aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp')
-rw-r--r--src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp972
1 files changed, 486 insertions, 486 deletions
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
index 6d2b6ee998..3adf8b0d9f 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/kernels/a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst/generic_direct.cpp
@@ -25,7 +25,7 @@
#include <cstddef>
#include <cstdint>
-#if __aarch64__
+#if defined(__aarch64__)
namespace arm_conv {
namespace depthwise {
@@ -124,9 +124,9 @@ void a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
"add x27, x10, x22, LSL #2\n"
"add x23, x5, x5\n"
"add x20, %x[params_struct], %[offsetof_args_min]\n"
- "ld1r { v15.4s }, [x20]\n"
+ "ld1r { v13.4s }, [x20]\n"
"add x20, %x[params_struct], %[offsetof_args_max]\n"
- "ld1r { v14.4s }, [x20]\n"
+ "ld1r { v15.4s }, [x20]\n"
"add x26, x9, x24, LSL #2\n"
"add x25, x28, x4\n"
"add x24, x27, x22, LSL #2\n"
@@ -134,7 +134,7 @@ void a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
"mov x21, #0x0\n"
"sub x20, XZR, x6\n"
"cbz x13, 4f\n"
- "ldr q13, [x16, #0x0]\n"
+ "ldr q14, [x16, #0x0]\n"
"ldr q0, [x16, #0x10]\n"
"cmp x6, x13, LSL #4\n"
"ldr q1, [x16, #0x20]\n"
@@ -152,499 +152,499 @@ void a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
"ldr q12, [x14, x11]\n"
"bge 3f\n"
"2:" // Tile loop: Channel loop
- "mov v21.16b, v13.16b\n fmla v21.4s, v4.4s, v9.4s\n"
- "mov v16.16b, v13.16b\n fmla v16.4s, v8.4s, v9.4s\n"
+ "mov v26.16b, v14.16b\n fmla v26.4s, v4.4s, v9.4s\n"
+ "mov v28.16b, v14.16b\n fmla v28.4s, v8.4s, v9.4s\n"
"add x6, x6, #0x10\n"
"cmp x6, x13, LSL #4\n"
- "mov v22.16b, v13.16b\n fmla v22.4s, v3.4s, v9.4s\n"
- "mov v25.16b, v13.16b\n fmla v25.4s, v1.4s, v9.4s\n"
+ "mov v16.16b, v14.16b\n fmla v16.4s, v3.4s, v9.4s\n"
+ "mov v22.16b, v14.16b\n fmla v22.4s, v1.4s, v9.4s\n"
"add x20, x20, #0x10\n"
"add x21, x21, #0x10\n"
- "mov v26.16b, v13.16b\n fmla v26.4s, v0.4s, v9.4s\n"
- "fmla v21.4s, v5.4s, v12.4s\n"
- "mov v17.16b, v13.16b\n fmla v17.4s, v7.4s, v9.4s\n"
- "mov v18.16b, v13.16b\n fmla v18.4s, v6.4s, v9.4s\n"
- "mov v20.16b, v13.16b\n fmla v20.4s, v5.4s, v9.4s\n"
- "mov v24.16b, v13.16b\n fmla v24.4s, v2.4s, v9.4s\n"
+ "mov v23.16b, v14.16b\n fmla v23.4s, v0.4s, v9.4s\n"
+ "fmla v26.4s, v5.4s, v12.4s\n"
+ "mov v25.16b, v14.16b\n fmla v25.4s, v7.4s, v9.4s\n"
+ "mov v17.16b, v14.16b\n fmla v17.4s, v6.4s, v9.4s\n"
+ "mov v31.16b, v14.16b\n fmla v31.4s, v5.4s, v9.4s\n"
+ "mov v20.16b, v14.16b\n fmla v20.4s, v2.4s, v9.4s\n"
"ldr q9, [x12, x17]\n"
- "fmla v16.4s, v0.4s, v10.4s\n"
- "ld1 { v10.4s }, [x26]\n"
- "mov v19.16b, v13.16b\n fmla v19.4s, v2.4s, v11.4s\n"
- "ldr q11, [x26, x25]\n"
- "fmla v22.4s, v4.4s, v12.4s\n"
- "fmla v25.4s, v2.4s, v12.4s\n"
- "fmla v26.4s, v1.4s, v12.4s\n"
- "mov v28.16b, v13.16b\n fmla v28.4s, v6.4s, v10.4s\n"
+ "fmla v28.4s, v0.4s, v10.4s\n"
+ "ld1 { v30.4s }, [x26]\n"
+ "mov v29.16b, v14.16b\n fmla v29.4s, v2.4s, v11.4s\n"
+ "ldr q27, [x26, x25]\n"
+ "fmla v16.4s, v4.4s, v12.4s\n"
+ "fmla v22.4s, v2.4s, v12.4s\n"
+ "fmla v23.4s, v1.4s, v12.4s\n"
+ "mov v21.16b, v14.16b\n fmla v21.4s, v6.4s, v30.4s\n"
"ldr q10, [x12, x11]\n"
- "fmla v21.4s, v7.4s, v9.4s\n"
- "fmla v17.4s, v8.4s, v12.4s\n"
- "fmla v18.4s, v7.4s, v12.4s\n"
- "fmla v19.4s, v6.4s, v12.4s\n"
- "mov v23.16b, v13.16b\n fmla v23.4s, v3.4s, v12.4s\n"
- "mov v27.16b, v13.16b\n fmla v27.4s, v0.4s, v12.4s\n"
- "ldr q12, [x7, x4]\n"
- "mov v31.16b, v13.16b\n fmla v31.4s, v8.4s, v11.4s\n"
- "ldr q11, [x7, x28]\n"
- "fmla v22.4s, v6.4s, v9.4s\n"
- "fmla v25.4s, v4.4s, v9.4s\n"
- "fmla v26.4s, v3.4s, v9.4s\n"
- "fmla v20.4s, v8.4s, v9.4s\n"
- "fmla v24.4s, v5.4s, v9.4s\n"
- "fmla v28.4s, v2.4s, v9.4s\n"
- "fmla v21.4s, v8.4s, v10.4s\n"
- "fmla v16.4s, v1.4s, v12.4s\n"
- "fmla v17.4s, v0.4s, v12.4s\n"
- "ldr q12, [x15, x25]\n"
- "fmla v18.4s, v2.4s, v11.4s\n"
- "fmla v19.4s, v1.4s, v11.4s\n"
- "ld1 { v11.4s }, [x9]\n"
- "fmla v22.4s, v7.4s, v10.4s\n"
- "fmla v23.4s, v6.4s, v10.4s\n"
- "fmla v25.4s, v5.4s, v10.4s\n"
- "fmla v26.4s, v4.4s, v10.4s\n"
- "fmla v27.4s, v3.4s, v10.4s\n"
- "fmla v31.4s, v0.4s, v10.4s\n"
- "fmla v24.4s, v6.4s, v11.4s\n"
- "fmla v28.4s, v3.4s, v11.4s\n"
- "ldr q11, [x9, x25]\n"
- "fmla v19.4s, v5.4s, v12.4s\n"
- "fmla v23.4s, v2.4s, v12.4s\n"
- "ldr q12, [x15, x11]\n"
- "fmla v27.4s, v8.4s, v11.4s\n"
- "fmla v31.4s, v5.4s, v11.4s\n"
- "mov v29.16b, v13.16b\n fmla v29.4s, v1.4s, v9.4s\n"
- "mov v30.16b, v13.16b\n fmla v30.4s, v0.4s, v9.4s\n"
+ "fmla v26.4s, v7.4s, v9.4s\n"
+ "fmla v25.4s, v8.4s, v12.4s\n"
+ "fmla v17.4s, v7.4s, v12.4s\n"
+ "fmla v29.4s, v6.4s, v12.4s\n"
+ "mov v24.16b, v14.16b\n fmla v24.4s, v3.4s, v12.4s\n"
+ "mov v19.16b, v14.16b\n fmla v19.4s, v0.4s, v12.4s\n"
+ "ldr q11, [x7, x4]\n"
+ "mov v30.16b, v14.16b\n fmla v30.4s, v8.4s, v27.4s\n"
+ "ldr q12, [x7, x28]\n"
+ "fmla v16.4s, v6.4s, v9.4s\n"
+ "fmla v22.4s, v4.4s, v9.4s\n"
+ "fmla v23.4s, v3.4s, v9.4s\n"
+ "mov v27.16b, v14.16b\n fmla v27.4s, v1.4s, v9.4s\n"
+ "mov v18.16b, v14.16b\n fmla v18.4s, v0.4s, v9.4s\n"
+ "ldr q14, [x16, #0x0]\n"
+ "fmla v31.4s, v8.4s, v9.4s\n"
+ "fmla v20.4s, v5.4s, v9.4s\n"
+ "fmla v21.4s, v2.4s, v9.4s\n"
"ld1 { v9.4s }, [x15]\n"
- "fmla v29.4s, v2.4s, v10.4s\n"
- "fmla v30.4s, v1.4s, v10.4s\n"
+ "fmla v26.4s, v8.4s, v10.4s\n"
+ "fmla v28.4s, v1.4s, v11.4s\n"
+ "fmla v25.4s, v0.4s, v11.4s\n"
+ "ldr q11, [x15, x25]\n"
+ "fmla v17.4s, v2.4s, v12.4s\n"
+ "fmla v29.4s, v1.4s, v12.4s\n"
+ "ld1 { v12.4s }, [x9]\n"
+ "fmla v16.4s, v7.4s, v10.4s\n"
+ "fmla v24.4s, v6.4s, v10.4s\n"
+ "fmla v22.4s, v5.4s, v10.4s\n"
+ "fmla v23.4s, v4.4s, v10.4s\n"
+ "fmla v19.4s, v3.4s, v10.4s\n"
+ "fmla v27.4s, v2.4s, v10.4s\n"
+ "fmla v18.4s, v1.4s, v10.4s\n"
+ "fmla v30.4s, v0.4s, v10.4s\n"
"ldr q10, [x15, x17]\n"
- "fmla v20.4s, v0.4s, v9.4s\n"
- "fmla v21.4s, v1.4s, v10.4s\n"
- "fmla v16.4s, v3.4s, v9.4s\n"
- "ldr q11, [x26, x4]\n"
- "fmla v17.4s, v4.4s, v10.4s\n"
- "fmla v18.4s, v3.4s, v10.4s\n"
- "fmla v22.4s, v0.4s, v10.4s\n"
- "fmla v20.4s, v2.4s, v10.4s\n"
- "fmla v21.4s, v2.4s, v12.4s\n"
- "fmla v16.4s, v5.4s, v10.4s\n"
+ "fmla v31.4s, v0.4s, v9.4s\n"
+ "fmla v20.4s, v6.4s, v12.4s\n"
+ "fmla v21.4s, v3.4s, v12.4s\n"
+ "ldr q12, [x9, x25]\n"
+ "fmla v26.4s, v1.4s, v10.4s\n"
+ "fmla v28.4s, v3.4s, v9.4s\n"
+ "fmla v29.4s, v5.4s, v11.4s\n"
+ "fmla v24.4s, v2.4s, v11.4s\n"
+ "ldr q11, [x15, x11]\n"
+ "fmla v25.4s, v4.4s, v10.4s\n"
+ "fmla v17.4s, v3.4s, v10.4s\n"
+ "fmla v16.4s, v0.4s, v10.4s\n"
+ "fmla v19.4s, v8.4s, v12.4s\n"
+ "fmla v30.4s, v5.4s, v12.4s\n"
+ "ldr q9, [x26, x4]\n"
+ "fmla v31.4s, v2.4s, v10.4s\n"
+ "fmla v26.4s, v2.4s, v11.4s\n"
+ "fmla v28.4s, v5.4s, v10.4s\n"
"ldr q10, [x14, x4]\n"
- "fmla v17.4s, v5.4s, v12.4s\n"
- "fmla v18.4s, v4.4s, v12.4s\n"
- "fmla v19.4s, v3.4s, v12.4s\n"
- "fmla v22.4s, v1.4s, v12.4s\n"
- "fmla v23.4s, v0.4s, v12.4s\n"
- "ldr q12, [x14, x28]\n"
- "fmla v28.4s, v7.4s, v11.4s\n"
- "fmla v29.4s, v6.4s, v11.4s\n"
- "ldr q11, [x26, x28]\n"
- "fmla v20.4s, v4.4s, v10.4s\n"
- "fmla v21.4s, v3.4s, v10.4s\n"
- "fmla v24.4s, v1.4s, v10.4s\n"
- "fmla v25.4s, v0.4s, v10.4s\n"
- "fmla v16.4s, v7.4s, v10.4s\n"
- "fmla v17.4s, v6.4s, v10.4s\n"
+ "fmla v25.4s, v5.4s, v11.4s\n"
+ "fmla v17.4s, v4.4s, v11.4s\n"
+ "fmla v29.4s, v3.4s, v11.4s\n"
+ "fmla v16.4s, v1.4s, v11.4s\n"
+ "fmla v24.4s, v0.4s, v11.4s\n"
+ "ldr q11, [x14, x28]\n"
+ "fmla v21.4s, v7.4s, v9.4s\n"
+ "fmla v27.4s, v6.4s, v9.4s\n"
+ "ldr q12, [x26, x28]\n"
+ "fmla v31.4s, v4.4s, v10.4s\n"
+ "fmla v26.4s, v3.4s, v10.4s\n"
+ "fmla v20.4s, v1.4s, v10.4s\n"
+ "fmla v22.4s, v0.4s, v10.4s\n"
+ "fmla v28.4s, v7.4s, v10.4s\n"
+ "fmla v25.4s, v6.4s, v10.4s\n"
"ldr q10, [x7, x17]\n"
- "fmla v30.4s, v8.4s, v11.4s\n"
- "fmla v31.4s, v7.4s, v11.4s\n"
- "ldr q11, [x12, x4]\n"
"fmla v18.4s, v8.4s, v12.4s\n"
- "fmla v19.4s, v7.4s, v12.4s\n"
- "fmla v22.4s, v5.4s, v12.4s\n"
- "fmla v23.4s, v4.4s, v12.4s\n"
- "fmla v26.4s, v2.4s, v12.4s\n"
- "fmla v27.4s, v1.4s, v12.4s\n"
+ "fmla v30.4s, v7.4s, v12.4s\n"
+ "ldr q9, [x12, x4]\n"
+ "fmla v17.4s, v8.4s, v11.4s\n"
+ "fmla v29.4s, v7.4s, v11.4s\n"
+ "fmla v16.4s, v5.4s, v11.4s\n"
+ "fmla v24.4s, v4.4s, v11.4s\n"
+ "fmla v23.4s, v2.4s, v11.4s\n"
+ "fmla v19.4s, v1.4s, v11.4s\n"
"ldr q12, [x7, x11]\n"
"add x7, x7, #0x10\n"
- "fmla v20.4s, v7.4s, v11.4s\n"
- "fmla v21.4s, v6.4s, v11.4s\n"
- "fmla v24.4s, v4.4s, v11.4s\n"
- "fmla v25.4s, v3.4s, v11.4s\n"
- "fmla v28.4s, v1.4s, v11.4s\n"
- "fmla v29.4s, v0.4s, v11.4s\n"
- "ldr q11, [x12, x28]\n"
- "fmla v16.4s, v2.4s, v10.4s\n"
- "fmla v17.4s, v1.4s, v10.4s\n"
- "fmla v18.4s, v0.4s, v10.4s\n"
+ "fmla v31.4s, v7.4s, v9.4s\n"
+ "fmla v26.4s, v6.4s, v9.4s\n"
+ "fmla v20.4s, v4.4s, v9.4s\n"
+ "fmla v22.4s, v3.4s, v9.4s\n"
+ "fmla v21.4s, v1.4s, v9.4s\n"
+ "fmla v27.4s, v0.4s, v9.4s\n"
+ "ldr q9, [x12, x28]\n"
+ "fmla v28.4s, v2.4s, v10.4s\n"
+ "fmla v25.4s, v1.4s, v10.4s\n"
+ "fmla v17.4s, v0.4s, v10.4s\n"
"ld1 { v10.4s }, [x14]\n"
- "fmla v30.4s, v2.4s, v11.4s\n"
- "fmla v19.4s, v0.4s, v12.4s\n"
- "fmla v20.4s, v3.4s, v10.4s\n"
- "fmla v24.4s, v0.4s, v10.4s\n"
- "fmla v22.4s, v8.4s, v11.4s\n"
- "fmla v23.4s, v7.4s, v11.4s\n"
- "fmla v26.4s, v5.4s, v11.4s\n"
- "fmla v27.4s, v4.4s, v11.4s\n"
- "fmla v31.4s, v1.4s, v11.4s\n"
+ "fmla v18.4s, v2.4s, v9.4s\n"
+ "fmla v29.4s, v0.4s, v12.4s\n"
+ "fmla v31.4s, v3.4s, v10.4s\n"
+ "fmla v20.4s, v0.4s, v10.4s\n"
+ "fmla v16.4s, v8.4s, v9.4s\n"
+ "fmla v24.4s, v7.4s, v9.4s\n"
+ "fmla v23.4s, v5.4s, v9.4s\n"
+ "fmla v19.4s, v4.4s, v9.4s\n"
+ "fmla v30.4s, v1.4s, v9.4s\n"
"ldr q11, [x9, x17]\n"
- "fmla v17.4s, v2.4s, v12.4s\n"
- "fmla v18.4s, v1.4s, v12.4s\n"
+ "fmla v25.4s, v2.4s, v12.4s\n"
+ "fmla v17.4s, v1.4s, v12.4s\n"
"ldr q12, [x14, x25]\n"
"add x14, x14, #0x10\n"
- "fmla v16.4s, v6.4s, v10.4s\n"
+ "ldr q9, [x14, x17]\n"
+ "fmla v28.4s, v6.4s, v10.4s\n"
"ld1 { v10.4s }, [x12]\n"
- "fmla v29.4s, v4.4s, v11.4s\n"
- "fmla v30.4s, v3.4s, v11.4s\n"
- "fmla v19.4s, v8.4s, v12.4s\n"
- "fmla v23.4s, v5.4s, v12.4s\n"
- "fmla v27.4s, v2.4s, v12.4s\n"
+ "fmla v27.4s, v4.4s, v11.4s\n"
+ "fmla v18.4s, v3.4s, v11.4s\n"
+ "fmla v29.4s, v8.4s, v12.4s\n"
+ "fmla v24.4s, v5.4s, v12.4s\n"
+ "fmla v19.4s, v2.4s, v12.4s\n"
"ldr q12, [x12, x25]\n"
"add x12, x12, #0x10\n"
- "fmla v20.4s, v6.4s, v10.4s\n"
- "fmla v24.4s, v3.4s, v10.4s\n"
- "fmla v28.4s, v0.4s, v10.4s\n"
+ "fmla v31.4s, v6.4s, v10.4s\n"
+ "fmla v20.4s, v3.4s, v10.4s\n"
+ "fmla v21.4s, v0.4s, v10.4s\n"
"ldr q10, [x26, x17]\n"
- "fmla v31.4s, v2.4s, v12.4s\n"
- "fmla v29.4s, v7.4s, v10.4s\n"
- "fmla v30.4s, v6.4s, v10.4s\n"
- "fmla v24.4s, v8.4s, v11.4s\n"
- "fmla v25.4s, v7.4s, v11.4s\n"
- "fmla v26.4s, v6.4s, v11.4s\n"
- "fmla v28.4s, v5.4s, v11.4s\n"
+ "fmla v30.4s, v2.4s, v12.4s\n"
+ "fmla v27.4s, v7.4s, v10.4s\n"
+ "fmla v18.4s, v6.4s, v10.4s\n"
+ "fmla v20.4s, v8.4s, v11.4s\n"
+ "fmla v22.4s, v7.4s, v11.4s\n"
+ "fmla v23.4s, v6.4s, v11.4s\n"
+ "fmla v21.4s, v5.4s, v11.4s\n"
"ldr q11, [x9, x11]\n"
- "fmla v27.4s, v5.4s, v12.4s\n"
- "fmla v29.4s, v5.4s, v11.4s\n"
- "fmla v30.4s, v4.4s, v11.4s\n"
- "fmla v31.4s, v3.4s, v11.4s\n"
- "fmla v23.4s, v8.4s, v12.4s\n"
+ "fmla v19.4s, v5.4s, v12.4s\n"
+ "fmla v27.4s, v5.4s, v11.4s\n"
+ "fmla v18.4s, v4.4s, v11.4s\n"
+ "fmla v30.4s, v3.4s, v11.4s\n"
+ "fmla v24.4s, v8.4s, v12.4s\n"
"ldr q12, [x26, x11]\n"
- "fmla v28.4s, v8.4s, v10.4s\n"
+ "fmla v21.4s, v8.4s, v10.4s\n"
"ldr q10, [x15, x4]\n"
- "fmla v25.4s, v8.4s, v11.4s\n"
- "fmla v26.4s, v7.4s, v11.4s\n"
+ "fmla v22.4s, v8.4s, v11.4s\n"
+ "fmla v23.4s, v7.4s, v11.4s\n"
"add x26, x26, #0x10\n"
- "fmla v27.4s, v6.4s, v11.4s\n"
+ "fmla v19.4s, v6.4s, v11.4s\n"
"ldr q11, [x15, x28]\n"
- "fmla v29.4s, v8.4s, v12.4s\n"
+ "fmla v27.4s, v8.4s, v12.4s\n"
"add x15, x15, #0x10\n"
- "fmla v30.4s, v7.4s, v12.4s\n"
- "fmla v31.4s, v6.4s, v12.4s\n"
+ "fmla v18.4s, v7.4s, v12.4s\n"
+ "fmla v30.4s, v6.4s, v12.4s\n"
"ldr q12, [x9, x4]\n"
- "fmla v16.4s, v4.4s, v10.4s\n"
- "fmla v17.4s, v3.4s, v10.4s\n"
- "fmax v16.4s, v16.4s, v15.4s\n"
- "fmla v20.4s, v1.4s, v10.4s\n"
- "fmla v21.4s, v0.4s, v10.4s\n"
+ "fmla v28.4s, v4.4s, v10.4s\n"
+ "fmla v25.4s, v3.4s, v10.4s\n"
+ "fmax v28.4s, v28.4s, v13.4s\n"
+ "fmla v31.4s, v1.4s, v10.4s\n"
+ "fmla v26.4s, v0.4s, v10.4s\n"
"ldr q10, [x9, x28]\n"
- "ldr q9, [x14, x17]\n"
- "fmla v18.4s, v5.4s, v11.4s\n"
- "fmla v19.4s, v4.4s, v11.4s\n"
- "fmax v17.4s, v17.4s, v15.4s\n"
+ "ldr q0, [x16, #0x10]\n"
+ "fmla v17.4s, v5.4s, v11.4s\n"
+ "fmla v29.4s, v4.4s, v11.4s\n"
+ "fmax v25.4s, v25.4s, v13.4s\n"
"add x9, x9, #0x10\n"
- "fmla v22.4s, v2.4s, v11.4s\n"
- "ldr q13, [x16, #0x0]\n"
- "fmla v23.4s, v1.4s, v11.4s\n"
+ "fmla v16.4s, v2.4s, v11.4s\n"
+ "ldr q2, [x16, #0x30]\n"
+ "fmla v24.4s, v1.4s, v11.4s\n"
"ldr q11, [x7, x25]\n"
- "ldr q0, [x16, #0x10]\n"
- "fmla v24.4s, v7.4s, v12.4s\n"
- "fmla v25.4s, v6.4s, v12.4s\n"
"ldr q1, [x16, #0x20]\n"
- "fmla v28.4s, v4.4s, v12.4s\n"
- "fmla v29.4s, v3.4s, v12.4s\n"
+ "fmla v20.4s, v7.4s, v12.4s\n"
+ "fmla v22.4s, v6.4s, v12.4s\n"
+ "ldr q6, [x16, #0x70]\n"
+ "fmla v21.4s, v4.4s, v12.4s\n"
+ "fmla v27.4s, v3.4s, v12.4s\n"
"ldr q12, [x14, x11]\n"
- "ldr q2, [x16, #0x30]\n"
- "fmla v26.4s, v8.4s, v10.4s\n"
"ldr q3, [x16, #0x40]\n"
- "fmla v27.4s, v7.4s, v10.4s\n"
- "ldr q6, [x16, #0x70]\n"
- "fmla v30.4s, v5.4s, v10.4s\n"
+ "fmla v23.4s, v8.4s, v10.4s\n"
+ "ldr q8, [x16, #0x90]\n"
+ "fmla v19.4s, v7.4s, v10.4s\n"
+ "ldr q7, [x16, #0x80]\n"
+ "fmla v18.4s, v5.4s, v10.4s\n"
"ldr q5, [x16, #0x60]\n"
- "fmla v31.4s, v4.4s, v10.4s\n"
+ "fmla v30.4s, v4.4s, v10.4s\n"
"ld1 { v10.4s }, [x7]\n"
"ldr q4, [x16, #0x50]\n"
- "fmax v18.4s, v18.4s, v15.4s\n"
- "fmax v19.4s, v19.4s, v15.4s\n"
- "fmax v20.4s, v20.4s, v15.4s\n"
- "fmax v21.4s, v21.4s, v15.4s\n"
- "fmax v22.4s, v22.4s, v15.4s\n"
- "fmax v23.4s, v23.4s, v15.4s\n"
- "fmax v24.4s, v24.4s, v15.4s\n"
- "fmax v25.4s, v25.4s, v15.4s\n"
- "fmax v26.4s, v26.4s, v15.4s\n"
- "fmax v27.4s, v27.4s, v15.4s\n"
- "fmax v28.4s, v28.4s, v15.4s\n"
- "fmax v29.4s, v29.4s, v15.4s\n"
- "fmax v30.4s, v30.4s, v15.4s\n"
- "fmax v31.4s, v31.4s, v15.4s\n"
- "fmin v16.4s, v16.4s, v14.4s\n"
- "fmin v17.4s, v17.4s, v14.4s\n"
- "st1 { v16.4s }, [x8]\n"
- "ldr q7, [x16, #0x80]\n"
- "fmin v18.4s, v18.4s, v14.4s\n"
- "fmin v19.4s, v19.4s, v14.4s\n"
- "str q17, [x8, x5]\n"
- "ldr q8, [x16, #0x90]\n"
- "fmin v20.4s, v20.4s, v14.4s\n"
- "fmin v21.4s, v21.4s, v14.4s\n"
- "str q18, [x8, x23]\n"
+ "fmax v17.4s, v17.4s, v13.4s\n"
+ "fmax v29.4s, v29.4s, v13.4s\n"
"add x16, x16, #0xa0\n"
- "fmin v22.4s, v22.4s, v14.4s\n"
- "fmin v23.4s, v23.4s, v14.4s\n"
- "str q19, [x8, x22]\n"
+ "fmax v31.4s, v31.4s, v13.4s\n"
+ "fmax v26.4s, v26.4s, v13.4s\n"
+ "fmax v16.4s, v16.4s, v13.4s\n"
+ "fmax v24.4s, v24.4s, v13.4s\n"
+ "fmax v20.4s, v20.4s, v13.4s\n"
+ "fmax v22.4s, v22.4s, v13.4s\n"
+ "fmax v23.4s, v23.4s, v13.4s\n"
+ "fmax v19.4s, v19.4s, v13.4s\n"
+ "fmax v21.4s, v21.4s, v13.4s\n"
+ "fmax v27.4s, v27.4s, v13.4s\n"
+ "fmax v18.4s, v18.4s, v13.4s\n"
+ "fmax v30.4s, v30.4s, v13.4s\n"
+ "fmin v28.4s, v28.4s, v15.4s\n"
+ "fmin v25.4s, v25.4s, v15.4s\n"
+ "st1 { v28.4s }, [x8]\n"
+ "fmin v17.4s, v17.4s, v15.4s\n"
+ "fmin v29.4s, v29.4s, v15.4s\n"
+ "str q25, [x8, x5]\n"
+ "fmin v31.4s, v31.4s, v15.4s\n"
+ "fmin v26.4s, v26.4s, v15.4s\n"
+ "str q17, [x8, x23]\n"
+ "fmin v16.4s, v16.4s, v15.4s\n"
+ "fmin v24.4s, v24.4s, v15.4s\n"
+ "str q29, [x8, x22]\n"
"add x8, x8, #0x10\n"
- "fmin v24.4s, v24.4s, v14.4s\n"
- "fmin v25.4s, v25.4s, v14.4s\n"
- "st1 { v20.4s }, [x10]\n"
- "fmin v26.4s, v26.4s, v14.4s\n"
- "fmin v27.4s, v27.4s, v14.4s\n"
- "str q21, [x10, x5]\n"
- "fmin v28.4s, v28.4s, v14.4s\n"
- "fmin v29.4s, v29.4s, v14.4s\n"
- "str q22, [x10, x23]\n"
- "fmin v30.4s, v30.4s, v14.4s\n"
- "fmin v31.4s, v31.4s, v14.4s\n"
- "str q23, [x10, x22]\n"
+ "fmin v20.4s, v20.4s, v15.4s\n"
+ "fmin v22.4s, v22.4s, v15.4s\n"
+ "st1 { v31.4s }, [x10]\n"
+ "fmin v23.4s, v23.4s, v15.4s\n"
+ "fmin v19.4s, v19.4s, v15.4s\n"
+ "str q26, [x10, x5]\n"
+ "fmin v21.4s, v21.4s, v15.4s\n"
+ "fmin v27.4s, v27.4s, v15.4s\n"
+ "str q16, [x10, x23]\n"
+ "fmin v18.4s, v18.4s, v15.4s\n"
+ "fmin v30.4s, v30.4s, v15.4s\n"
+ "str q24, [x10, x22]\n"
"add x10, x10, #0x10\n"
- "st1 { v24.4s }, [x27]\n"
- "str q25, [x27, x5]\n"
- "str q26, [x27, x23]\n"
- "str q27, [x27, x22]\n"
+ "st1 { v20.4s }, [x27]\n"
+ "str q22, [x27, x5]\n"
+ "str q23, [x27, x23]\n"
+ "str q19, [x27, x22]\n"
"add x27, x27, #0x10\n"
- "st1 { v28.4s }, [x24]\n"
- "str q29, [x24, x5]\n"
- "str q30, [x24, x23]\n"
- "str q31, [x24, x22]\n"
+ "st1 { v21.4s }, [x24]\n"
+ "str q27, [x24, x5]\n"
+ "str q18, [x24, x23]\n"
+ "str q30, [x24, x22]\n"
"add x24, x24, #0x10\n"
"blt 2b\n"
"3:" // Tile loop: Channel tail
- "mov v21.16b, v13.16b\n fmla v21.4s, v4.4s, v9.4s\n"
- "mov v16.16b, v13.16b\n fmla v16.4s, v8.4s, v9.4s\n"
- "mov v22.16b, v13.16b\n fmla v22.4s, v3.4s, v9.4s\n"
- "mov v25.16b, v13.16b\n fmla v25.4s, v1.4s, v9.4s\n"
- "mov v26.16b, v13.16b\n fmla v26.4s, v0.4s, v9.4s\n"
- "fmla v21.4s, v5.4s, v12.4s\n"
- "mov v17.16b, v13.16b\n fmla v17.4s, v7.4s, v9.4s\n"
- "mov v18.16b, v13.16b\n fmla v18.4s, v6.4s, v9.4s\n"
- "mov v20.16b, v13.16b\n fmla v20.4s, v5.4s, v9.4s\n"
- "mov v24.16b, v13.16b\n fmla v24.4s, v2.4s, v9.4s\n"
- "ldr q9, [x12, x17]\n"
- "fmla v16.4s, v0.4s, v10.4s\n"
- "ld1 { v10.4s }, [x26]\n"
- "mov v19.16b, v13.16b\n fmla v19.4s, v2.4s, v11.4s\n"
- "ldr q11, [x26, x25]\n"
- "fmla v22.4s, v4.4s, v12.4s\n"
- "fmla v25.4s, v2.4s, v12.4s\n"
- "fmla v26.4s, v1.4s, v12.4s\n"
- "mov v28.16b, v13.16b\n fmla v28.4s, v6.4s, v10.4s\n"
- "ldr q10, [x12, x11]\n"
- "fmla v21.4s, v7.4s, v9.4s\n"
+ "mov v16.16b, v14.16b\n fmla v16.4s, v4.4s, v9.4s\n"
+ "mov v23.16b, v14.16b\n fmla v23.4s, v8.4s, v9.4s\n"
+ "mov v31.16b, v14.16b\n fmla v31.4s, v3.4s, v9.4s\n"
+ "mov v30.16b, v14.16b\n fmla v30.4s, v1.4s, v9.4s\n"
+ "mov v18.16b, v14.16b\n fmla v18.4s, v0.4s, v9.4s\n"
+ "fmla v16.4s, v5.4s, v12.4s\n"
+ "mov v17.16b, v14.16b\n fmla v17.4s, v7.4s, v9.4s\n"
+ "mov v19.16b, v14.16b\n fmla v19.4s, v6.4s, v9.4s\n"
+ "mov v28.16b, v14.16b\n fmla v28.4s, v5.4s, v9.4s\n"
+ "mov v27.16b, v14.16b\n fmla v27.4s, v2.4s, v9.4s\n"
+ "ldr q24, [x12, x17]\n"
+ "fmla v23.4s, v0.4s, v10.4s\n"
+ "ld1 { v21.4s }, [x26]\n"
+ "mov v29.16b, v14.16b\n fmla v29.4s, v2.4s, v11.4s\n"
+ "ldr q20, [x26, x25]\n"
+ "fmla v31.4s, v4.4s, v12.4s\n"
+ "fmla v30.4s, v2.4s, v12.4s\n"
+ "fmla v18.4s, v1.4s, v12.4s\n"
+ "mov v26.16b, v14.16b\n fmla v26.4s, v6.4s, v21.4s\n"
+ "ldr q9, [x12, x11]\n"
+ "fmla v16.4s, v7.4s, v24.4s\n"
"fmla v17.4s, v8.4s, v12.4s\n"
- "fmla v18.4s, v7.4s, v12.4s\n"
- "fmla v19.4s, v6.4s, v12.4s\n"
- "mov v23.16b, v13.16b\n fmla v23.4s, v3.4s, v12.4s\n"
- "mov v27.16b, v13.16b\n fmla v27.4s, v0.4s, v12.4s\n"
- "ldr q12, [x7, x4]\n"
- "mov v31.16b, v13.16b\n fmla v31.4s, v8.4s, v11.4s\n"
- "ldr q11, [x7, x28]\n"
- "fmla v22.4s, v6.4s, v9.4s\n"
- "fmla v25.4s, v4.4s, v9.4s\n"
- "fmla v26.4s, v3.4s, v9.4s\n"
- "fmla v20.4s, v8.4s, v9.4s\n"
- "fmla v24.4s, v5.4s, v9.4s\n"
- "fmla v28.4s, v2.4s, v9.4s\n"
- "fmla v21.4s, v8.4s, v10.4s\n"
- "fmla v16.4s, v1.4s, v12.4s\n"
- "fmla v17.4s, v0.4s, v12.4s\n"
- "ldr q12, [x15, x25]\n"
- "fmla v18.4s, v2.4s, v11.4s\n"
- "fmla v19.4s, v1.4s, v11.4s\n"
- "ld1 { v11.4s }, [x9]\n"
- "fmla v22.4s, v7.4s, v10.4s\n"
- "fmla v23.4s, v6.4s, v10.4s\n"
- "fmla v25.4s, v5.4s, v10.4s\n"
- "fmla v26.4s, v4.4s, v10.4s\n"
- "fmla v27.4s, v3.4s, v10.4s\n"
- "fmla v31.4s, v0.4s, v10.4s\n"
- "fmla v24.4s, v6.4s, v11.4s\n"
- "fmla v28.4s, v3.4s, v11.4s\n"
- "ldr q11, [x9, x25]\n"
- "fmla v19.4s, v5.4s, v12.4s\n"
- "fmla v23.4s, v2.4s, v12.4s\n"
- "ldr q12, [x15, x11]\n"
- "fmla v27.4s, v8.4s, v11.4s\n"
- "fmla v31.4s, v5.4s, v11.4s\n"
- "mov v29.16b, v13.16b\n fmla v29.4s, v1.4s, v9.4s\n"
- "mov v30.16b, v13.16b\n fmla v30.4s, v0.4s, v9.4s\n"
- "ld1 { v9.4s }, [x15]\n"
- "fmla v29.4s, v2.4s, v10.4s\n"
- "fmla v30.4s, v1.4s, v10.4s\n"
- "ldr q10, [x15, x17]\n"
- "fmla v20.4s, v0.4s, v9.4s\n"
- "fmla v21.4s, v1.4s, v10.4s\n"
- "fmla v16.4s, v3.4s, v9.4s\n"
- "ldr q11, [x26, x4]\n"
- "fmla v17.4s, v4.4s, v10.4s\n"
- "fmla v18.4s, v3.4s, v10.4s\n"
- "fmla v22.4s, v0.4s, v10.4s\n"
- "fmla v20.4s, v2.4s, v10.4s\n"
- "fmla v21.4s, v2.4s, v12.4s\n"
- "fmla v16.4s, v5.4s, v10.4s\n"
- "ldr q10, [x14, x4]\n"
- "fmla v17.4s, v5.4s, v12.4s\n"
- "fmla v18.4s, v4.4s, v12.4s\n"
- "fmla v19.4s, v3.4s, v12.4s\n"
- "fmla v22.4s, v1.4s, v12.4s\n"
- "fmla v23.4s, v0.4s, v12.4s\n"
- "ldr q12, [x14, x28]\n"
- "fmla v28.4s, v7.4s, v11.4s\n"
- "fmla v29.4s, v6.4s, v11.4s\n"
- "ldr q11, [x26, x28]\n"
- "fmla v20.4s, v4.4s, v10.4s\n"
- "fmla v21.4s, v3.4s, v10.4s\n"
- "fmla v24.4s, v1.4s, v10.4s\n"
- "fmla v25.4s, v0.4s, v10.4s\n"
- "fmla v16.4s, v7.4s, v10.4s\n"
- "fmla v17.4s, v6.4s, v10.4s\n"
- "ldr q10, [x7, x17]\n"
- "fmla v30.4s, v8.4s, v11.4s\n"
- "fmla v31.4s, v7.4s, v11.4s\n"
- "ldr q11, [x12, x4]\n"
- "fmla v18.4s, v8.4s, v12.4s\n"
"fmla v19.4s, v7.4s, v12.4s\n"
- "fmla v22.4s, v5.4s, v12.4s\n"
- "fmla v23.4s, v4.4s, v12.4s\n"
- "fmla v26.4s, v2.4s, v12.4s\n"
- "fmla v27.4s, v1.4s, v12.4s\n"
- "ldr q12, [x7, x11]\n"
+ "fmla v29.4s, v6.4s, v12.4s\n"
+ "mov v11.16b, v14.16b\n fmla v11.4s, v3.4s, v12.4s\n"
+ "mov v10.16b, v14.16b\n fmla v10.4s, v0.4s, v12.4s\n"
+ "ldr q22, [x7, x4]\n"
+ "mov v25.16b, v14.16b\n fmla v25.4s, v8.4s, v20.4s\n"
+ "ldr q21, [x7, x28]\n"
+ "fmla v31.4s, v6.4s, v24.4s\n"
+ "fmla v30.4s, v4.4s, v24.4s\n"
+ "fmla v18.4s, v3.4s, v24.4s\n"
+ "mov v12.16b, v14.16b\n fmla v12.4s, v1.4s, v24.4s\n"
+ "fmla v14.4s, v0.4s, v24.4s\n"
+ "fmla v28.4s, v8.4s, v24.4s\n"
+ "fmla v27.4s, v5.4s, v24.4s\n"
+ "fmla v26.4s, v2.4s, v24.4s\n"
+ "ld1 { v24.4s }, [x15]\n"
+ "fmla v16.4s, v8.4s, v9.4s\n"
+ "fmla v23.4s, v1.4s, v22.4s\n"
+ "fmla v17.4s, v0.4s, v22.4s\n"
+ "ldr q22, [x15, x25]\n"
+ "fmla v19.4s, v2.4s, v21.4s\n"
+ "fmla v29.4s, v1.4s, v21.4s\n"
+ "ld1 { v20.4s }, [x9]\n"
+ "fmla v31.4s, v7.4s, v9.4s\n"
+ "fmla v11.4s, v6.4s, v9.4s\n"
+ "fmla v30.4s, v5.4s, v9.4s\n"
+ "fmla v18.4s, v4.4s, v9.4s\n"
+ "fmla v10.4s, v3.4s, v9.4s\n"
+ "fmla v12.4s, v2.4s, v9.4s\n"
+ "fmla v14.4s, v1.4s, v9.4s\n"
+ "fmla v25.4s, v0.4s, v9.4s\n"
+ "ldr q21, [x15, x17]\n"
+ "fmla v28.4s, v0.4s, v24.4s\n"
+ "fmla v27.4s, v6.4s, v20.4s\n"
+ "fmla v26.4s, v3.4s, v20.4s\n"
+ "ldr q20, [x9, x25]\n"
+ "fmla v16.4s, v1.4s, v21.4s\n"
+ "fmla v23.4s, v3.4s, v24.4s\n"
+ "fmla v29.4s, v5.4s, v22.4s\n"
+ "fmla v11.4s, v2.4s, v22.4s\n"
+ "ldr q22, [x15, x11]\n"
+ "fmla v17.4s, v4.4s, v21.4s\n"
+ "fmla v19.4s, v3.4s, v21.4s\n"
+ "fmla v31.4s, v0.4s, v21.4s\n"
+ "fmla v10.4s, v8.4s, v20.4s\n"
+ "fmla v25.4s, v5.4s, v20.4s\n"
+ "ldr q20, [x26, x4]\n"
+ "fmla v28.4s, v2.4s, v21.4s\n"
+ "fmla v16.4s, v2.4s, v22.4s\n"
+ "fmla v23.4s, v5.4s, v21.4s\n"
+ "ldr q21, [x14, x4]\n"
+ "fmla v17.4s, v5.4s, v22.4s\n"
+ "fmla v19.4s, v4.4s, v22.4s\n"
+ "fmla v29.4s, v3.4s, v22.4s\n"
+ "fmla v31.4s, v1.4s, v22.4s\n"
+ "fmla v11.4s, v0.4s, v22.4s\n"
+ "ldr q22, [x14, x28]\n"
+ "fmla v26.4s, v7.4s, v20.4s\n"
+ "fmla v12.4s, v6.4s, v20.4s\n"
+ "ldr q20, [x26, x28]\n"
+ "fmla v28.4s, v4.4s, v21.4s\n"
+ "fmla v16.4s, v3.4s, v21.4s\n"
+ "fmla v27.4s, v1.4s, v21.4s\n"
+ "fmla v30.4s, v0.4s, v21.4s\n"
+ "fmla v23.4s, v7.4s, v21.4s\n"
+ "fmla v17.4s, v6.4s, v21.4s\n"
+ "ldr q21, [x7, x17]\n"
+ "fmla v14.4s, v8.4s, v20.4s\n"
+ "fmla v25.4s, v7.4s, v20.4s\n"
+ "ldr q20, [x12, x4]\n"
+ "fmla v19.4s, v8.4s, v22.4s\n"
+ "fmla v29.4s, v7.4s, v22.4s\n"
+ "fmla v31.4s, v5.4s, v22.4s\n"
+ "fmla v11.4s, v4.4s, v22.4s\n"
+ "fmla v18.4s, v2.4s, v22.4s\n"
+ "fmla v10.4s, v1.4s, v22.4s\n"
+ "ldr q22, [x7, x11]\n"
"add x7, x7, #0x10\n"
- "fmla v20.4s, v7.4s, v11.4s\n"
- "fmla v21.4s, v6.4s, v11.4s\n"
- "fmla v24.4s, v4.4s, v11.4s\n"
- "fmla v25.4s, v3.4s, v11.4s\n"
- "fmla v28.4s, v1.4s, v11.4s\n"
- "fmla v29.4s, v0.4s, v11.4s\n"
- "ldr q11, [x12, x28]\n"
- "fmla v16.4s, v2.4s, v10.4s\n"
- "fmla v17.4s, v1.4s, v10.4s\n"
- "fmla v18.4s, v0.4s, v10.4s\n"
- "ld1 { v10.4s }, [x14]\n"
- "fmla v30.4s, v2.4s, v11.4s\n"
- "fmla v19.4s, v0.4s, v12.4s\n"
- "fmla v20.4s, v3.4s, v10.4s\n"
- "fmla v24.4s, v0.4s, v10.4s\n"
- "fmla v22.4s, v8.4s, v11.4s\n"
- "fmla v23.4s, v7.4s, v11.4s\n"
- "fmla v26.4s, v5.4s, v11.4s\n"
- "fmla v27.4s, v4.4s, v11.4s\n"
- "fmla v31.4s, v1.4s, v11.4s\n"
- "ldr q11, [x9, x17]\n"
- "fmla v17.4s, v2.4s, v12.4s\n"
- "fmla v18.4s, v1.4s, v12.4s\n"
- "ldr q12, [x14, x25]\n"
+ "fmla v28.4s, v7.4s, v20.4s\n"
+ "fmla v16.4s, v6.4s, v20.4s\n"
+ "fmla v27.4s, v4.4s, v20.4s\n"
+ "fmla v30.4s, v3.4s, v20.4s\n"
+ "fmla v26.4s, v1.4s, v20.4s\n"
+ "fmla v12.4s, v0.4s, v20.4s\n"
+ "ldr q20, [x12, x28]\n"
+ "fmla v23.4s, v2.4s, v21.4s\n"
+ "fmla v17.4s, v1.4s, v21.4s\n"
+ "fmla v19.4s, v0.4s, v21.4s\n"
+ "ld1 { v21.4s }, [x14]\n"
+ "fmla v14.4s, v2.4s, v20.4s\n"
+ "fmla v29.4s, v0.4s, v22.4s\n"
+ "fmla v28.4s, v3.4s, v21.4s\n"
+ "fmla v27.4s, v0.4s, v21.4s\n"
+ "fmla v31.4s, v8.4s, v20.4s\n"
+ "fmla v11.4s, v7.4s, v20.4s\n"
+ "fmla v18.4s, v5.4s, v20.4s\n"
+ "fmla v10.4s, v4.4s, v20.4s\n"
+ "fmla v25.4s, v1.4s, v20.4s\n"
+ "ldr q24, [x9, x17]\n"
+ "fmla v17.4s, v2.4s, v22.4s\n"
+ "fmla v19.4s, v1.4s, v22.4s\n"
+ "ldr q20, [x14, x25]\n"
"add x14, x14, #0x10\n"
- "fmla v16.4s, v6.4s, v10.4s\n"
- "ld1 { v10.4s }, [x12]\n"
- "fmla v29.4s, v4.4s, v11.4s\n"
- "fmla v30.4s, v3.4s, v11.4s\n"
- "fmla v19.4s, v8.4s, v12.4s\n"
- "fmla v23.4s, v5.4s, v12.4s\n"
- "fmla v27.4s, v2.4s, v12.4s\n"
- "ldr q12, [x12, x25]\n"
+ "fmla v23.4s, v6.4s, v21.4s\n"
+ "ld1 { v21.4s }, [x12]\n"
+ "fmla v12.4s, v4.4s, v24.4s\n"
+ "fmla v14.4s, v3.4s, v24.4s\n"
+ "fmla v29.4s, v8.4s, v20.4s\n"
+ "fmla v11.4s, v5.4s, v20.4s\n"
+ "fmla v10.4s, v2.4s, v20.4s\n"
+ "ldr q20, [x12, x25]\n"
"add x12, x12, #0x10\n"
- "fmla v20.4s, v6.4s, v10.4s\n"
- "fmla v24.4s, v3.4s, v10.4s\n"
- "fmla v28.4s, v0.4s, v10.4s\n"
- "ldr q10, [x26, x17]\n"
- "fmla v31.4s, v2.4s, v12.4s\n"
- "fmla v29.4s, v7.4s, v10.4s\n"
- "fmla v30.4s, v6.4s, v10.4s\n"
- "fmla v24.4s, v8.4s, v11.4s\n"
- "fmla v25.4s, v7.4s, v11.4s\n"
- "fmla v26.4s, v6.4s, v11.4s\n"
- "fmla v28.4s, v5.4s, v11.4s\n"
- "ldr q11, [x9, x11]\n"
- "fmla v27.4s, v5.4s, v12.4s\n"
- "fmla v29.4s, v5.4s, v11.4s\n"
- "fmla v30.4s, v4.4s, v11.4s\n"
- "fmla v31.4s, v3.4s, v11.4s\n"
- "fmla v23.4s, v8.4s, v12.4s\n"
- "ldr q12, [x26, x11]\n"
- "fmla v28.4s, v8.4s, v10.4s\n"
- "ldr q10, [x15, x4]\n"
- "fmla v25.4s, v8.4s, v11.4s\n"
- "fmla v26.4s, v7.4s, v11.4s\n"
+ "fmla v28.4s, v6.4s, v21.4s\n"
+ "fmla v27.4s, v3.4s, v21.4s\n"
+ "fmla v26.4s, v0.4s, v21.4s\n"
+ "ldr q22, [x26, x17]\n"
+ "fmla v25.4s, v2.4s, v20.4s\n"
+ "fmla v12.4s, v7.4s, v22.4s\n"
+ "fmla v14.4s, v6.4s, v22.4s\n"
+ "fmla v27.4s, v8.4s, v24.4s\n"
+ "fmla v30.4s, v7.4s, v24.4s\n"
+ "fmla v18.4s, v6.4s, v24.4s\n"
+ "fmla v26.4s, v5.4s, v24.4s\n"
+ "ldr q21, [x9, x11]\n"
+ "fmla v10.4s, v5.4s, v20.4s\n"
+ "fmla v12.4s, v5.4s, v21.4s\n"
+ "fmla v14.4s, v4.4s, v21.4s\n"
+ "fmla v25.4s, v3.4s, v21.4s\n"
+ "fmla v11.4s, v8.4s, v20.4s\n"
+ "ldr q20, [x26, x11]\n"
+ "fmla v26.4s, v8.4s, v22.4s\n"
+ "ldr q9, [x15, x4]\n"
+ "fmla v30.4s, v8.4s, v21.4s\n"
+ "fmla v18.4s, v7.4s, v21.4s\n"
"add x26, x26, #0x10\n"
- "fmla v27.4s, v6.4s, v11.4s\n"
- "ldr q11, [x15, x28]\n"
- "fmla v29.4s, v8.4s, v12.4s\n"
+ "fmla v10.4s, v6.4s, v21.4s\n"
+ "ldr q21, [x15, x28]\n"
+ "fmla v12.4s, v8.4s, v20.4s\n"
"add x15, x15, #0x10\n"
- "fmla v30.4s, v7.4s, v12.4s\n"
- "fmla v31.4s, v6.4s, v12.4s\n"
- "ldr q12, [x9, x4]\n"
- "fmla v16.4s, v4.4s, v10.4s\n"
- "fmla v17.4s, v3.4s, v10.4s\n"
- "fmax v16.4s, v16.4s, v15.4s\n"
- "fmla v20.4s, v1.4s, v10.4s\n"
- "fmla v21.4s, v0.4s, v10.4s\n"
- "ldr q10, [x9, x28]\n"
- "fmax v17.4s, v17.4s, v15.4s\n"
- "fmla v18.4s, v5.4s, v11.4s\n"
- "fmla v19.4s, v4.4s, v11.4s\n"
- "fmax v18.4s, v18.4s, v15.4s\n"
+ "fmla v14.4s, v7.4s, v20.4s\n"
+ "fmla v25.4s, v6.4s, v20.4s\n"
+ "ldr q24, [x9, x4]\n"
+ "fmla v23.4s, v4.4s, v9.4s\n"
+ "fmla v17.4s, v3.4s, v9.4s\n"
+ "fmax v23.4s, v23.4s, v13.4s\n"
+ "fmla v28.4s, v1.4s, v9.4s\n"
+ "fmla v16.4s, v0.4s, v9.4s\n"
+ "ldr q0, [x9, x28]\n"
+ "fmax v17.4s, v17.4s, v13.4s\n"
+ "fmla v19.4s, v5.4s, v21.4s\n"
+ "fmla v29.4s, v4.4s, v21.4s\n"
+ "fmax v19.4s, v19.4s, v13.4s\n"
"add x9, x9, #0x10\n"
- "fmla v22.4s, v2.4s, v11.4s\n"
- "fmla v23.4s, v1.4s, v11.4s\n"
- "fmax v19.4s, v19.4s, v15.4s\n"
- "fmla v24.4s, v7.4s, v12.4s\n"
- "fmla v25.4s, v6.4s, v12.4s\n"
- "fmax v20.4s, v20.4s, v15.4s\n"
- "fmla v28.4s, v4.4s, v12.4s\n"
- "fmla v29.4s, v3.4s, v12.4s\n"
- "fmax v21.4s, v21.4s, v15.4s\n"
- "fmla v26.4s, v8.4s, v10.4s\n"
- "fmla v27.4s, v7.4s, v10.4s\n"
- "fmax v22.4s, v22.4s, v15.4s\n"
- "fmla v30.4s, v5.4s, v10.4s\n"
- "fmla v31.4s, v4.4s, v10.4s\n"
- "fmax v23.4s, v23.4s, v15.4s\n"
- "fmax v24.4s, v24.4s, v15.4s\n"
- "fmax v25.4s, v25.4s, v15.4s\n"
- "fmax v26.4s, v26.4s, v15.4s\n"
- "fmax v27.4s, v27.4s, v15.4s\n"
- "fmax v28.4s, v28.4s, v15.4s\n"
- "fmax v29.4s, v29.4s, v15.4s\n"
- "fmax v30.4s, v30.4s, v15.4s\n"
- "fmax v31.4s, v31.4s, v15.4s\n"
- "fmin v16.4s, v16.4s, v14.4s\n"
- "fmin v17.4s, v17.4s, v14.4s\n"
- "st1 { v16.4s }, [x8]\n"
- "fmin v18.4s, v18.4s, v14.4s\n"
- "fmin v19.4s, v19.4s, v14.4s\n"
+ "fmla v31.4s, v2.4s, v21.4s\n"
+ "fmla v11.4s, v1.4s, v21.4s\n"
+ "fmax v29.4s, v29.4s, v13.4s\n"
+ "fmla v27.4s, v7.4s, v24.4s\n"
+ "fmla v30.4s, v6.4s, v24.4s\n"
+ "fmax v28.4s, v28.4s, v13.4s\n"
+ "fmla v26.4s, v4.4s, v24.4s\n"
+ "fmla v12.4s, v3.4s, v24.4s\n"
+ "fmax v16.4s, v16.4s, v13.4s\n"
+ "fmla v18.4s, v8.4s, v0.4s\n"
+ "fmla v10.4s, v7.4s, v0.4s\n"
+ "fmax v31.4s, v31.4s, v13.4s\n"
+ "fmla v14.4s, v5.4s, v0.4s\n"
+ "fmla v25.4s, v4.4s, v0.4s\n"
+ "fmax v11.4s, v11.4s, v13.4s\n"
+ "fmax v27.4s, v27.4s, v13.4s\n"
+ "fmax v30.4s, v30.4s, v13.4s\n"
+ "fmax v18.4s, v18.4s, v13.4s\n"
+ "fmax v10.4s, v10.4s, v13.4s\n"
+ "fmax v26.4s, v26.4s, v13.4s\n"
+ "fmax v12.4s, v12.4s, v13.4s\n"
+ "fmax v14.4s, v14.4s, v13.4s\n"
+ "fmax v25.4s, v25.4s, v13.4s\n"
+ "fmin v23.4s, v23.4s, v15.4s\n"
+ "fmin v17.4s, v17.4s, v15.4s\n"
+ "st1 { v23.4s }, [x8]\n"
+ "fmin v19.4s, v19.4s, v15.4s\n"
+ "fmin v29.4s, v29.4s, v15.4s\n"
"str q17, [x8, x5]\n"
- "fmin v20.4s, v20.4s, v14.4s\n"
- "fmin v21.4s, v21.4s, v14.4s\n"
- "str q18, [x8, x23]\n"
- "fmin v22.4s, v22.4s, v14.4s\n"
- "fmin v23.4s, v23.4s, v14.4s\n"
- "str q19, [x8, x22]\n"
+ "fmin v28.4s, v28.4s, v15.4s\n"
+ "fmin v16.4s, v16.4s, v15.4s\n"
+ "str q19, [x8, x23]\n"
+ "fmin v31.4s, v31.4s, v15.4s\n"
+ "fmin v11.4s, v11.4s, v15.4s\n"
+ "str q29, [x8, x22]\n"
"add x8, x8, #0x10\n"
- "fmin v24.4s, v24.4s, v14.4s\n"
- "fmin v25.4s, v25.4s, v14.4s\n"
- "st1 { v20.4s }, [x10]\n"
- "fmin v26.4s, v26.4s, v14.4s\n"
- "fmin v27.4s, v27.4s, v14.4s\n"
- "str q21, [x10, x5]\n"
- "fmin v28.4s, v28.4s, v14.4s\n"
- "fmin v29.4s, v29.4s, v14.4s\n"
- "str q22, [x10, x23]\n"
- "fmin v30.4s, v30.4s, v14.4s\n"
- "fmin v31.4s, v31.4s, v14.4s\n"
- "str q23, [x10, x22]\n"
+ "fmin v27.4s, v27.4s, v15.4s\n"
+ "fmin v30.4s, v30.4s, v15.4s\n"
+ "st1 { v28.4s }, [x10]\n"
+ "fmin v18.4s, v18.4s, v15.4s\n"
+ "fmin v10.4s, v10.4s, v15.4s\n"
+ "str q16, [x10, x5]\n"
+ "fmin v26.4s, v26.4s, v15.4s\n"
+ "fmin v12.4s, v12.4s, v15.4s\n"
+ "str q31, [x10, x23]\n"
+ "fmin v14.4s, v14.4s, v15.4s\n"
+ "fmin v25.4s, v25.4s, v15.4s\n"
+ "str q11, [x10, x22]\n"
"add x10, x10, #0x10\n"
- "st1 { v24.4s }, [x27]\n"
- "str q25, [x27, x5]\n"
- "str q26, [x27, x23]\n"
- "str q27, [x27, x22]\n"
+ "st1 { v27.4s }, [x27]\n"
+ "str q30, [x27, x5]\n"
+ "str q18, [x27, x23]\n"
+ "str q10, [x27, x22]\n"
"add x27, x27, #0x10\n"
- "st1 { v28.4s }, [x24]\n"
- "str q29, [x24, x5]\n"
- "str q30, [x24, x23]\n"
- "str q31, [x24, x22]\n"
+ "st1 { v26.4s }, [x24]\n"
+ "str q12, [x24, x5]\n"
+ "str q14, [x24, x23]\n"
+ "str q25, [x24, x22]\n"
"add x24, x24, #0x10\n"
"4:" // Tile loop: Oddments
"tst %x[n_channels], #0x3\n"
"beq 73f\n"
- "ldr q13, [x16, #0x0]\n"
+ "ldr q14, [x16, #0x0]\n"
"ldr q0, [x16, #0x10]\n"
"add x23, x14, x17\n"
"add x22, x7, XZR\n"
@@ -675,27 +675,27 @@ void a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
"ldr s11, [x21, #0x0]\n"
"ldr s12, [x20, #0x0]\n"
"6:" // Tile loop: Oddments: Load inputs: (2, 2), (0, 0), (0, 5), (2, 3): Bit 1: End
- "mov v16.16b, v13.16b\n fmla v16.4s, v8.4s, v9.4s\n"
- "mov v17.16b, v13.16b\n fmla v17.4s, v7.4s, v9.4s\n"
+ "mov v16.16b, v14.16b\n fmla v16.4s, v8.4s, v9.4s\n"
+ "mov v17.16b, v14.16b\n fmla v17.4s, v7.4s, v9.4s\n"
"add x20, x26, XZR\n"
- "mov v18.16b, v13.16b\n fmla v18.4s, v6.4s, v9.4s\n"
- "mov v21.16b, v13.16b\n fmla v21.4s, v4.4s, v9.4s\n"
- "mov v22.16b, v13.16b\n fmla v22.4s, v3.4s, v9.4s\n"
- "mov v25.16b, v13.16b\n fmla v25.4s, v1.4s, v9.4s\n"
- "mov v26.16b, v13.16b\n fmla v26.4s, v0.4s, v9.4s\n"
- "mov v19.16b, v13.16b\n fmla v19.4s, v2.4s, v11.4s\n"
- "mov v20.16b, v13.16b\n fmla v20.4s, v5.4s, v9.4s\n"
- "mov v24.16b, v13.16b\n fmla v24.4s, v2.4s, v9.4s\n"
+ "mov v18.16b, v14.16b\n fmla v18.4s, v6.4s, v9.4s\n"
+ "mov v21.16b, v14.16b\n fmla v21.4s, v4.4s, v9.4s\n"
+ "mov v22.16b, v14.16b\n fmla v22.4s, v3.4s, v9.4s\n"
+ "mov v25.16b, v14.16b\n fmla v25.4s, v1.4s, v9.4s\n"
+ "mov v26.16b, v14.16b\n fmla v26.4s, v0.4s, v9.4s\n"
+ "mov v19.16b, v14.16b\n fmla v19.4s, v2.4s, v11.4s\n"
+ "mov v20.16b, v14.16b\n fmla v20.4s, v5.4s, v9.4s\n"
+ "mov v24.16b, v14.16b\n fmla v24.4s, v2.4s, v9.4s\n"
"fmla v16.4s, v0.4s, v10.4s\n"
"fmla v17.4s, v8.4s, v12.4s\n"
"fmla v18.4s, v7.4s, v12.4s\n"
"fmla v19.4s, v6.4s, v12.4s\n"
"fmla v21.4s, v5.4s, v12.4s\n"
"fmla v22.4s, v4.4s, v12.4s\n"
- "mov v23.16b, v13.16b\n fmla v23.4s, v3.4s, v12.4s\n"
+ "mov v23.16b, v14.16b\n fmla v23.4s, v3.4s, v12.4s\n"
"fmla v25.4s, v2.4s, v12.4s\n"
"fmla v26.4s, v1.4s, v12.4s\n"
- "mov v27.16b, v13.16b\n fmla v27.4s, v0.4s, v12.4s\n"
+ "mov v27.16b, v14.16b\n fmla v27.4s, v0.4s, v12.4s\n"
"tbz %x[n_channels], #1, 7f\n"
"ldr d10, [x20], #0x8\n"
"tbz %x[n_channels], #0, 8f\n"
@@ -704,7 +704,7 @@ void a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
"7:" // Tile loop: Oddments: Load inputs: (5, 0): Bit 1: Unset
"ldr s10, [x20, #0x0]\n"
"8:" // Tile loop: Oddments: Load inputs: (5, 0): Bit 1: End
- "mov v28.16b, v13.16b\n fmla v28.4s, v6.4s, v10.4s\n"
+ "mov v28.16b, v14.16b\n fmla v28.4s, v6.4s, v10.4s\n"
"add x20, x26, x25\n"
"tbz %x[n_channels], #1, 9f\n"
"ldr d11, [x20], #0x8\n"
@@ -714,7 +714,7 @@ void a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
"9:" // Tile loop: Oddments: Load inputs: (5, 5): Bit 1: Unset
"ldr s11, [x20, #0x0]\n"
"10:" // Tile loop: Oddments: Load inputs: (5, 5): Bit 1: End
- "mov v31.16b, v13.16b\n fmla v31.4s, v8.4s, v11.4s\n"
+ "mov v31.16b, v14.16b\n fmla v31.4s, v8.4s, v11.4s\n"
"add x20, x12, x17\n"
"tbz %x[n_channels], #1, 11f\n"
"ldr d9, [x20], #0x8\n"
@@ -732,8 +732,8 @@ void a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
"fmla v25.4s, v4.4s, v9.4s\n"
"fmla v26.4s, v3.4s, v9.4s\n"
"fmla v28.4s, v2.4s, v9.4s\n"
- "mov v29.16b, v13.16b\n fmla v29.4s, v1.4s, v9.4s\n"
- "mov v30.16b, v13.16b\n fmla v30.4s, v0.4s, v9.4s\n"
+ "mov v29.16b, v14.16b\n fmla v29.4s, v1.4s, v9.4s\n"
+ "mov v30.16b, v14.16b\n fmla v30.4s, v0.4s, v9.4s\n"
"tbz %x[n_channels], #1, 13f\n"
"ldr d12, [x20], #0x8\n"
"tbz %x[n_channels], #0, 14f\n"
@@ -1105,40 +1105,40 @@ void a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
"70:" // Tile loop: Oddments: Load inputs: (4, 4): Bit 1: End
"fmla v26.4s, v8.4s, v10.4s\n"
"fmla v27.4s, v7.4s, v10.4s\n"
- "fmax v16.4s, v16.4s, v15.4s\n"
+ "fmax v16.4s, v16.4s, v13.4s\n"
"fmla v30.4s, v5.4s, v10.4s\n"
"fmla v31.4s, v4.4s, v10.4s\n"
- "fmax v17.4s, v17.4s, v15.4s\n"
- "fmax v18.4s, v18.4s, v15.4s\n"
- "fmax v19.4s, v19.4s, v15.4s\n"
- "fmax v20.4s, v20.4s, v15.4s\n"
- "fmax v21.4s, v21.4s, v15.4s\n"
- "fmax v22.4s, v22.4s, v15.4s\n"
- "fmax v23.4s, v23.4s, v15.4s\n"
- "fmax v24.4s, v24.4s, v15.4s\n"
- "fmax v25.4s, v25.4s, v15.4s\n"
- "fmax v26.4s, v26.4s, v15.4s\n"
- "fmax v27.4s, v27.4s, v15.4s\n"
- "fmax v28.4s, v28.4s, v15.4s\n"
- "fmax v29.4s, v29.4s, v15.4s\n"
- "fmax v30.4s, v30.4s, v15.4s\n"
- "fmax v31.4s, v31.4s, v15.4s\n"
- "fmin v16.4s, v16.4s, v14.4s\n"
- "fmin v17.4s, v17.4s, v14.4s\n"
- "fmin v18.4s, v18.4s, v14.4s\n"
- "fmin v19.4s, v19.4s, v14.4s\n"
- "fmin v20.4s, v20.4s, v14.4s\n"
- "fmin v21.4s, v21.4s, v14.4s\n"
- "fmin v22.4s, v22.4s, v14.4s\n"
- "fmin v23.4s, v23.4s, v14.4s\n"
- "fmin v24.4s, v24.4s, v14.4s\n"
- "fmin v25.4s, v25.4s, v14.4s\n"
- "fmin v26.4s, v26.4s, v14.4s\n"
- "fmin v27.4s, v27.4s, v14.4s\n"
- "fmin v28.4s, v28.4s, v14.4s\n"
- "fmin v29.4s, v29.4s, v14.4s\n"
- "fmin v30.4s, v30.4s, v14.4s\n"
- "fmin v31.4s, v31.4s, v14.4s\n"
+ "fmax v17.4s, v17.4s, v13.4s\n"
+ "fmax v18.4s, v18.4s, v13.4s\n"
+ "fmax v19.4s, v19.4s, v13.4s\n"
+ "fmax v20.4s, v20.4s, v13.4s\n"
+ "fmax v21.4s, v21.4s, v13.4s\n"
+ "fmax v22.4s, v22.4s, v13.4s\n"
+ "fmax v23.4s, v23.4s, v13.4s\n"
+ "fmax v24.4s, v24.4s, v13.4s\n"
+ "fmax v25.4s, v25.4s, v13.4s\n"
+ "fmax v26.4s, v26.4s, v13.4s\n"
+ "fmax v27.4s, v27.4s, v13.4s\n"
+ "fmax v28.4s, v28.4s, v13.4s\n"
+ "fmax v29.4s, v29.4s, v13.4s\n"
+ "fmax v30.4s, v30.4s, v13.4s\n"
+ "fmax v31.4s, v31.4s, v13.4s\n"
+ "fmin v16.4s, v16.4s, v15.4s\n"
+ "fmin v17.4s, v17.4s, v15.4s\n"
+ "fmin v18.4s, v18.4s, v15.4s\n"
+ "fmin v19.4s, v19.4s, v15.4s\n"
+ "fmin v20.4s, v20.4s, v15.4s\n"
+ "fmin v21.4s, v21.4s, v15.4s\n"
+ "fmin v22.4s, v22.4s, v15.4s\n"
+ "fmin v23.4s, v23.4s, v15.4s\n"
+ "fmin v24.4s, v24.4s, v15.4s\n"
+ "fmin v25.4s, v25.4s, v15.4s\n"
+ "fmin v26.4s, v26.4s, v15.4s\n"
+ "fmin v27.4s, v27.4s, v15.4s\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, 71f\n"
"mov x23, x8\n"
"mov x22, x10\n"
@@ -1229,4 +1229,4 @@ void a64_fp32_nhwc_3x3_s1_output4x4_mla_depthfirst_direct_impl(
} // namespace depthwise
} // namespace arm_conv
-#endif // __aarch64__
+#endif // defined(__aarch64__)