aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_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_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp810
1 files changed, 404 insertions, 406 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp
index 860adac326..32939eb6dc 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst/generic_indirect.cpp
@@ -103,16 +103,16 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"lsr x9, %x[n_channels], #0x2\n"
"ldr x16, [%x[params_struct], %[offsetof_args_params]]\n"
"add x20, %x[params_struct], %[offsetof_args_min]\n"
- "ld1r { v18.4s }, [x20]\n"
+ "ld1r { v27.4s }, [x20]\n"
"add x20, %x[params_struct], %[offsetof_args_max]\n"
- "ld1r { v17.4s }, [x20]\n"
+ "ld1r { v15.4s }, [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"
- "ldr q16, [x16, #0x0]\n"
+ "ldr q26, [x16, #0x0]\n"
"ldr q0, [x16, #0x10]\n"
"cmp x17, x9, LSL #4\n"
"ldr q1, [x16, #0x20]\n"
@@ -120,436 +120,436 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ldr q3, [x16, #0x40]\n"
"ldr q4, [x16, #0x50]\n"
"add x16, x16, #0x60\n"
- "ldp x27, x26, [x15, #0x0]\n"
- "ldr q5, [x27, x10]\n"
- "ldr q6, [x26, x10]\n"
- "ldp x25, x24, [x15, #0x10]\n"
- "ldr q7, [x25, x10]\n"
- "ldr q8, [x24, x10]\n"
- "ldp x23, x22, [x15, #0x20]\n"
- "ldr q9, [x23, x10]\n"
- "ldr q13, [x22, x10]\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 x21, x20, [x15, #0x30]\n"
"ldr q11, [x21, x10]\n"
"ldr q12, [x20, x10]\n"
- "ldp x27, x26, [x15, #0x40]\n"
- "ldr q10, [x27, x10]\n"
- "ldr q14, [x26, x10]\n"
+ "ldp x21, x20, [x15, #0x40]\n"
+ "ldr q10, [x21, x10]\n"
+ "ldr q14, [x20, x10]\n"
"bge 2f\n"
"1:" // Channel loop
- "mov v28.16b, v16.16b\n fmla v28.4s, v0.4s, v5.4s\n"
- "mov v29.16b, v16.16b\n fmla v29.4s, v0.4s, v6.4s\n"
- "ldr x25, [x15, #0x50]\n"
- "ldr q5, [x25, x10]\n"
- "mov v30.16b, v16.16b\n fmla v30.4s, v0.4s, v7.4s\n"
- "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v8.4s\n"
- "ldr q0, [x16, #0x0]\n"
- "ldr q16, [x16, #0x140]\n"
- "fmla v28.4s, v1.4s, v6.4s\n"
- "fmla v29.4s, v1.4s, v9.4s\n"
- "ldr x24, [x15, #0x58]\n"
- "ldr q6, [x24, x10]\n"
- "fmla v30.4s, v1.4s, v8.4s\n"
- "fmla v31.4s, v1.4s, v13.4s\n"
- "ldr q1, [x16, #0x10]\n"
- "ldr x23, [x15, #0x60]\n"
- "fmla v28.4s, v2.4s, v9.4s\n"
- "ldr q9, [x23, x10]\n"
- "fmla v29.4s, v2.4s, v11.4s\n"
- "ldr x22, [x15, #0x68]\n"
- "fmla v30.4s, v2.4s, v13.4s\n"
- "fmla v31.4s, v2.4s, v5.4s\n"
- "ldr q2, [x16, #0x20]\n"
- "ldr x21, [x15, #0x70]\n"
- "fmla v28.4s, v3.4s, v11.4s\n"
- "ldr q11, [x22, x10]\n"
- "fmla v29.4s, v3.4s, v12.4s\n"
- "ldr x20, [x15, #0x78]\n"
- "fmla v30.4s, v3.4s, v5.4s\n"
- "fmla v31.4s, v3.4s, v6.4s\n"
- "ldr q3, [x16, #0x30]\n"
- "ldr x27, [x15, #0x80]\n"
- "fmla v28.4s, v4.4s, v12.4s\n"
- "ldr q12, [x21, x10]\n"
- "fmla v29.4s, v4.4s, v9.4s\n"
- "ldr q9, [x20, x10]\n"
- "fmla v30.4s, v4.4s, v6.4s\n"
- "fmla v31.4s, v4.4s, v10.4s\n"
- "ldr q4, [x16, #0x40]\n"
- "ldr x26, [x15, #0x88]\n"
- "fmla v28.4s, v0.4s, v7.4s\n"
- "fmla v29.4s, v0.4s, v8.4s\n"
- "ldr x25, [x15, #0x90]\n"
- "ldr x24, [x15, #0x98]\n"
- "fmla v30.4s, v0.4s, v14.4s\n"
- "fmla v31.4s, v0.4s, v11.4s\n"
- "ldr q0, [x16, #0x50]\n"
- "ldr x23, [x15, #0xa0]\n"
+ "mov v30.16b, v26.16b\n fmla v30.4s, v0.4s, v5.4s\n"
+ "mov v31.16b, v26.16b\n fmla v31.4s, v0.4s, v6.4s\n"
+ "ldr x20, [x15, #0x50]\n"
+ "ldr q24, [x20, x10]\n"
+ "mov v28.16b, v26.16b\n fmla v28.4s, v0.4s, v7.4s\n"
+ "mov v29.16b, v26.16b\n fmla v29.4s, v0.4s, v8.4s\n"
+ "ldr q23, [x16, #0x0]\n"
+ "ldr q26, [x16, #0x140]\n"
+ "fmla v30.4s, v1.4s, v6.4s\n"
+ "fmla v31.4s, v1.4s, v9.4s\n"
+ "ldr x20, [x15, #0x58]\n"
+ "ldr q22, [x20, x10]\n"
"fmla v28.4s, v1.4s, v8.4s\n"
- "ldr q8, [x26, x10]\n"
"fmla v29.4s, v1.4s, v13.4s\n"
- "ldr x22, [x15, #0xa8]\n"
- "fmla v30.4s, v1.4s, v11.4s\n"
- "fmla v31.4s, v1.4s, v12.4s\n"
- "ldr q1, [x16, #0x60]\n"
- "ldr x21, [x15, #0xb0]\n"
- "fmla v28.4s, v2.4s, v13.4s\n"
- "ldr q13, [x27, x10]\n"
- "fmla v29.4s, v2.4s, v5.4s\n"
- "ldr x20, [x15, #0xb8]\n"
- "fmla v30.4s, v2.4s, v12.4s\n"
- "fmla v31.4s, v2.4s, v9.4s\n"
- "ldr q2, [x16, #0x70]\n"
- "ldr x27, [x15, #0xc0]\n"
- "fmla v28.4s, v3.4s, v5.4s\n"
- "ldr q5, [x25, x10]\n"
- "fmla v29.4s, v3.4s, v6.4s\n"
- "ldr x26, [x15, #0xc8]\n"
- "fmla v30.4s, v3.4s, v9.4s\n"
- "fmla v31.4s, v3.4s, v13.4s\n"
- "ldr q3, [x16, #0x80]\n"
- "ldr x25, [x15, #0xd0]\n"
- "fmla v28.4s, v4.4s, v6.4s\n"
- "ldr q6, [x24, x10]\n"
- "fmla v29.4s, v4.4s, v10.4s\n"
- "ldr q10, [x23, x10]\n"
- "fmla v30.4s, v4.4s, v13.4s\n"
- "fmla v31.4s, v4.4s, v8.4s\n"
- "ldr q4, [x16, #0x90]\n"
- "ldr x24, [x15, #0xd8]\n"
- "fmla v28.4s, v0.4s, v14.4s\n"
- "ldr q14, [x20, x10]\n"
- "fmla v29.4s, v0.4s, v11.4s\n"
- "ldr x23, [x15, #0xe0]\n"
- "fmla v30.4s, v0.4s, v5.4s\n"
- "fmla v31.4s, v0.4s, v6.4s\n"
- "ldr q0, [x16, #0xa0]\n"
- "ldr x20, [x15, #0xf8]\n"
- "fmla v28.4s, v1.4s, v11.4s\n"
- "ldr q11, [x22, x10]\n"
- "fmla v29.4s, v1.4s, v12.4s\n"
- "ldr x22, [x15, #0xe8]\n"
- "fmla v30.4s, v1.4s, v6.4s\n"
- "fmla v31.4s, v1.4s, v10.4s\n"
- "ldr q1, [x16, #0xb0]\n"
- "add x28, x28, #0x10\n"
- "fmla v28.4s, v2.4s, v12.4s\n"
- "ldr q12, [x21, x10]\n"
- "fmla v29.4s, v2.4s, v9.4s\n"
- "ldr x21, [x15, #0xf0]\n"
- "fmla v30.4s, v2.4s, v10.4s\n"
+ "ldr q21, [x16, #0x10]\n"
+ "ldr x20, [x15, #0x60]\n"
+ "fmla v30.4s, v2.4s, v9.4s\n"
+ "ldr q17, [x20, x10]\n"
"fmla v31.4s, v2.4s, v11.4s\n"
- "ldr q2, [x16, #0xc0]\n"
- "fmla v28.4s, v3.4s, v9.4s\n"
- "ldr q9, [x27, x10]\n"
- "fmla v29.4s, v3.4s, v13.4s\n"
- "ldr x27, [x15, #0x100]\n"
+ "ldr x20, [x15, #0x68]\n"
+ "fmla v28.4s, v2.4s, v13.4s\n"
+ "fmla v29.4s, v2.4s, v24.4s\n"
+ "ldr q16, [x16, #0x20]\n"
+ "ldr x22, [x15, #0x70]\n"
"fmla v30.4s, v3.4s, v11.4s\n"
+ "ldr q5, [x20, x10]\n"
"fmla v31.4s, v3.4s, v12.4s\n"
- "ldr q3, [x16, #0xd0]\n"
- "fmla v28.4s, v4.4s, v13.4s\n"
- "ldr q13, [x26, x10]\n"
- "fmla v29.4s, v4.4s, v8.4s\n"
- "ldr q8, [x23, x10]\n"
+ "ldr x20, [x15, #0x78]\n"
+ "fmla v28.4s, v3.4s, v24.4s\n"
+ "fmla v29.4s, v3.4s, v22.4s\n"
+ "ldr q20, [x16, #0x30]\n"
+ "ldr x21, [x15, #0x80]\n"
"fmla v30.4s, v4.4s, v12.4s\n"
- "fmla v31.4s, v4.4s, v14.4s\n"
- "ldr q4, [x16, #0xe0]\n"
- "ldr x26, [x15, #0x108]\n"
- "fmla v28.4s, v0.4s, v5.4s\n"
- "ldr q5, [x25, x10]\n"
- "fmla v29.4s, v0.4s, v6.4s\n"
- "ldr x25, [x15, #0x110]\n"
- "fmla v30.4s, v0.4s, v9.4s\n"
- "fmla v31.4s, v0.4s, v13.4s\n"
- "ldr q0, [x16, #0xf0]\n"
- "fmla v28.4s, v1.4s, v6.4s\n"
- "ldr q6, [x24, x10]\n"
- "fmla v29.4s, v1.4s, v10.4s\n"
- "ldr x24, [x15, #0x118]\n"
- "fmla v30.4s, v1.4s, v13.4s\n"
+ "ldr q19, [x22, x10]\n"
+ "fmla v31.4s, v4.4s, v17.4s\n"
+ "ldr q2, [x20, x10]\n"
+ "fmla v28.4s, v4.4s, v22.4s\n"
+ "fmla v29.4s, v4.4s, v10.4s\n"
+ "ldr q18, [x16, #0x40]\n"
+ "ldr x20, [x15, #0x88]\n"
+ "fmla v30.4s, v23.4s, v7.4s\n"
+ "fmla v31.4s, v23.4s, v8.4s\n"
+ "ldr x23, [x15, #0x90]\n"
+ "ldr x26, [x15, #0x98]\n"
+ "fmla v28.4s, v23.4s, v14.4s\n"
+ "fmla v29.4s, v23.4s, v5.4s\n"
+ "ldr q1, [x16, #0x50]\n"
+ "ldr x22, [x15, #0xa0]\n"
+ "fmla v30.4s, v21.4s, v8.4s\n"
+ "ldr q25, [x20, x10]\n"
+ "fmla v31.4s, v21.4s, v13.4s\n"
+ "ldr x25, [x15, #0xa8]\n"
+ "fmla v28.4s, v21.4s, v5.4s\n"
+ "fmla v29.4s, v21.4s, v19.4s\n"
+ "ldr q17, [x16, #0x60]\n"
+ "ldr x24, [x15, #0xb0]\n"
+ "fmla v30.4s, v16.4s, v13.4s\n"
+ "ldr q8, [x21, x10]\n"
+ "fmla v31.4s, v16.4s, v24.4s\n"
+ "ldr x20, [x15, #0xb8]\n"
+ "fmla v28.4s, v16.4s, v19.4s\n"
+ "fmla v29.4s, v16.4s, v2.4s\n"
+ "ldr q16, [x16, #0x70]\n"
+ "ldr x21, [x15, #0xc0]\n"
+ "fmla v30.4s, v20.4s, v24.4s\n"
+ "ldr q24, [x23, x10]\n"
+ "fmla v31.4s, v20.4s, v22.4s\n"
+ "ldr x27, [x15, #0xc8]\n"
+ "fmla v28.4s, v20.4s, v2.4s\n"
+ "fmla v29.4s, v20.4s, v8.4s\n"
+ "ldr q23, [x16, #0x80]\n"
+ "ldr x23, [x15, #0xd0]\n"
+ "fmla v30.4s, v18.4s, v22.4s\n"
+ "ldr q22, [x26, x10]\n"
+ "fmla v31.4s, v18.4s, v10.4s\n"
+ "ldr q21, [x22, x10]\n"
+ "fmla v28.4s, v18.4s, v8.4s\n"
+ "fmla v29.4s, v18.4s, v25.4s\n"
+ "ldr q20, [x16, #0x90]\n"
+ "ldr x22, [x15, #0xd8]\n"
+ "fmla v30.4s, v1.4s, v14.4s\n"
+ "ldr q0, [x20, x10]\n"
"fmla v31.4s, v1.4s, v5.4s\n"
- "ldr q1, [x16, #0x100]\n"
- "fmla v28.4s, v2.4s, v10.4s\n"
- "ldr q10, [x22, x10]\n"
- "fmla v29.4s, v2.4s, v11.4s\n"
- "fmla v30.4s, v2.4s, v5.4s\n"
- "fmla v31.4s, v2.4s, v6.4s\n"
- "ldr q2, [x16, #0x110]\n"
- "fmla v28.4s, v3.4s, v11.4s\n"
- "ldr q11, [x21, x10]\n"
- "fmla v29.4s, v3.4s, v12.4s\n"
- "fmla v30.4s, v3.4s, v6.4s\n"
- "fmla v31.4s, v3.4s, v8.4s\n"
- "ldr q3, [x16, #0x120]\n"
- "fmla v28.4s, v4.4s, v12.4s\n"
- "ldr q12, [x20, x10]\n"
- "fmla v29.4s, v4.4s, v14.4s\n"
- "fmla v30.4s, v4.4s, v8.4s\n"
- "fmla v31.4s, v4.4s, v10.4s\n"
- "ldr q4, [x16, #0x130]\n"
- "fmla v28.4s, v0.4s, v9.4s\n"
- "ldr q9, [x27, x10]\n"
- "fmla v29.4s, v0.4s, v13.4s\n"
- "fmla v30.4s, v0.4s, v11.4s\n"
- "ldr q11, [x26, x10]\n"
- "fmla v31.4s, v0.4s, v12.4s\n"
+ "ldr x20, [x15, #0xe0]\n"
+ "fmla v28.4s, v1.4s, v24.4s\n"
+ "fmla v29.4s, v1.4s, v22.4s\n"
+ "ldr q6, [x16, #0xa0]\n"
+ "ldr x26, [x15, #0xf8]\n"
+ "fmla v30.4s, v17.4s, v5.4s\n"
+ "ldr q1, [x25, x10]\n"
+ "fmla v31.4s, v17.4s, v19.4s\n"
+ "ldr x25, [x15, #0xe8]\n"
+ "fmla v28.4s, v17.4s, v22.4s\n"
+ "fmla v29.4s, v17.4s, v21.4s\n"
+ "ldr q18, [x16, #0xb0]\n"
+ "add x28, x28, #0x10\n"
+ "fmla v30.4s, v16.4s, v19.4s\n"
+ "ldr q19, [x24, x10]\n"
+ "fmla v31.4s, v16.4s, v2.4s\n"
+ "ldr x24, [x15, #0xf0]\n"
+ "fmla v28.4s, v16.4s, v21.4s\n"
+ "fmla v29.4s, v16.4s, v1.4s\n"
+ "ldr q17, [x16, #0xc0]\n"
+ "fmla v30.4s, v23.4s, v2.4s\n"
+ "ldr q16, [x21, x10]\n"
+ "fmla v31.4s, v23.4s, v8.4s\n"
+ "ldr x21, [x15, #0x100]\n"
+ "fmla v28.4s, v23.4s, v1.4s\n"
+ "fmla v29.4s, v23.4s, v19.4s\n"
+ "ldr q13, [x16, #0xd0]\n"
+ "fmla v30.4s, v20.4s, v8.4s\n"
+ "ldr q2, [x27, x10]\n"
+ "fmla v31.4s, v20.4s, v25.4s\n"
+ "ldr q10, [x20, x10]\n"
+ "fmla v28.4s, v20.4s, v19.4s\n"
+ "fmla v29.4s, v20.4s, v0.4s\n"
+ "ldr q9, [x16, #0xe0]\n"
+ "ldr x20, [x15, #0x108]\n"
+ "fmla v30.4s, v6.4s, v24.4s\n"
+ "ldr q5, [x23, x10]\n"
+ "fmla v31.4s, v6.4s, v22.4s\n"
+ "ldr x23, [x15, #0x110]\n"
+ "fmla v28.4s, v6.4s, v16.4s\n"
+ "fmla v29.4s, v6.4s, v2.4s\n"
+ "ldr q24, [x16, #0xf0]\n"
+ "fmla v30.4s, v18.4s, v22.4s\n"
+ "ldr q25, [x22, x10]\n"
+ "fmla v31.4s, v18.4s, v21.4s\n"
+ "ldr x22, [x15, #0x118]\n"
+ "fmla v28.4s, v18.4s, v2.4s\n"
+ "fmla v29.4s, v18.4s, v5.4s\n"
+ "ldr q23, [x16, #0x100]\n"
+ "fmla v30.4s, v17.4s, v21.4s\n"
+ "ldr q22, [x25, x10]\n"
+ "fmla v31.4s, v17.4s, v1.4s\n"
+ "fmla v28.4s, v17.4s, v5.4s\n"
+ "fmla v29.4s, v17.4s, v25.4s\n"
+ "ldr q21, [x16, #0x110]\n"
+ "fmla v30.4s, v13.4s, v1.4s\n"
+ "ldr q18, [x24, x10]\n"
+ "fmla v31.4s, v13.4s, v19.4s\n"
+ "fmla v28.4s, v13.4s, v25.4s\n"
+ "fmla v29.4s, v13.4s, v10.4s\n"
+ "ldr q20, [x16, #0x120]\n"
+ "fmla v30.4s, v9.4s, v19.4s\n"
+ "ldr q17, [x26, x10]\n"
+ "fmla v31.4s, v9.4s, v0.4s\n"
+ "fmla v28.4s, v9.4s, v10.4s\n"
+ "fmla v29.4s, v9.4s, v22.4s\n"
+ "ldr q19, [x16, #0x130]\n"
+ "fmla v30.4s, v24.4s, v16.4s\n"
+ "ldr q16, [x21, x10]\n"
+ "fmla v31.4s, v24.4s, v2.4s\n"
+ "fmla v28.4s, v24.4s, v18.4s\n"
+ "ldr q18, [x20, x10]\n"
+ "fmla v29.4s, v24.4s, v17.4s\n"
"ldr q0, [x16, #0x150]\n"
- "fmla v28.4s, v1.4s, v13.4s\n"
- "fmla v29.4s, v1.4s, v5.4s\n"
- "ldp x27, x26, [x15, #0x0]\n"
- "fmla v30.4s, v1.4s, v12.4s\n"
- "ldr q12, [x25, x10]\n"
- "fmla v31.4s, v1.4s, v9.4s\n"
+ "fmla v30.4s, v23.4s, v2.4s\n"
+ "fmla v31.4s, v23.4s, v5.4s\n"
+ "ldp x21, x20, [x15, #0x0]\n"
+ "fmla v28.4s, v23.4s, v17.4s\n"
+ "ldr q17, [x23, x10]\n"
+ "fmla v29.4s, v23.4s, v16.4s\n"
"ldr q1, [x16, #0x160]\n"
- "fmla v28.4s, v2.4s, v5.4s\n"
- "ldr q5, [x27, x17]\n"
- "fmla v29.4s, v2.4s, v6.4s\n"
- "fmla v30.4s, v2.4s, v9.4s\n"
- "ldr q9, [x24, x10]\n"
- "fmla v31.4s, v2.4s, v11.4s\n"
+ "fmla v30.4s, v21.4s, v5.4s\n"
+ "ldr q5, [x21, x17]\n"
+ "fmla v31.4s, v21.4s, v25.4s\n"
+ "fmla v28.4s, v21.4s, v16.4s\n"
+ "ldr q16, [x22, x10]\n"
+ "fmla v29.4s, v21.4s, v18.4s\n"
"ldr q2, [x16, #0x170]\n"
- "fmla v28.4s, v3.4s, v6.4s\n"
- "ldr q6, [x26, x17]\n"
- "fmla v29.4s, v3.4s, v8.4s\n"
- "ldp x25, x24, [x15, #0x10]\n"
- "ldr q7, [x25, x17]\n"
- "fmla v30.4s, v3.4s, v11.4s\n"
- "fmla v31.4s, v3.4s, v12.4s\n"
+ "fmla v30.4s, v20.4s, v25.4s\n"
+ "ldr q6, [x20, x17]\n"
+ "fmla v31.4s, v20.4s, v10.4s\n"
+ "ldp x21, x20, [x15, #0x10]\n"
+ "ldr q7, [x21, x17]\n"
+ "fmla v28.4s, v20.4s, v18.4s\n"
+ "fmla v29.4s, v20.4s, v17.4s\n"
"ldr q3, [x16, #0x180]\n"
- "fmla v28.4s, v4.4s, v8.4s\n"
- "ldr q8, [x24, x17]\n"
- "fmla v29.4s, v4.4s, v10.4s\n"
- "ldp x23, x22, [x15, #0x20]\n"
- "ldr q13, [x22, x17]\n"
- "fmla v30.4s, v4.4s, v12.4s\n"
- "fmla v31.4s, v4.4s, v9.4s\n"
- "ldr q9, [x23, x17]\n"
+ "fmla v30.4s, v19.4s, v10.4s\n"
+ "ldr q8, [x20, x17]\n"
+ "fmla v31.4s, v19.4s, v22.4s\n"
+ "ldp x21, x20, [x15, #0x20]\n"
+ "ldr q13, [x20, x17]\n"
+ "fmla v28.4s, v19.4s, v17.4s\n"
+ "fmla v29.4s, v19.4s, v16.4s\n"
+ "ldr q9, [x21, x17]\n"
"ldr q4, [x16, #0x190]\n"
"ldp x21, x20, [x15, #0x30]\n"
- "fmax v28.4s, v28.4s, v18.4s\n"
- "fmax v29.4s, v29.4s, v18.4s\n"
+ "fmax v30.4s, v30.4s, v27.4s\n"
+ "fmax v31.4s, v31.4s, v27.4s\n"
"ldr q11, [x21, x17]\n"
"ldr q12, [x20, x17]\n"
- "fmax v30.4s, v30.4s, v18.4s\n"
- "fmax v31.4s, v31.4s, v18.4s\n"
- "ldp x27, x26, [x15, #0x40]\n"
- "ldr q10, [x27, x17]\n"
- "fmin v28.4s, v28.4s, v17.4s\n"
- "fmin v29.4s, v29.4s, v17.4s\n"
- "ldr q14, [x26, x17]\n"
+ "fmax v28.4s, v28.4s, v27.4s\n"
+ "fmax v29.4s, v29.4s, v27.4s\n"
+ "ldp x21, x20, [x15, #0x40]\n"
+ "ldr q10, [x21, x17]\n"
+ "fmin v30.4s, v30.4s, v15.4s\n"
+ "fmin v31.4s, v31.4s, v15.4s\n"
+ "ldr q14, [x20, x17]\n"
"add x17, x17, #0x10\n"
"cmp x17, x9, LSL #4\n"
- "fmin v30.4s, v30.4s, v17.4s\n"
- "fmin v31.4s, v31.4s, v17.4s\n"
+ "fmin v28.4s, v28.4s, v15.4s\n"
+ "fmin v29.4s, v29.4s, v15.4s\n"
"add x10, x10, #0x10\n"
- "str q28, [x14, x28]\n"
+ "str q30, [x14, x28]\n"
"add x16, x16, #0x1a0\n"
- "str q29, [x13, x28]\n"
- "str q30, [x12, x28]\n"
- "str q31, [x11, x28]\n"
+ "str q31, [x13, x28]\n"
+ "str q28, [x12, x28]\n"
+ "str q29, [x11, x28]\n"
"blt 1b\n"
"2:" // Channel tail
- "mov v28.16b, v16.16b\n fmla v28.4s, v0.4s, v5.4s\n"
- "mov v29.16b, v16.16b\n fmla v29.4s, v0.4s, v6.4s\n"
- "ldr x25, [x15, #0x50]\n"
- "ldr q5, [x25, x10]\n"
- "mov v30.16b, v16.16b\n fmla v30.4s, v0.4s, v7.4s\n"
- "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v8.4s\n"
- "ldr q0, [x16, #0x0]\n"
- "ldr x24, [x15, #0x58]\n"
- "fmla v28.4s, v1.4s, v6.4s\n"
- "ldr q6, [x24, x10]\n"
- "fmla v29.4s, v1.4s, v9.4s\n"
- "ldr x23, [x15, #0x60]\n"
+ "mov v31.16b, v26.16b\n fmla v31.4s, v0.4s, v5.4s\n"
+ "mov v5.16b, v26.16b\n fmla v5.4s, v0.4s, v6.4s\n"
+ "ldr x20, [x15, #0x50]\n"
+ "ldr q22, [x20, x10]\n"
+ "mov v30.16b, v26.16b\n fmla v30.4s, v0.4s, v7.4s\n"
+ "mov v29.16b, v26.16b\n fmla v29.4s, v0.4s, v8.4s\n"
+ "ldr q19, [x16, #0x0]\n"
+ "ldr x20, [x15, #0x58]\n"
+ "fmla v31.4s, v1.4s, v6.4s\n"
+ "ldr q21, [x20, x10]\n"
+ "fmla v5.4s, v1.4s, v9.4s\n"
+ "ldr x21, [x15, #0x60]\n"
"fmla v30.4s, v1.4s, v8.4s\n"
- "fmla v31.4s, v1.4s, v13.4s\n"
- "ldr q1, [x16, #0x10]\n"
- "ldr x22, [x15, #0x68]\n"
- "fmla v28.4s, v2.4s, v9.4s\n"
- "ldr q9, [x23, x10]\n"
- "fmla v29.4s, v2.4s, v11.4s\n"
- "ldr x21, [x15, #0x70]\n"
- "fmla v30.4s, v2.4s, v13.4s\n"
- "fmla v31.4s, v2.4s, v5.4s\n"
- "ldr q2, [x16, #0x20]\n"
- "ldr x20, [x15, #0x78]\n"
- "fmla v28.4s, v3.4s, v11.4s\n"
- "ldr q11, [x22, x10]\n"
- "fmla v29.4s, v3.4s, v12.4s\n"
- "ldr x27, [x15, #0x80]\n"
- "fmla v30.4s, v3.4s, v5.4s\n"
- "fmla v31.4s, v3.4s, v6.4s\n"
- "ldr q3, [x16, #0x30]\n"
- "ldr x26, [x15, #0x88]\n"
- "fmla v28.4s, v4.4s, v12.4s\n"
- "ldr q12, [x21, x10]\n"
- "fmla v29.4s, v4.4s, v9.4s\n"
- "ldr q9, [x20, x10]\n"
- "fmla v30.4s, v4.4s, v6.4s\n"
- "fmla v31.4s, v4.4s, v10.4s\n"
- "ldr q4, [x16, #0x40]\n"
- "ldr x25, [x15, #0x90]\n"
- "fmla v28.4s, v0.4s, v7.4s\n"
- "fmla v29.4s, v0.4s, v8.4s\n"
- "ldr x24, [x15, #0x98]\n"
- "ldr x23, [x15, #0xa0]\n"
- "fmla v30.4s, v0.4s, v14.4s\n"
- "fmla v31.4s, v0.4s, v11.4s\n"
- "ldr q0, [x16, #0x50]\n"
- "ldr x22, [x15, #0xa8]\n"
- "fmla v28.4s, v1.4s, v8.4s\n"
- "ldr q8, [x26, x10]\n"
"fmla v29.4s, v1.4s, v13.4s\n"
- "ldr x21, [x15, #0xb0]\n"
- "fmla v30.4s, v1.4s, v11.4s\n"
- "fmla v31.4s, v1.4s, v12.4s\n"
- "ldr q1, [x16, #0x60]\n"
- "ldr x20, [x15, #0xb8]\n"
- "fmla v28.4s, v2.4s, v13.4s\n"
- "ldr q13, [x27, x10]\n"
- "fmla v29.4s, v2.4s, v5.4s\n"
- "ldr x27, [x15, #0xc0]\n"
- "fmla v30.4s, v2.4s, v12.4s\n"
+ "ldr q18, [x16, #0x10]\n"
+ "ldr x20, [x15, #0x68]\n"
"fmla v31.4s, v2.4s, v9.4s\n"
- "ldr q2, [x16, #0x70]\n"
- "ldr x26, [x15, #0xc8]\n"
- "fmla v28.4s, v3.4s, v5.4s\n"
- "ldr q5, [x25, x10]\n"
- "fmla v29.4s, v3.4s, v6.4s\n"
- "ldr x25, [x15, #0xd0]\n"
- "fmla v30.4s, v3.4s, v9.4s\n"
- "fmla v31.4s, v3.4s, v13.4s\n"
- "ldr q3, [x16, #0x80]\n"
- "add x28, x28, #0x10\n"
- "fmla v28.4s, v4.4s, v6.4s\n"
- "ldr q6, [x24, x10]\n"
+ "ldr q16, [x21, x10]\n"
+ "fmla v5.4s, v2.4s, v11.4s\n"
+ "ldr x23, [x15, #0x70]\n"
+ "fmla v30.4s, v2.4s, v13.4s\n"
+ "fmla v29.4s, v2.4s, v22.4s\n"
+ "ldr q17, [x16, #0x20]\n"
+ "ldr x21, [x15, #0x78]\n"
+ "fmla v31.4s, v3.4s, v11.4s\n"
+ "ldr q6, [x20, x10]\n"
+ "fmla v5.4s, v3.4s, v12.4s\n"
+ "ldr x22, [x15, #0x80]\n"
+ "fmla v30.4s, v3.4s, v22.4s\n"
+ "fmla v29.4s, v3.4s, v21.4s\n"
+ "ldr q20, [x16, #0x30]\n"
+ "ldr x20, [x15, #0x88]\n"
+ "fmla v31.4s, v4.4s, v12.4s\n"
+ "ldr q2, [x23, x10]\n"
+ "fmla v5.4s, v4.4s, v16.4s\n"
+ "ldr q28, [x21, x10]\n"
+ "fmla v30.4s, v4.4s, v21.4s\n"
"fmla v29.4s, v4.4s, v10.4s\n"
- "ldr q10, [x23, x10]\n"
- "fmla v30.4s, v4.4s, v13.4s\n"
- "fmla v31.4s, v4.4s, v8.4s\n"
- "ldr q4, [x16, #0x90]\n"
- "ldr x24, [x15, #0xd8]\n"
- "fmla v28.4s, v0.4s, v14.4s\n"
- "ldr q14, [x20, x10]\n"
- "fmla v29.4s, v0.4s, v11.4s\n"
- "ldr x23, [x15, #0xe0]\n"
- "fmla v30.4s, v0.4s, v5.4s\n"
- "fmla v31.4s, v0.4s, v6.4s\n"
- "ldr q0, [x16, #0xa0]\n"
- "ldr x20, [x15, #0xf8]\n"
- "fmla v28.4s, v1.4s, v11.4s\n"
- "ldr q11, [x22, x10]\n"
- "fmla v29.4s, v1.4s, v12.4s\n"
- "ldr x22, [x15, #0xe8]\n"
- "fmla v30.4s, v1.4s, v6.4s\n"
- "fmla v31.4s, v1.4s, v10.4s\n"
- "ldr q1, [x16, #0xb0]\n"
- "fmla v28.4s, v2.4s, v12.4s\n"
- "ldr q12, [x21, x10]\n"
- "fmla v29.4s, v2.4s, v9.4s\n"
- "ldr x21, [x15, #0xf0]\n"
- "fmla v30.4s, v2.4s, v10.4s\n"
- "fmla v31.4s, v2.4s, v11.4s\n"
- "ldr q2, [x16, #0xc0]\n"
- "fmla v28.4s, v3.4s, v9.4s\n"
- "ldr q9, [x27, x10]\n"
- "fmla v29.4s, v3.4s, v13.4s\n"
- "ldr x27, [x15, #0x100]\n"
- "fmla v30.4s, v3.4s, v11.4s\n"
- "fmla v31.4s, v3.4s, v12.4s\n"
+ "ldr q16, [x16, #0x40]\n"
+ "ldr x21, [x15, #0x90]\n"
+ "fmla v31.4s, v19.4s, v7.4s\n"
+ "fmla v5.4s, v19.4s, v8.4s\n"
+ "ldr x27, [x15, #0x98]\n"
+ "ldr x26, [x15, #0xa0]\n"
+ "fmla v30.4s, v19.4s, v14.4s\n"
+ "fmla v29.4s, v19.4s, v6.4s\n"
+ "ldr q19, [x16, #0x50]\n"
+ "ldr x25, [x15, #0xa8]\n"
+ "fmla v31.4s, v18.4s, v8.4s\n"
+ "ldr q1, [x20, x10]\n"
+ "fmla v5.4s, v18.4s, v13.4s\n"
+ "ldr x24, [x15, #0xb0]\n"
+ "fmla v30.4s, v18.4s, v6.4s\n"
+ "fmla v29.4s, v18.4s, v2.4s\n"
+ "ldr q18, [x16, #0x60]\n"
+ "ldr x20, [x15, #0xb8]\n"
+ "fmla v31.4s, v17.4s, v13.4s\n"
+ "ldr q26, [x22, x10]\n"
+ "fmla v5.4s, v17.4s, v22.4s\n"
+ "ldr x23, [x15, #0xc0]\n"
+ "fmla v30.4s, v17.4s, v2.4s\n"
+ "fmla v29.4s, v17.4s, v28.4s\n"
+ "ldr q17, [x16, #0x70]\n"
+ "ldr x22, [x15, #0xc8]\n"
+ "fmla v31.4s, v20.4s, v22.4s\n"
+ "ldr q25, [x21, x10]\n"
+ "fmla v5.4s, v20.4s, v21.4s\n"
+ "ldr x21, [x15, #0xd0]\n"
+ "fmla v30.4s, v20.4s, v28.4s\n"
+ "fmla v29.4s, v20.4s, v26.4s\n"
+ "ldr q24, [x16, #0x80]\n"
+ "add x28, x28, #0x10\n"
+ "fmla v31.4s, v16.4s, v21.4s\n"
+ "ldr q23, [x27, x10]\n"
+ "fmla v5.4s, v16.4s, v10.4s\n"
+ "ldr q0, [x26, x10]\n"
+ "fmla v30.4s, v16.4s, v26.4s\n"
+ "fmla v29.4s, v16.4s, v1.4s\n"
+ "ldr q22, [x16, #0x90]\n"
+ "ldr x27, [x15, #0xd8]\n"
+ "fmla v31.4s, v19.4s, v14.4s\n"
+ "ldr q16, [x20, x10]\n"
+ "fmla v5.4s, v19.4s, v6.4s\n"
+ "ldr x20, [x15, #0xe0]\n"
+ "fmla v30.4s, v19.4s, v25.4s\n"
+ "fmla v29.4s, v19.4s, v23.4s\n"
+ "ldr q21, [x16, #0xa0]\n"
+ "ldr x26, [x15, #0xf8]\n"
+ "fmla v31.4s, v18.4s, v6.4s\n"
+ "ldr q20, [x25, x10]\n"
+ "fmla v5.4s, v18.4s, v2.4s\n"
+ "ldr x25, [x15, #0xe8]\n"
+ "fmla v30.4s, v18.4s, v23.4s\n"
+ "fmla v29.4s, v18.4s, v0.4s\n"
+ "ldr q18, [x16, #0xb0]\n"
+ "fmla v31.4s, v17.4s, v2.4s\n"
+ "ldr q19, [x24, x10]\n"
+ "fmla v5.4s, v17.4s, v28.4s\n"
+ "ldr x24, [x15, #0xf0]\n"
+ "fmla v30.4s, v17.4s, v0.4s\n"
+ "fmla v29.4s, v17.4s, v20.4s\n"
+ "ldr q17, [x16, #0xc0]\n"
+ "fmla v31.4s, v24.4s, v28.4s\n"
+ "ldr q7, [x23, x10]\n"
+ "fmla v5.4s, v24.4s, v26.4s\n"
+ "ldr x23, [x15, #0x100]\n"
+ "fmla v30.4s, v24.4s, v20.4s\n"
+ "fmla v29.4s, v24.4s, v19.4s\n"
"ldr q3, [x16, #0xd0]\n"
- "fmla v28.4s, v4.4s, v13.4s\n"
- "ldr q13, [x26, x10]\n"
- "fmla v29.4s, v4.4s, v8.4s\n"
- "ldr q8, [x23, x10]\n"
- "fmla v30.4s, v4.4s, v12.4s\n"
- "fmla v31.4s, v4.4s, v14.4s\n"
- "ldr q4, [x16, #0xe0]\n"
- "ldr x26, [x15, #0x108]\n"
- "fmla v28.4s, v0.4s, v5.4s\n"
- "ldr q5, [x25, x10]\n"
- "fmla v29.4s, v0.4s, v6.4s\n"
- "ldr x25, [x15, #0x110]\n"
- "fmla v30.4s, v0.4s, v9.4s\n"
- "fmla v31.4s, v0.4s, v13.4s\n"
- "ldr q0, [x16, #0xf0]\n"
- "fmla v28.4s, v1.4s, v6.4s\n"
- "ldr q6, [x24, x10]\n"
- "fmla v29.4s, v1.4s, v10.4s\n"
- "ldr x24, [x15, #0x118]\n"
- "fmla v30.4s, v1.4s, v13.4s\n"
- "fmla v31.4s, v1.4s, v5.4s\n"
- "ldr q1, [x16, #0x100]\n"
- "fmla v28.4s, v2.4s, v10.4s\n"
- "ldr q10, [x22, x10]\n"
- "fmla v29.4s, v2.4s, v11.4s\n"
- "fmla v30.4s, v2.4s, v5.4s\n"
- "fmla v31.4s, v2.4s, v6.4s\n"
- "ldr q2, [x16, #0x110]\n"
- "fmla v28.4s, v3.4s, v11.4s\n"
- "ldr q11, [x21, x10]\n"
- "fmla v29.4s, v3.4s, v12.4s\n"
- "fmla v30.4s, v3.4s, v6.4s\n"
- "fmla v31.4s, v3.4s, v8.4s\n"
- "ldr q3, [x16, #0x120]\n"
- "fmla v28.4s, v4.4s, v12.4s\n"
- "ldr q12, [x20, x10]\n"
- "fmla v29.4s, v4.4s, v14.4s\n"
- "fmla v30.4s, v4.4s, v8.4s\n"
- "fmla v31.4s, v4.4s, v10.4s\n"
- "ldr q4, [x16, #0x130]\n"
+ "fmla v31.4s, v22.4s, v26.4s\n"
+ "ldr q28, [x22, x10]\n"
+ "fmla v5.4s, v22.4s, v1.4s\n"
+ "ldr q13, [x20, x10]\n"
+ "fmla v30.4s, v22.4s, v19.4s\n"
+ "fmla v29.4s, v22.4s, v16.4s\n"
+ "ldr q11, [x16, #0xe0]\n"
+ "ldr x22, [x15, #0x108]\n"
+ "fmla v31.4s, v21.4s, v25.4s\n"
+ "ldr q26, [x21, x10]\n"
+ "fmla v5.4s, v21.4s, v23.4s\n"
+ "ldr x21, [x15, #0x110]\n"
+ "fmla v30.4s, v21.4s, v7.4s\n"
+ "fmla v29.4s, v21.4s, v28.4s\n"
+ "ldr q25, [x16, #0xf0]\n"
+ "fmla v31.4s, v18.4s, v23.4s\n"
+ "ldr q24, [x27, x10]\n"
+ "fmla v5.4s, v18.4s, v0.4s\n"
+ "ldr x20, [x15, #0x118]\n"
+ "fmla v30.4s, v18.4s, v28.4s\n"
+ "fmla v29.4s, v18.4s, v26.4s\n"
+ "ldr q23, [x16, #0x100]\n"
+ "fmla v31.4s, v17.4s, v0.4s\n"
+ "ldr q22, [x25, x10]\n"
+ "fmla v5.4s, v17.4s, v20.4s\n"
+ "fmla v30.4s, v17.4s, v26.4s\n"
+ "fmla v29.4s, v17.4s, v24.4s\n"
+ "ldr q21, [x16, #0x110]\n"
+ "fmla v31.4s, v3.4s, v20.4s\n"
+ "ldr q18, [x24, x10]\n"
+ "fmla v5.4s, v3.4s, v19.4s\n"
+ "fmla v30.4s, v3.4s, v24.4s\n"
+ "fmla v29.4s, v3.4s, v13.4s\n"
+ "ldr q20, [x16, #0x120]\n"
+ "fmla v31.4s, v11.4s, v19.4s\n"
+ "ldr q17, [x26, x10]\n"
+ "fmla v5.4s, v11.4s, v16.4s\n"
+ "fmla v30.4s, v11.4s, v13.4s\n"
+ "fmla v29.4s, v11.4s, v22.4s\n"
+ "ldr q19, [x16, #0x130]\n"
"add x16, x16, #0x140\n"
- "fmla v28.4s, v0.4s, v9.4s\n"
- "ldr q9, [x27, x10]\n"
- "fmla v29.4s, v0.4s, v13.4s\n"
- "fmla v30.4s, v0.4s, v11.4s\n"
- "ldr q11, [x26, x10]\n"
- "fmla v31.4s, v0.4s, v12.4s\n"
- "fmla v28.4s, v1.4s, v13.4s\n"
- "fmla v29.4s, v1.4s, v5.4s\n"
- "fmla v30.4s, v1.4s, v12.4s\n"
- "ldr q12, [x25, x10]\n"
- "fmla v31.4s, v1.4s, v9.4s\n"
- "fmla v28.4s, v2.4s, v5.4s\n"
- "fmla v29.4s, v2.4s, v6.4s\n"
- "fmla v30.4s, v2.4s, v9.4s\n"
- "ldr q9, [x24, x10]\n"
- "fmla v31.4s, v2.4s, v11.4s\n"
+ "fmla v31.4s, v25.4s, v7.4s\n"
+ "ldr q16, [x23, x10]\n"
+ "fmla v5.4s, v25.4s, v28.4s\n"
+ "fmla v30.4s, v25.4s, v18.4s\n"
+ "ldr q18, [x22, x10]\n"
+ "fmla v29.4s, v25.4s, v17.4s\n"
+ "fmla v31.4s, v23.4s, v28.4s\n"
+ "fmla v5.4s, v23.4s, v26.4s\n"
+ "fmla v30.4s, v23.4s, v17.4s\n"
+ "ldr q17, [x21, x10]\n"
+ "fmla v29.4s, v23.4s, v16.4s\n"
+ "fmla v31.4s, v21.4s, v26.4s\n"
+ "fmla v5.4s, v21.4s, v24.4s\n"
+ "fmla v30.4s, v21.4s, v16.4s\n"
+ "ldr q16, [x20, x10]\n"
+ "fmla v29.4s, v21.4s, v18.4s\n"
"add x10, x10, #0x10\n"
- "fmla v28.4s, v3.4s, v6.4s\n"
- "fmla v29.4s, v3.4s, v8.4s\n"
- "fmla v30.4s, v3.4s, v11.4s\n"
- "fmla v31.4s, v3.4s, v12.4s\n"
- "fmla v28.4s, v4.4s, v8.4s\n"
- "fmla v29.4s, v4.4s, v10.4s\n"
- "fmax v28.4s, v28.4s, v18.4s\n"
- "fmla v30.4s, v4.4s, v12.4s\n"
- "fmla v31.4s, v4.4s, v9.4s\n"
- "fmax v29.4s, v29.4s, v18.4s\n"
- "fmax v30.4s, v30.4s, v18.4s\n"
- "fmax v31.4s, v31.4s, v18.4s\n"
- "fmin v28.4s, v28.4s, v17.4s\n"
- "fmin v29.4s, v29.4s, v17.4s\n"
- "str q28, [x14, x28]\n"
- "fmin v30.4s, v30.4s, v17.4s\n"
- "fmin v31.4s, v31.4s, v17.4s\n"
- "str q29, [x13, x28]\n"
+ "fmla v31.4s, v20.4s, v24.4s\n"
+ "fmla v5.4s, v20.4s, v13.4s\n"
+ "fmla v30.4s, v20.4s, v18.4s\n"
+ "fmla v29.4s, v20.4s, v17.4s\n"
+ "fmla v31.4s, v19.4s, v13.4s\n"
+ "fmla v5.4s, v19.4s, v22.4s\n"
+ "fmax v31.4s, v31.4s, v27.4s\n"
+ "fmla v30.4s, v19.4s, v17.4s\n"
+ "fmla v29.4s, v19.4s, v16.4s\n"
+ "fmax v5.4s, v5.4s, v27.4s\n"
+ "fmax v30.4s, v30.4s, v27.4s\n"
+ "fmax v29.4s, v29.4s, v27.4s\n"
+ "fmin v31.4s, v31.4s, v15.4s\n"
+ "fmin v5.4s, v5.4s, v15.4s\n"
+ "str q31, [x14, x28]\n"
+ "fmin v30.4s, v30.4s, v15.4s\n"
+ "fmin v29.4s, v29.4s, v15.4s\n"
+ "str q5, [x13, x28]\n"
"str q30, [x12, x28]\n"
- "str q31, [x11, x28]\n"
+ "str q29, [x11, x28]\n"
"3:" // Oddments
"tst %x[n_channels], #0x3\n"
"beq 60f\n"
- "ldr q16, [x16, #0x0]\n"
+ "ldr q26, [x16, #0x0]\n"
"ldr q0, [x16, #0x10]\n"
- "mov x28, x10\n"
- "add x14, x14, x28\n"
+ "mov x20, x10\n"
+ "add x14, x14, x20\n"
"ldr q1, [x16, #0x20]\n"
"ldr q2, [x16, #0x30]\n"
- "add x13, x13, x28\n"
- "add x12, x12, x28\n"
+ "add x13, x13, x20\n"
+ "add x12, x12, x20\n"
"ldr q3, [x16, #0x40]\n"
"ldr q4, [x16, #0x50]\n"
- "add x11, x11, x28\n"
+ "add x11, x11, x20\n"
"ldr x9, [x15, #0x0]\n"
"ldr x28, [x15, #0x8]\n"
"add x9, x9, x10\n"
@@ -606,12 +606,12 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ld1 { v10.s }[0], [x21], #0x4\n"
"ld1 { v14.s }[0], [x20], #0x4\n"
"5:" // Oddments: Load inputs (0, 0), (0, 1), (1, 0), (1, 1), (0, 2), (1, 2), (0, 3), (0, 4), (1, 5), (2, 0): Bit 1: End
- "mov v28.16b, v16.16b\n fmla v28.4s, v0.4s, v5.4s\n"
- "mov v29.16b, v16.16b\n fmla v29.4s, v0.4s, v6.4s\n"
+ "mov v28.16b, v26.16b\n fmla v28.4s, v0.4s, v5.4s\n"
+ "mov v29.16b, v26.16b\n fmla v29.4s, v0.4s, v6.4s\n"
"ldr x20, [x15, #0x50]\n"
"add x20, x20, x10\n"
- "mov v30.16b, v16.16b\n fmla v30.4s, v0.4s, v7.4s\n"
- "mov v31.16b, v16.16b\n fmla v31.4s, v0.4s, v8.4s\n"
+ "mov v30.16b, v26.16b\n fmla v30.4s, v0.4s, v7.4s\n"
+ "mov v31.16b, v26.16b\n fmla v31.4s, v0.4s, v8.4s\n"
"fmla v28.4s, v1.4s, v6.4s\n"
"fmla v29.4s, v1.4s, v9.4s\n"
"fmla v30.4s, v1.4s, v8.4s\n"
@@ -1005,14 +1005,14 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"ld1 { v9.s }[0], [x20], #0x4\n"
"57:" // Oddments: Load input (5, 5): Bit 1: End
"fmla v31.4s, v4.4s, v9.4s\n"
- "fmax v28.4s, v28.4s, v18.4s\n"
- "fmax v29.4s, v29.4s, v18.4s\n"
- "fmax v30.4s, v30.4s, v18.4s\n"
- "fmax v31.4s, v31.4s, v18.4s\n"
- "fmin v28.4s, v28.4s, v17.4s\n"
- "fmin v29.4s, v29.4s, v17.4s\n"
- "fmin v30.4s, v30.4s, v17.4s\n"
- "fmin v31.4s, v31.4s, v17.4s\n"
+ "fmax v28.4s, v28.4s, v27.4s\n"
+ "fmax v29.4s, v29.4s, v27.4s\n"
+ "fmax v30.4s, v30.4s, v27.4s\n"
+ "fmax v31.4s, v31.4s, v27.4s\n"
+ "fmin v28.4s, v28.4s, v15.4s\n"
+ "fmin v29.4s, v29.4s, v15.4s\n"
+ "fmin v30.4s, v30.4s, v15.4s\n"
+ "fmin v31.4s, v31.4s, v15.4s\n"
"tbz %x[n_channels], #1, 58f\n"
"st1 { v28.d }[0], [x14], #0x8\n"
"st1 { v29.d }[0], [x13], #0x8\n"
@@ -1030,12 +1030,10 @@ void a64_fp32_nhwc_5x5_s1_output2x2_mla_depthfirst_indirect_impl(
"st1 { v30.s }[0], [x12], #0x4\n"
"st1 { v31.s }[0], [x11], #0x4\n"
"59:" // Oddments: Store: Bit 1: End
-
"60:" // 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", "v16", "v17", "v18", "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", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
);
}