v10.8h\n"
- "fmla v29.8h, v1.8h, v11.8h\n"
- "ldr q11, [x9, x13]\n"
- "fmla v23.8h, v1.8h, v13.8h\n"
- "ldr q13, [x28, x13]\n"
- "fmla v24.8h, v2.8h, v12.8h\n"
- "fmla v25.8h, v1.8h, v12.8h\n"
- "ldr q12, [x27, x13]\n"
- "ldr x9, [x14, #0x90]\n"
- "fmla v27.8h, v5.8h, v10.8h\n"
- "fmla v30.8h, v2.8h, v10.8h\n"
- "ldr x27, [x14, #0xa0]\n"
- "ldr x28, [x14, #0x98]\n"
- "fmla v26.8h, v0.8h, v11.8h\n"
- "fmla v28.8h, v2.8h, v13.8h\n"
- "fmla v24.8h, v8.8h, v10.8h\n"
- "fmla v25.8h, v7.8h, v10.8h\n"
- "fmla v31.8h, v1.8h, v10.8h\n"
- "ldr q10, [x24, x13]\n"
- "fmla v29.8h, v3.8h, v12.8h\n"
- "ldr x24, [x14, #0xa8]\n"
- "fmla v26.8h, v6.8h, v12.8h\n"
- "ldr q12, [x11, x13]\n"
- "fmla v27.8h, v7.8h, v10.8h\n"
- "ldr x11, [x14, #0xc0]\n"
- "fmla v28.8h, v6.8h, v10.8h\n"
- "fmla v30.8h, v4.8h, v10.8h\n"
- "fmla v23.8h, v3.8h, v11.8h\n"
- "ldr q11, [x26, x13]\n"
- "fmla v25.8h, v5.8h, v13.8h\n"
- "ldr q13, [x25, x13]\n"
- "fmla v29.8h, v5.8h, v10.8h\n"
- "fmla v31.8h, v3.8h, v10.8h\n"
- "ldr x26, [x14, #0xb0]\n"
- "ldr x25, [x14, #0xb8]\n"
- "fmla v26.8h, v8.8h, v10.8h\n"
- "fmla v28.8h, v8.8h, v11.8h\n"
- "fmla v30.8h, v6.8h, v13.8h\n"
- "fmla v24.8h, v3.8h, v12.8h\n"
- "fmla v27.8h, v0.8h, v12.8h\n"
- "fmla v31.8h, v5.8h, v11.8h\n"
- "ldr q11, [x10, x13]\n"
- "fmla v29.8h, v7.8h, v13.8h\n"
- "ldr q13, [x9, x13]\n"
- "fmla v23.8h, v4.8h, v12.8h\n"
- "fmla v26.8h, v1.8h, v12.8h\n"
- "ldr q12, [x28, x13]\n"
- "fmla v24.8h, v5.8h, v11.8h\n"
- "fmla v25.8h, v4.8h, v11.8h\n"
- "fmla v27.8h, v2.8h, v11.8h\n"
- "fmla v28.8h, v1.8h, v11.8h\n"
- "ldr q11, [x27, x13]\n"
- "fmla v30.8h, v8.8h, v13.8h\n"
- "fmla v31.8h, v7.8h, v13.8h\n"
- "ldr q13, [x24, x13]\n"
- "fmla v23.8h, v2.8h, v11.8h\n"
- "fmla v26.8h, v7.8h, v12.8h\n"
- "fmla v27.8h, v6.8h, v12.8h\n"
- "fmla v29.8h, v4.8h, v12.8h\n"
- "fmla v30.8h, v3.8h, v12.8h\n"
- "ldr q12, [x26, x13]\n"
- "fmla v31.8h, v4.8h, v13.8h\n"
- "fmla v24.8h, v1.8h, v11.8h\n"
- "fmax v24.8h, v24.8h, v18.8h\n"
- "fmla v25.8h, v0.8h, v11.8h\n"
- "ldr q11, [x25, x13]\n"
+ "ldr q18, [x22, x14]\n"
"fmla v23.8h, v6.8h, v12.8h\n"
- "fmax v23.8h, v23.8h, v18.8h\n"
- "fmla v28.8h, v7.8h, v13.8h\n"
- "fmla v30.8h, v5.8h, v13.8h\n"
- "fmin v23.8h, v23.8h, v17.8h\n"
- "str q23, [x23, x12]\n"
- "fmla v29.8h, v0.8h, v12.8h\n"
- "fmla v31.8h, v2.8h, v11.8h\n"
- "ldr x23, [x16, #0x20]\n"
- "fmin v24.8h, v24.8h, v17.8h\n"
- "fmla v27.8h, v8.8h, v13.8h\n"
- "ldr q13, [x11, x13]\n"
- "fmla v26.8h, v3.8h, v12.8h\n"
- "fmax v26.8h, v26.8h, v18.8h\n"
- "fmla v25.8h, v8.8h, v11.8h\n"
- "fmla v28.8h, v5.8h, v11.8h\n"
- "fmax v25.8h, v25.8h, v18.8h\n"
- "str q24, [x22, x12]\n"
- "fmla v29.8h, v8.8h, v13.8h\n"
- "fmla v30.8h, v7.8h, v13.8h\n"
- "fmax v27.8h, v27.8h, v18.8h\n"
- "ldr x22, [x16, #0x28]\n"
- "fmla v31.8h, v6.8h, v13.8h\n"
- "fmin v25.8h, v25.8h, v17.8h\n"
- "str q25, [x21, x12]\n"
- "ldr x21, [x16, #0x30]\n"
- "fmin v26.8h, v26.8h, v17.8h\n"
- "fmin v27.8h, v27.8h, v17.8h\n"
- "str q26, [x20, x12]\n"
- "ldr x20, [x16, #0x38]\n"
- "fmax v28.8h, v28.8h, v18.8h\n"
- "fmax v29.8h, v29.8h, v18.8h\n"
- "str q27, [x23, x12]\n"
- "ldr x23, [x16, #0x40]\n"
- "fmax v30.8h, v30.8h, v18.8h\n"
- "fmax v31.8h, v31.8h, v18.8h\n"
- "add x13, x13, #0x10\n"
- "fmin v28.8h, v28.8h, v17.8h\n"
- "fmin v29.8h, v29.8h, v17.8h\n"
- "str q28, [x22, x12]\n"
- "fmin v30.8h, v30.8h, v17.8h\n"
- "fmin v31.8h, v31.8h, v17.8h\n"
- "str q29, [x21, x12]\n"
- "str q30, [x20, x12]\n"
- "str q31, [x23, x12]\n"
+ "ldr q16, [x21, x14]\n"
+ "mov v21.16b, v31.16b\n fmla v21.8h, v1.8h, v9.8h\n"
+ "fmla v29.8h, v7.8h, v17.8h\n"
+ "ldr x22, [x15, #0x68]\n"
+ "ldr x21, [x15, #0x78]\n"
+ "fmla v28.8h, v0.8h, v18.8h\n"
+ "fmla v22.8h, v8.8h, v16.8h\n"
+ "ldr q16, [x20, x14]\n"
+ "ldr x20, [x15, #0x80]\n"
+ "fmla v26.8h, v4.8h, v17.8h\n"
+ "fmla v25.8h, v3.8h, v17.8h\n"
+ "ldr x27, [x17, #0x18]\n"
+ "fmla v21.8h, v0.8h, v17.8h\n"
+ "fmla v24.8h, v4.8h, v19.8h\n"
+ "fmla v23.8h, v1.8h, v17.8h\n"
+ "ldr q17, [x25, x14]\n"
+ "fmla v29.8h, v1.8h, v18.8h\n"
+ "ldr q20, [x24, x14]\n"
+ "fmla v28.8h, v2.8h, v16.8h\n"
+ "fmla v27.8h, v1.8h, v16.8h\n"
+ "ldr q16, [x23, x14]\n"
+ "ldr x26, [x15, #0x90]\n"
+ "fmla v25.8h, v5.8h, v19.8h\n"
+ "fmla v21.8h, v2.8h, v19.8h\n"
+ "ldr x25, [x15, #0xa0]\n"
+ "ldr x24, [x15, #0x98]\n"
+ "fmla v26.8h, v0.8h, v17.8h\n"
+ "fmla v24.8h, v2.8h, v20.8h\n"
+ "fmla v28.8h, v8.8h, v19.8h\n"
+ "fmla v27.8h, v7.8h, v19.8h\n"
+ "fmla v22.8h, v1.8h, v19.8h\n"
+ "ldr q19, [x22, x14]\n"
+ "fmla v23.8h, v3.8h, v16.8h\n"
+ "ldr x23, [x15, #0xa8]\n"
+ "fmla v26.8h, v6.8h, v16.8h\n"
+ "ldr q18, [x20, x14]\n"
+ "fmla v25.8h, v7.8h, v19.8h\n"
+ "ldr x22, [x15, #0xc0]\n"
+ "fmla v24.8h, v6.8h, v19.8h\n"
+ "fmla v21.8h, v4.8h, v19.8h\n"
+ "fmla v29.8h, v3.8h, v17.8h\n"
+ "ldr q17, [x12, x14]\n"
+ "fmla v27.8h, v5.8h, v20.8h\n"
+ "ldr q16, [x21, x14]\n"
+ "fmla v23.8h, v5.8h, v19.8h\n"
+ "fmla v22.8h, v3.8h, v19.8h\n"
+ "ldr x21, [x15, #0xb0]\n"
+ "ldr x20, [x15, #0xb8]\n"
+ "fmla v26.8h, v8.8h, v19.8h\n"
+ "fmla v24.8h, v8.8h, v17.8h\n"
+ "fmla v21.8h, v6.8h, v16.8h\n"
+ "fmla v28.8h, v3.8h, v18.8h\n"
+ "fmla v25.8h, v0.8h, v18.8h\n"
+ "fmla v22.8h, v5.8h, v17.8h\n"
+ "ldr q17, [x11, x14]\n"
+ "fmla v23.8h, v7.8h, v16.8h\n"
+ "ldr q16, [x26, x14]\n"
+ "fmla v29.8h, v4.8h, v18.8h\n"
+ "fmla v26.8h, v1.8h, v18.8h\n"
+ "ldr q18, [x24, x14]\n"
+ "fmla v28.8h, v5.8h, v17.8h\n"
+ "fmla v27.8h, v4.8h, v17.8h\n"
+ "fmla v25.8h, v2.8h, v17.8h\n"
+ "fmla v24.8h, v1.8h, v17.8h\n"
+ "ldr q17, [x25, x14]\n"
+ "fmla v21.8h, v8.8h, v16.8h\n"
+ "fmla v22.8h, v7.8h, v16.8h\n"
+ "ldr q16, [x23, x14]\n"
+ "fmla v29.8h, v2.8h, v17.8h\n"
+ "fmla v26.8h, v7.8h, v18.8h\n"
+ "fmla v25.8h, v6.8h, v18.8h\n"
+ "fmla v23.8h, v4.8h, v18.8h\n"
+ "fmla v21.8h, v3.8h, v18.8h\n"
+ "ldr q18, [x21, x14]\n"
+ "fmla v22.8h, v4.8h, v16.8h\n"
+ "fmla v28.8h, v1.8h, v17.8h\n"
+ "fmax v28.8h, v28.8h, v15.8h\n"
+ "fmla v27.8h, v0.8h, v17.8h\n"
+ "ldr q17, [x20, x14]\n"
+ "fmla v29.8h, v6.8h, v18.8h\n"
+ "fmax v29.8h, v29.8h, v15.8h\n"
+ "fmla v24.8h, v7.8h, v16.8h\n"
+ "fmla v21.8h, v5.8h, v16.8h\n"
+ "fmin v29.8h, v29.8h, v14.8h\n"
+ "str q29, [x10, x13]\n"
+ "fmla v23.8h, v0.8h, v18.8h\n"
+ "fmla v22.8h, v2.8h, v17.8h\n"
+ "ldr x20, [x17, #0x20]\n"
+ "fmin v28.8h, v28.8h, v14.8h\n"
+ "fmla v25.8h, v8.8h, v16.8h\n"
+ "ldr q16, [x22, x14]\n"
+ "fmla v26.8h, v3.8h, v18.8h\n"
+ "fmax v26.8h, v26.8h, v15.8h\n"
+ "fmla v27.8h, v8.8h, v17.8h\n"
+ "fmla v24.8h, v5.8h, v17.8h\n"
+ "fmax v27.8h, v27.8h, v15.8h\n"
+ "str q28, [x9, x13]\n"
+ "fmla v23.8h, v8.8h, v16.8h\n"
+ "fmla v21.8h, v7.8h, v16.8h\n"
+ "fmax v25.8h, v25.8h, v15.8h\n"
+ "ldr x23, [x17, #0x28]\n"
+ "fmla v22.8h, v6.8h, v16.8h\n"
+ "fmin v27.8h, v27.8h, v14.8h\n"
+ "str q27, [x28, x13]\n"
+ "ldr x22, [x17, #0x30]\n"
+ "fmin v26.8h, v26.8h, v14.8h\n"
+ "fmin v25.8h, v25.8h, v14.8h\n"
+ "str q26, [x27, x13]\n"
+ "ldr x21, [x17, #0x38]\n"
+ "fmax v24.8h, v24.8h, v15.8h\n"
+ "fmax v23.8h, v23.8h, v15.8h\n"
+ "str q25, [x20, x13]\n"
+ "ldr x20, [x17, #0x40]\n"
+ "fmax v21.8h, v21.8h, v15.8h\n"
+ "fmax v22.8h, v22.8h, v15.8h\n"
+ "add x14, x14, #0x10\n"
+ "fmin v24.8h, v24.8h, v14.8h\n"
+ "fmin v23.8h, v23.8h, v14.8h\n"
+ "str q24, [x23, x13]\n"
+ "fmin v21.8h, v21.8h, v14.8h\n"
+ "fmin v22.8h, v22.8h, v14.8h\n"
+ "str q23, [x22, x13]\n"
+ "str q21, [x21, x13]\n"
+ "str q22, [x20, x13]\n"
"3:" // Oddments
"tst %x[n_channels], #0x7\n"
"beq 92f\n"
- "ldr q16, [x15, #0x0]\n"
- "ldr q0, [x15, #0x10]\n"
- "mov x12, x13\n"
- "ldr q1, [x15, #0x20]\n"
- "ldr q2, [x15, #0x30]\n"
- "ldr q3, [x15, #0x40]\n"
- "ldr q4, [x15, #0x50]\n"
- "ldr q5, [x15, #0x60]\n"
- "ldr q6, [x15, #0x70]\n"
- "ldr q7, [x15, #0x80]\n"
- "ldr q8, [x15, #0x90]\n"
- "ldr x24, [x14, #0x0]\n"
- "ldr x23, [x14, #0x8]\n"
- "add x24, x24, x13\n"
- "add x23, x23, x13\n"
- "ldr x22, [x14, #0x10]\n"
- "ldr x21, [x14, #0x18]\n"
- "add x22, x22, x13\n"
- "add x21, x21, x13\n"
- "ldr x20, [x14, #0x20]\n"
- "add x20, x20, x13\n"
+ "ldr q31, [x16, #0x0]\n"
+ "ldr q0, [x16, #0x10]\n"
+ "mov x13, x14\n"
+ "ldr q1, [x16, #0x20]\n"
+ "ldr q2, [x16, #0x30]\n"
+ "ldr q3, [x16, #0x40]\n"
+ "ldr q4, [x16, #0x50]\n"
+ "ldr q5, [x16, #0x60]\n"
+ "ldr q6, [x16, #0x70]\n"
+ "ldr q7, [x16, #0x80]\n"
+ "ldr q8, [x16, #0x90]\n"
+ "ldr x24, [x15, #0x0]\n"
+ "ldr x23, [x15, #0x8]\n"
+ "add x24, x24, x14\n"
+ "add x23, x23, x14\n"
+ "ldr x22, [x15, #0x10]\n"
+ "ldr x21, [x15, #0x18]\n"
+ "add x22, x22, x14\n"
+ "add x21, x21, x14\n"
+ "ldr x20, [x15, #0x20]\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 5f\n"
"ld1 { v9.d }[0], [x24], #0x8\n"
"ld1 { v10.d }[0], [x23], #0x8\n"
@@ -534,19 +534,19 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"ld1 { v12.h }[0], [x21], #0x2\n"
"ld1 { v13.h }[0], [x20], #0x2\n"
"7:" // Oddments: Load inputs (2, 2), (0, 0), (0, 4), (4, 0), (1, 2): Bit 2: End
- "mov v23.16b, v16.16b\n fmla v23.8h, v8.8h, v9.8h\n"
- "mov v25.16b, v16.16b\n fmla v25.8h, v6.8h, v9.8h\n"
- "ldr x20, [x14, #0x28]\n"
- "add x20, x20, x13\n"
- "mov v24.16b, v16.16b\n fmla v24.8h, v7.8h, v9.8h\n"
- "mov v26.16b, v16.16b\n fmla v26.8h, v5.8h, v9.8h\n"
- "mov v27.16b, v16.16b\n fmla v27.8h, v4.8h, v9.8h\n"
- "mov v28.16b, v16.16b\n fmla v28.8h, v3.8h, v9.8h\n"
- "mov v29.16b, v16.16b\n fmla v29.8h, v2.8h, v9.8h\n"
+ "mov v23.16b, v31.16b\n fmla v23.8h, v8.8h, v9.8h\n"
+ "mov v25.16b, v31.16b\n fmla v25.8h, v6.8h, v9.8h\n"
+ "ldr x20, [x15, #0x28]\n"
+ "add x20, x20, x14\n"
+ "mov v24.16b, v31.16b\n fmla v24.8h, v7.8h, v9.8h\n"
+ "mov v26.16b, v31.16b\n fmla v26.8h, v5.8h, v9.8h\n"
+ "mov v27.16b, v31.16b\n fmla v27.8h, v4.8h, v9.8h\n"
+ "mov v28.16b, v31.16b\n fmla v28.8h, v3.8h, v9.8h\n"
+ "mov v29.16b, v31.16b\n fmla v29.8h, v2.8h, v9.8h\n"
"fmla v23.8h, v0.8h, v10.8h\n"
"fmla v25.8h, v2.8h, v11.8h\n"
- "mov v30.16b, v16.16b\n fmla v30.8h, v1.8h, v9.8h\n"
- "mov v31.16b, v16.16b\n fmla v31.8h, v0.8h, v9.8h\n"
+ "mov v30.16b, v31.16b\n fmla v30.8h, v1.8h, v9.8h\n"
+ "fmla v31.8h, v0.8h, v9.8h\n"
"fmla v29.8h, v6.8h, v12.8h\n"
"fmla v23.8h, v5.8h, v13.8h\n"
"fmla v24.8h, v4.8h, v13.8h\n"
@@ -574,9 +574,9 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"10:" // Oddments: Load input (4, 4): Bit 2: Unset: Bit 1: Unset
"ld1 { v12.h }[0], [x20], #0x2\n"
"11:" // Oddments: Load input (4, 4): Bit 2: End
- "ldr x20, [x14, #0x30]\n"
+ "ldr x20, [x15, #0x30]\n"
"fmla v31.8h, v8.8h, v12.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 13f\n"
"ld1 { v11.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 12f\n"
@@ -597,10 +597,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"14:" // Oddments: Load input (2, 1): Bit 2: Unset: Bit 1: Unset
"ld1 { v11.h }[0], [x20], #0x2\n"
"15:" // Oddments: Load input (2, 1): Bit 2: End
- "ldr x20, [x14, #0x38]\n"
+ "ldr x20, [x15, #0x38]\n"
"fmla v23.8h, v7.8h, v11.8h\n"
"fmla v24.8h, v6.8h, v11.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"fmla v26.8h, v4.8h, v11.8h\n"
"fmla v27.8h, v3.8h, v11.8h\n"
"fmla v29.8h, v1.8h, v11.8h\n"
@@ -625,10 +625,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"18:" // Oddments: Load input (0, 1): Bit 2: Unset: Bit 1: Unset
"ld1 { v13.h }[0], [x20], #0x2\n"
"19:" // Oddments: Load input (0, 1): Bit 2: End
- "ldr x20, [x14, #0x40]\n"
+ "ldr x20, [x15, #0x40]\n"
"fmla v23.8h, v1.8h, v13.8h\n"
"fmla v24.8h, v0.8h, v13.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 21f\n"
"ld1 { v12.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 20f\n"
@@ -649,10 +649,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"22:" // Oddments: Load input (0, 3): Bit 2: Unset: Bit 1: Unset
"ld1 { v12.h }[0], [x20], #0x2\n"
"23:" // Oddments: Load input (0, 3): Bit 2: End
- "ldr x20, [x14, #0x48]\n"
+ "ldr x20, [x15, #0x48]\n"
"fmla v24.8h, v2.8h, v12.8h\n"
"fmla v25.8h, v1.8h, v12.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 25f\n"
"ld1 { v10.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 24f\n"
@@ -673,10 +673,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"26:" // Oddments: Load input (2, 3): Bit 2: Unset: Bit 1: Unset
"ld1 { v10.h }[0], [x20], #0x2\n"
"27:" // Oddments: Load input (2, 3): Bit 2: End
- "ldr x20, [x14, #0x50]\n"
+ "ldr x20, [x15, #0x50]\n"
"fmla v24.8h, v8.8h, v10.8h\n"
"fmla v25.8h, v7.8h, v10.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"fmla v27.8h, v5.8h, v10.8h\n"
"fmla v28.8h, v4.8h, v10.8h\n"
"fmla v30.8h, v2.8h, v10.8h\n"
@@ -701,10 +701,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"30:" // Oddments: Load input (1, 0): Bit 2: Unset: Bit 1: Unset
"ld1 { v11.h }[0], [x20], #0x2\n"
"31:" // Oddments: Load input (1, 0): Bit 2: End
- "ldr x20, [x14, #0x58]\n"
+ "ldr x20, [x15, #0x58]\n"
"fmla v23.8h, v3.8h, v11.8h\n"
"fmla v26.8h, v0.8h, v11.8h\n"
- "add x20, x20, x13\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"
@@ -725,10 +725,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"34:" // Oddments: Load input (1, 4): Bit 2: Unset: Bit 1: Unset
"ld1 { v13.h }[0], [x20], #0x2\n"
"35:" // Oddments: Load input (1, 4): Bit 2: End
- "ldr x20, [x14, #0x60]\n"
+ "ldr x20, [x15, #0x60]\n"
"fmla v25.8h, v5.8h, v13.8h\n"
"fmla v28.8h, v2.8h, v13.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 37f\n"
"ld1 { v12.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 36f\n"
@@ -749,10 +749,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"38:" // Oddments: Load input (3, 0): Bit 2: Unset: Bit 1: Unset
"ld1 { v12.h }[0], [x20], #0x2\n"
"39:" // Oddments: Load input (3, 0): Bit 2: End
- "ldr x20, [x14, #0x68]\n"
+ "ldr x20, [x15, #0x68]\n"
"fmla v26.8h, v6.8h, v12.8h\n"
"fmla v29.8h, v3.8h, v12.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 41f\n"
"ld1 { v10.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 40f\n"
@@ -773,10 +773,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"42:" // Oddments: Load input (3, 2): Bit 2: Unset: Bit 1: Unset
"ld1 { v10.h }[0], [x20], #0x2\n"
"43:" // Oddments: Load input (3, 2): Bit 2: End
- "ldr x20, [x14, #0x70]\n"
+ "ldr x20, [x15, #0x70]\n"
"fmla v26.8h, v8.8h, v10.8h\n"
"fmla v27.8h, v7.8h, v10.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"fmla v28.8h, v6.8h, v10.8h\n"
"fmla v29.8h, v5.8h, v10.8h\n"
"fmla v30.8h, v4.8h, v10.8h\n"
@@ -801,10 +801,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"46:" // Oddments: Load input (3, 4): Bit 2: Unset: Bit 1: Unset
"ld1 { v11.h }[0], [x20], #0x2\n"
"47:" // Oddments: Load input (3, 4): Bit 2: End
- "ldr x20, [x14, #0x78]\n"
+ "ldr x20, [x15, #0x78]\n"
"fmla v28.8h, v8.8h, v11.8h\n"
"fmla v31.8h, v5.8h, v11.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 49f\n"
"ld1 { v13.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 48f\n"
@@ -825,10 +825,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"50:" // Oddments: Load input (4, 1): Bit 2: Unset: Bit 1: Unset
"ld1 { v13.h }[0], [x20], #0x2\n"
"51:" // Oddments: Load input (4, 1): Bit 2: End
- "ldr x20, [x14, #0x80]\n"
+ "ldr x20, [x15, #0x80]\n"
"fmla v29.8h, v7.8h, v13.8h\n"
"fmla v30.8h, v6.8h, v13.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 53f\n"
"ld1 { v12.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 52f\n"
@@ -849,10 +849,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"54:" // Oddments: Load input (1, 1): Bit 2: Unset: Bit 1: Unset
"ld1 { v12.h }[0], [x20], #0x2\n"
"55:" // Oddments: Load input (1, 1): Bit 2: End
- "ldr x20, [x14, #0x88]\n"
+ "ldr x20, [x15, #0x88]\n"
"fmla v23.8h, v4.8h, v12.8h\n"
"fmla v24.8h, v3.8h, v12.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"fmla v26.8h, v1.8h, v12.8h\n"
"fmla v27.8h, v0.8h, v12.8h\n"
"tbz %x[n_channels], #2, 57f\n"
@@ -875,10 +875,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"58:" // Oddments: Load input (1, 3): Bit 2: Unset: Bit 1: Unset
"ld1 { v11.h }[0], [x20], #0x2\n"
"59:" // Oddments: Load input (1, 3): Bit 2: End
- "ldr x20, [x14, #0x90]\n"
+ "ldr x20, [x15, #0x90]\n"
"fmla v24.8h, v5.8h, v11.8h\n"
"fmla v25.8h, v4.8h, v11.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"fmla v27.8h, v2.8h, v11.8h\n"
"fmla v28.8h, v1.8h, v11.8h\n"
"tbz %x[n_channels], #2, 61f\n"
@@ -901,10 +901,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"62:" // Oddments: Load input (4, 3): Bit 2: Unset: Bit 1: Unset
"ld1 { v13.h }[0], [x20], #0x2\n"
"63:" // Oddments: Load input (4, 3): Bit 2: End
- "ldr x20, [x14, #0x98]\n"
+ "ldr x20, [x15, #0x98]\n"
"fmla v30.8h, v8.8h, v13.8h\n"
"fmla v31.8h, v7.8h, v13.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"tbz %x[n_channels], #2, 65f\n"
"ld1 { v12.d }[0], [x20], #0x8\n"
"tbz %x[n_channels], #1, 64f\n"
@@ -925,10 +925,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"66:" // Oddments: Load input (3, 1): Bit 2: Unset: Bit 1: Unset
"ld1 { v12.h }[0], [x20], #0x2\n"
"67:" // Oddments: Load input (3, 1): Bit 2: End
- "ldr x20, [x14, #0xa0]\n"
+ "ldr x20, [x15, #0xa0]\n"
"fmla v26.8h, v7.8h, v12.8h\n"
"fmla v27.8h, v6.8h, v12.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"fmla v29.8h, v4.8h, v12.8h\n"
"fmla v30.8h, v3.8h, v12.8h\n"
"tbz %x[n_channels], #2, 69f\n"
@@ -951,10 +951,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"70:" // Oddments: Load input (0, 2): Bit 2: Unset: Bit 1: Unset
"ld1 { v11.h }[0], [x20], #0x2\n"
"71:" // Oddments: Load input (0, 2): Bit 2: End
- "ldr x20, [x14, #0xa8]\n"
+ "ldr x20, [x15, #0xa8]\n"
"fmla v23.8h, v2.8h, v11.8h\n"
"fmla v24.8h, v1.8h, v11.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"fmla v25.8h, v0.8h, v11.8h\n"
"tbz %x[n_channels], #2, 73f\n"
"ld1 { v13.d }[0], [x20], #0x8\n"
@@ -976,10 +976,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"74:" // Oddments: Load input (3, 3): Bit 2: Unset: Bit 1: Unset
"ld1 { v13.h }[0], [x20], #0x2\n"
"75:" // Oddments: Load input (3, 3): Bit 2: End
- "ldr x20, [x14, #0xb0]\n"
+ "ldr x20, [x15, #0xb0]\n"
"fmla v27.8h, v8.8h, v13.8h\n"
"fmla v28.8h, v7.8h, v13.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"fmla v30.8h, v5.8h, v13.8h\n"
"fmla v31.8h, v4.8h, v13.8h\n"
"tbz %x[n_channels], #2, 77f\n"
@@ -1002,10 +1002,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"78:" // Oddments: Load input (2, 0): Bit 2: Unset: Bit 1: Unset
"ld1 { v12.h }[0], [x20], #0x2\n"
"79:" // Oddments: Load input (2, 0): Bit 2: End
- "ldr x20, [x14, #0xb8]\n"
+ "ldr x20, [x15, #0xb8]\n"
"fmla v23.8h, v6.8h, v12.8h\n"
"fmla v26.8h, v3.8h, v12.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"fmla v29.8h, v0.8h, v12.8h\n"
"tbz %x[n_channels], #2, 81f\n"
"ld1 { v11.d }[0], [x20], #0x8\n"
@@ -1027,10 +1027,10 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"82:" // Oddments: Load input (2, 4): Bit 2: Unset: Bit 1: Unset
"ld1 { v11.h }[0], [x20], #0x2\n"
"83:" // Oddments: Load input (2, 4): Bit 2: End
- "ldr x20, [x14, #0xc0]\n"
+ "ldr x20, [x15, #0xc0]\n"
"fmla v25.8h, v8.8h, v11.8h\n"
"fmla v28.8h, v5.8h, v11.8h\n"
- "add x20, x20, x13\n"
+ "add x20, x20, x14\n"
"fmla v31.8h, v2.8h, v11.8h\n"
"tbz %x[n_channels], #2, 85f\n"
"ld1 { v13.d }[0], [x20], #0x8\n"
@@ -1054,236 +1054,234 @@ void a64_fp16_nhwc_3x3_s1_output3x3_mla_depthfirst_indirect_impl(
"87:" // Oddments: Load input (4, 2): Bit 2: End
"fmla v29.8h, v8.8h, v13.8h\n"
"fmla v30.8h, v7.8h, v13.8h\n"
- "fmax v23.8h, v23.8h, v18.8h\n"
+ "fmax v23.8h, v23.8h, v15.8h\n"
"fmla v31.8h, v6.8h, v13.8h\n"
- "fmax v24.8h, v24.8h, v18.8h\n"
- "fmax v25.8h, v25.8h, v18.8h\n"
- "fmax v26.8h, v26.8h, v18.8h\n"
- "fmax v27.8h, v27.8h, v18.8h\n"
- "fmax v28.8h, v28.8h, v18.8h\n"
- "fmax v29.8h, v29.8h, v18.8h\n"
- "fmax v30.8h, v30.8h, v18.8h\n"
- "fmax v31.8h, v31.8h, v18.8h\n"
- "fmin v23.8h, v23.8h, v17.8h\n"
- "fmin v24.8h, v24.8h, v17.8h\n"
- "fmin v25.8h, v25.8h, v17.8h\n"
- "fmin v26.8h, v26.8h, v17.8h\n"
- "fmin v27.8h, v27.8h, v17.8h\n"
- "fmin v28.8h, v28.8h, v17.8h\n"
- "fmin v29.8h, v29.8h, v17.8h\n"
- "fmin v30.8h, v30.8h, v17.8h\n"
- "fmin v31.8h, v31.8h, v17.8h\n"
+ "fmax v24.8h, v24.8h, v15.8h\n"
+ "fmax v25.8h, v25.8h, v15.8h\n"
+ "fmax v26.8h, v26.8h, v15.8h\n"
+ "fmax v27.8h, v27.8h, v15.8h\n"
+ "fmax v28.8h, v28.8h, v15.8h\n"
+ "fmax v29.8h, v29.8h, v15.8h\n"
+ "fmax v30.8h, v30.8h, v15.8h\n"
+ "fmax v31.8h, v31.8h, v15.8h\n"
+ "fmin v23.8h, v23.8h, v14.8h\n"
+ "fmin v24.8h, v24.8h, v14.8h\n"
+ "fmin v25.8h, v25.8h, v14.8h\n"
+ "fmin v26.8h, v26.8h, v14.8h\n"
+ "fmin v27.8h, v27.8h, v14.8h\n"
+ "fmin v28.8h, v28.8h, v14.8h\n"
+ "fmin v29.8h, v29.8h, v14.8h\n"
+ "fmin v30.8h, v30.8h, v14.8h\n"
+ "fmin v31.8h, v31.8h, v14.8h\n"
"tbz %x[n_channels], #2, 89f\n"
- "ldr x23, [x16, #0x0]\n"
- "add x23, x23, x12\n"
- "st1 { v23.d }[0], [x23]\n"
- "ldr x22, [x16, #0x8]\n"
- "ldr x21, [x16, #0x10]\n"
- "ldr x20, [x16, #0x18]\n"
- "add x22, x22, x12\n"
- "add x21, x21, x12\n"
- "ldr x23, [x16, #0x20]\n"
- "add x20, x20, x12\n"
- "add x23, x23, x12\n"
- "st1 { v24.d }[0], [x22]\n"
- "st1 { v25.d }[0], [x21]\n"
- "ldr x22, [x16, #0x28]\n"
- "ldr x21, [x16, #0x30]\n"
- "add x22, x22, x12\n"
- "st1 { v26.d }[0], [x20]\n"
- "ldr x20, [x16, #0x38]\n"
- "add x21, x21, x12\n"
- "add x20, x20, x12\n"
- "st1 { v27.d }[0], [x23]\n"
- "ldr x23, [x16, #0x40]\n"
- "add x23, x23, x12\n"
- "add x12, x12, #0x8\n"
- "st1 { v28.d }[0], [x22]\n"
- "st1 { v29.d }[0], [x21]\n"
- "st1 { v30.d }[0], [x20]\n"
- "st1 { v31.d }[0], [x23]\n"
+ "ldr x20, [x17, #0x0]\n"
+ "add x20, x20, x13\n"
+ "st1 { v23.d }[0], [x20]\n"
+ "ldr x23, [x17, #0x8]\n"
+ "ldr x22, [x17, #0x10]\n"
+ "ldr x21, [x17, #0x18]\n"
+ "add x23, x23, x13\n"
+ "add x22, x22, x13\n"
+ "ldr x20, [x17, #0x20]\n"
+ "add x21, x21, x13\n"
+ "add x20, x20, x13\n"
+ "st1 { v24.d }[0], [x23]\n"
+ "st1 { v25.d }[0], [x22]\n"
+ "ldr x23, [x17, #0x28]\n"
+ "ldr x22, [x17, #0x30]\n"
+ "add x23, x23, x13\n"
+ "st1 { v26.d }[0], [x21]\n"
+ "ldr x21, [x17, #0x38]\n"
+ "add x22, x22, x13\n"
+ "add x21, x21, x13\n"
+ "st1 { v27.d }[0], [x20]\n"
+ "ldr x20, [x17, #0x40]\n"
+ "add x20, x20, x13\n"
+ "add x13, x13, #0x8\n"
+ "st1 { v28.d }[0], [x23]\n"
+ "st1 { v29.d }[0], [x22]\n"
+ "st1 { v30.d }[0], [x21]\n"
+ "st1 { v31.d }[0], [x20]\n"
"tbz %x[n_channels], #1, 88f\n"
- "ldr x23, [x16, #0x0]\n"
- "add x23, x23, x12\n"
- "st1 { v23.s }[2], [x23]\n"
- "ldr x22, [x16, #0x8]\n"
- "ldr x21, [x16, #0x10]\n"
- "ldr x20, [x16, #0x18]\n"
- "add x22, x22, x12\n"
- "add x21, x21, x12\n"
- "ldr x23, [x16, #0x20]\n"
- "add x20, x20, x12\n"
- "add x23, x23, x12\n"
- "st1 { v24.s }[2], [x22]\n"
- "st1 { v25.s }[2], [x21]\n"
- "ldr x22, [x16, #0x28]\n"
- "ldr x21, [x16, #0x30]\n"
- "add x22, x22, x12\n"
- "st1 { v26.s }[2], [x20]\n"
- "ldr x20, [x16, #0x38]\n"
- "add x21, x21, x12\n"
- "add x20, x20, x12\n"
- "st1 { v27.s }[2], [x23]\n"
- "ldr x23, [x16, #0x40]\n"
- "add x23, x23, x12\n"
- "add x12, x12, #0x4\n"
- "st1 { v28.s }[2], [x22]\n"
- "st1 { v29.s }[2], [x21]\n"
- "st1 { v30.s }[2], [x20]\n"
- "st1 { v31.s }[2], [x23]\n"
+ "ldr x20, [x17, #0x0]\n"
+ "add x20, x20, x13\n"
+ "st1 { v23.s }[2], [x20]\n"
+ "ldr x23, [x17, #0x8]\n"
+ "ldr x22, [x17, #0x10]\n"
+ "ldr x21, [x17, #0x18]\n"
+ "add x23, x23, x13\n"
+ "add x22, x22, x13\n"
+ "ldr x20, [x17, #0x20]\n"
+ "add x21, x21, x13\n"
+ "add x20, x20, x13\n"
+ "st1 { v24.s }[2], [x23]\n"
+ "st1 { v25.s }[2], [x22]\n"
+ "ldr x23, [x17, #0x28]\n"
+ "ldr x22, [x17, #0x30]\n"
+ "add x23, x23, x13\n"
+ "st1 { v26.s }[2], [x21]\n"
+ "ldr x21, [x17, #0x38]\n"
+ "add x22, x22, x13\n"
+ "add x21, x21, x13\n"
+ "st1 { v27.s }[2], [x20]\n"
+ "ldr x20, [x17, #0x40]\n"
+ "add x20, x20, x13\n"
+ "add x13, x13, #0x4\n"
+ "st1 { v28.s }[2], [x23]\n"
+ "st1 { v29.s }[2], [x22]\n"
+ "st1 { v30.s }[2], [x21]\n"
+ "st1 { v31.s }[2], [x20]\n"
"tbz %x[n_channels], #0, 91f\n"
- "ldr x23, [x16, #0x0]\n"
- "add x23, x23, x12\n"
- "st1 { v23.h }[6], [x23]\n"
- "ldr x22, [x16, #0x8]\n"
- "ldr x21, [x16, #0x10]\n"
- "ldr x20, [x16, #0x18]\n"
- "add x22, x22, x12\n"
- "add x21, x21, x12\n"
- "ldr x23, [x16, #0x20]\n"
- "add x20, x20, x12\n"
- "add x23, x23, x12\n"
- "st1 { v24.h }[6], [x22]\n"
- "st1 { v25.h }[6], [x21]\n"
- "ldr x22, [x16, #0x28]\n"
- "ldr x21, [x16, #0x30]\n"
- "add x22, x22, x12\n"
- "st1 { v26.h }[6], [x20]\n"
- "ldr x20, [x16, #0x38]\n"
- "add x21, x21, x12\n"
- "add x20, x20, x12\n"
- "st1 { v27.h }[6], [x23]\n"
- "ldr x23, [x16, #0x40]\n"
- "add x23, x23, x12\n"
- "st1 { v28.h }[6], [x22]\n"
- "st1 { v29.h }[6], [x21]\n"
- "st1 { v30.h }[6], [x20]\n"
- "st1 { v31.h }[6], [x23]\n"
+ "ldr x20, [x17, #0x0]\n"
+ "add x20, x20, x13\n"
+ "st1 { v23.h }[6], [x20]\n"
+ "ldr x23, [x17, #0x8]\n"
+ "ldr x22, [x17, #0x10]\n"
+ "ldr x21, [x17, #0x18]\n"
+ "add x23, x23, x13\n"
+ "add x22, x22, x13\n"
+ "ldr x20, [x17, #0x20]\n"
+ "add x21, x21, x13\n"
+ "add x20, x20, x13\n"
+ "st1 { v24.h }[6], [x23]\n"
+ "st1 { v25.h }[6], [x22]\n"
+ "ldr x23, [x17, #0x28]\n"
+ "ldr x22, [x17, #0x30]\n"
+ "add x23, x23, x13\n"
+ "st1 { v26.h }[6], [x21]\n"
+ "ldr x21, [x17, #0x38]\n"
+ "add x22, x22, x13\n"
+ "add x21, x21, x13\n"
+ "st1 { v27.h }[6], [x20]\n"
+ "ldr x20, [x17, #0x40]\n"
+ "add x20, x20, x13\n"
+ "st1 { v28.h }[6], [x23]\n"
+ "st1 { v29.h }[6], [x22]\n"
+ "st1 { v30.h }[6], [x21]\n"
+ "st1 { v31.h }[6], [x20]\n"
"b 91f\n"
"88:" // Oddments: Store: Bit 2: Bit 1: Unset
"tbz %x[n_channels], #0, 91f\n"
- "ldr x23, [x16, #0x0]\n"
- "add x23, x23, x12\n"
- "st1 { v23.h }[4], [x23]\n"
- "ldr x22, [x16, #0x8]\n"
- "ldr x21, [x16, #0x10]\n"
- "ldr x20, [x16, #0x18]\n"
- "add x22, x22, x12\n"
- "add x21, x21, x12\n"
- "ldr x23, [x16, #0x20]\n"
- "add x20, x20, x12\n"
- "add x23, x23, x12\n"
- "st1 { v24.h }[4], [x22]\n"
- "st1 { v25.h }[4], [x21]\n"
- "ldr x22, [x16, #0x28]\n"
- "ldr x21, [x16, #0x30]\n"
- "add x22, x22, x12\n"
- "st1 { v26.h }[4], [x20]\n"
- "ldr x20, [x16, #0x38]\n"
- "add x21, x21, x12\n"
- "add x20, x20, x12\n"
- "st1 { v27.h }[4], [x23]\n"
- "ldr x23, [x16, #0x40]\n"
- "add x23, x23, x12\n"
- "st1 { v28.h }[4], [x22]\n"
- "st1 { v29.h }[4], [x21]\n"
- "st1 { v30.h }[4], [x20]\n"
- "st1 { v31.h }[4], [x23]\n"
+ "ldr x20, [x17, #0x0]\n"
+ "add x20, x20, x13\n"
+ "st1 { v23.h }[4], [x20]\n"
+ "ldr x23, [x17, #0x8]\n"
+ "ldr x22, [x17, #0x10]\n"
+ "ldr x21, [x17, #0x18]\n"
+ "add x23, x23, x13\n"
+ "add x22, x22, x13\n"
+ "ldr x20, [x17, #0x20]\n"
+ "add x21, x21, x13\n"
+ "add x20, x20, x13\n"
+ "st1 { v24.h }[4], [x23]\n"
+ "st1 { v25.h }[4], [x22]\n"
+ "ldr x23, [x17, #0x28]\n"
+ "ldr x22, [x17, #0x30]\n"
+ "add x23, x23, x13\n"
+ "st1 { v26.h }[4], [x21]\n"
+ "ldr x21, [x17, #0x38]\n"
+ "add x22, x22, x13\n"
+ "add x21, x21, x13\n"
+ "st1 { v27.h }[4], [x20]\n"
+ "ldr x20, [x17, #0x40]\n"
+ "add x20, x20, x13\n"
+ "st1 { v28.h }[4], [x23]\n"
+ "st1 { v29.h }[4], [x22]\n"
+ "st1 { v30.h }[4], [x21]\n"
+ "st1 { v31.h }[4], [x20]\n"
"b 91f\n"
"89:" // Oddments: Store: Bit 2: Unset
"tbz %x[n_channels], #1, 90f\n"
- "ldr x23, [x16, #0x0]\n"
- "add x23, x23, x12\n"
- "st1 { v23.s }[0], [x23]\n"
- "ldr x22, [x16, #0x8]\n"
- "ldr x21, [x16, #0x10]\n"
- "ldr x20, [x16, #0x18]\n"
- "add x22, x22, x12\n"
- "add x21, x21, x12\n"
- "ldr x23, [x16, #0x20]\n"
- "add x20, x20, x12\n"
- "add x23, x23, x12\n"
- "st1 { v24.s }[0], [x22]\n"
- "st1 { v25.s }[0], [x21]\n"
- "ldr x22, [x16, #0x28]\n"
- "ldr x21, [x16, #0x30]\n"
- "add x22, x22, x12\n"
- "st1 { v26.s }[0], [x20]\n"
- "ldr x20, [x16, #0x38]\n"
- "add x21, x21, x12\n"
- "add x20, x20, x12\n"
- "st1 { v27.s }[0], [x23]\n"
- "ldr x23, [x16, #0x40]\n"
- "add x23, x23, x12\n"
- "add x12, x12, #0x4\n"
- "st1 { v28.s }[0], [x22]\n"
- "st1 { v29.s }[0], [x21]\n"
- "st1 { v30.s }[0], [x20]\n"
- "st1 { v31.s }[0], [x23]\n"
+ "ldr x20, [x17, #0x0]\n"
+ "add x20, x20, x13\n"
+ "st1 { v23.s }[0], [x20]\n"
+ "ldr x23, [x17, #0x8]\n"
+ "ldr x22, [x17, #0x10]\n"
+ "ldr x21, [x17, #0x18]\n"
+ "add x23, x23, x13\n"
+ "add x22, x22, x13\n"
+ "ldr x20, [x17, #0x20]\n"
+ "add x21, x21, x13\n"
+ "add x20, x20, x13\n"
+ "st1 { v24.s }[0], [x23]\n"
+ "st1 { v25.s }[0], [x22]\n"
+ "ldr x23, [x17, #0x28]\n"
+ "ldr x22, [x17, #0x30]\n"
+ "add x23, x23, x13\n"
+ "st1 { v26.s }[0], [x21]\n"
+ "ldr x21, [x17, #0x38]\n"
+ "add x22, x22, x13\n"
+ "add x21, x21, x13\n"
+ "st1 { v27.s }[0], [x20]\n"
+ "ldr x20, [x17, #0x40]\n"
+ "add x20, x20, x13\n"
+ "add x13, x13, #0x4\n"
+ "st1 { v28.s }[0], [x23]\n"
+ "st1 { v29.s }[0], [x22]\n"
+ "st1 { v30.s }[0], [x21]\n"
+ "st1 { v31.s }[0], [x20]\n"
"tbz %x[n_channels], #0, 91f\n"
- "ldr x23, [x16, #0x0]\n"
- "add x23, x23, x12\n"
- "st1 { v23.h }[2], [x23]\n"
- "ldr x22, [x16, #0x8]\n"
- "ldr x21, [x16, #0x10]\n"
- "ldr x20, [x16, #0x18]\n"
- "add x22, x22, x12\n"
- "add x21, x21, x12\n"
- "ldr x23, [x16, #0x20]\n"
- "add x20, x20, x12\n"
- "add x23, x23, x12\n"
- "st1 { v24.h }[2], [x22]\n"
- "st1 { v25.h }[2], [x21]\n"
- "ldr x22, [x16, #0x28]\n"
- "ldr x21, [x16, #0x30]\n"
- "add x22, x22, x12\n"
- "st1 { v26.h }[2], [x20]\n"
- "ldr x20, [x16, #0x38]\n"
- "add x21, x21, x12\n"
- "add x20, x20, x12\n"
- "st1 { v27.h }[2], [x23]\n"
- "ldr x23, [x16, #0x40]\n"
- "add x23, x23, x12\n"
- "st1 { v28.h }[2], [x22]\n"
- "st1 { v29.h }[2], [x21]\n"
- "st1 { v30.h }[2], [x20]\n"
- "st1 { v31.h }[2], [x23]\n"
+ "ldr x20, [x17, #0x0]\n"
+ "add x20, x20, x13\n"
+ "st1 { v23.h }[2], [x20]\n"
+ "ldr x23, [x17, #0x8]\n"
+ "ldr x22, [x17, #0x10]\n"
+ "ldr x21, [x17, #0x18]\n"
+ "add x23, x23, x13\n"
+ "add x22, x22, x13\n"
+ "ldr x20, [x17, #0x20]\n"
+ "add x21, x21, x13\n"
+ "add x20, x20, x13\n"
+ "st1 { v24.h }[2], [x23]\n"
+ "st1 { v25.h }[2], [x22]\n"
+ "ldr x23, [x17, #0x28]\n"
+ "ldr x22, [x17, #0x30]\n"
+ "add x23, x23, x13\n"
+ "st1 { v26.h }[2], [x21]\n"
+ "ldr x21, [x17, #0x38]\n"
+ "add x22, x22, x13\n"
+ "add x21, x21, x13\n"
+ "st1 { v27.h }[2], [x20]\n"
+ "ldr x20, [x17, #0x40]\n"
+ "add x20, x20, x13\n"
+ "st1 { v28.h }[2], [x23]\n"
+ "st1 { v29.h }[2], [x22]\n"
+ "st1 { v30.h }[2], [x21]\n"
+ "st1 { v31.h }[2], [x20]\n"
"b 91f\n"
"90:" // Oddments: Store: Bit 2: Unset: Bit 1: Unset
- "ldr x23, [x16, #0x0]\n"
- "add x23, x23, x12\n"
- "st1 { v23.h }[0], [x23]\n"
- "ldr x22, [x16, #0x8]\n"
- "ldr x21, [x16, #0x10]\n"
- "ldr x20, [x16, #0x18]\n"
- "add x22, x22, x12\n"
- "add x21, x21, x12\n"
- "ldr x23, [x16, #0x20]\n"
- "add x20, x20, x12\n"
- "add x23, x23, x12\n"
- "st1 { v24.h }[0], [x22]\n"
- "st1 { v25.h }[0], [x21]\n"
- "ldr x22, [x16, #0x28]\n"
- "ldr x21, [x16, #0x30]\n"
- "add x22, x22, x12\n"
- "st1 { v26.h }[0], [x20]\n"
- "ldr x20, [x16, #0x38]\n"
- "add x21, x21, x12\n"
- "add x20, x20, x12\n"
- "st1 { v27.h }[0], [x23]\n"
- "ldr x23, [x16, #0x40]\n"
- "add x23, x23, x12\n"
- "st1 { v28.h }[0], [x22]\n"
- "st1 { v29.h }[0], [x21]\n"
- "st1 { v30.h }[0], [x20]\n"
- "st1 { v31.h }[0], [x23]\n"
+ "ldr x20, [x17, #0x0]\n"
+ "add x20, x20, x13\n"
+ "st1 { v23.h }[0], [x20]\n"
+ "ldr x23, [x17, #0x8]\n"
+ "ldr x22, [x17, #0x10]\n"
+ "ldr x21, [x17, #0x18]\n"
+ "add x23, x23, x13\n"
+ "add x22, x22, x13\n"
+ "ldr x20, [x17, #0x20]\n"
+ "add x21, x21, x13\n"
+ "add x20, x20, x13\n"
+ "st1 { v24.h }[0], [x23]\n"
+ "st1 { v25.h }[0], [x22]\n"
+ "ldr x23, [x17, #0x28]\n"
+ "ldr x22, [x17, #0x30]\n"
+ "add x23, x23, x13\n"
+ "st1 { v26.h }[0], [x21]\n"
+ "ldr x21, [x17, #0x38]\n"
+ "add x22, x22, x13\n"
+ "add x21, x21, x13\n"
+ "st1 { v27.h }[0], [x20]\n"
+ "ldr x20, [x17, #0x40]\n"
+ "add x20, x20, x13\n"
+ "st1 { v28.h }[0], [x23]\n"
+ "st1 { v29.h }[0], [x22]\n"
+ "st1 { v30.h }[0], [x21]\n"
+ "st1 { v31.h }[0], [x20]\n"
"91:" // Oddments: Store: Bit 2: End
-
"92:" // 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", "v16", "v17", "v18", "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"
+ : "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", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
);
}
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-10 15:41:27 +0000