aboutsummaryrefslogtreecommitdiff
path: root/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_2x4_fp32bf16.hpp
diff options
context:
space:
mode:
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_2x4_fp32bf16.hpp')
-rw-r--r--src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_2x4_fp32bf16.hpp1268
1 files changed, 655 insertions, 613 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_2x4_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_2x4_fp32bf16.hpp
index 1cb7bc4445..3e0ab6d955 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_2x4_fp32bf16.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_24_2x4_fp32bf16.hpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021, 2023 Arm Limited.
+ * Copyright (c) 2021, 2023-2024 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -40,634 +40,675 @@ void a64_transpose_interleave_24_2x4_fp32bf16(bfloat16 *out, const float *in, si
__asm__ __volatile__(
"cmp %x[height], #0x8\n"
- "blt 10f\n"
+ "blt 11f\n"
"1:" // Main row loop: Head
"mov x9, %x[in]\n"
- "add x28, x9, %x[in_stride]\n"
- "add x27, x28, %x[in_stride]\n"
- "add x26, x27, %x[in_stride]\n"
+ "mov x28, %x[width]\n"
+ "mov x27, %x[out]\n"
+ "sub %x[height], %x[height], #0x8\n"
+ "add x26, x9, %x[in_stride]\n"
"add x25, x26, %x[in_stride]\n"
- "mov x24, %x[width]\n"
- "add x23, x25, %x[in_stride]\n"
+ "add x24, x25, %x[in_stride]\n"
+ "cmp x28, #0x18\n"
+ "add x23, x24, %x[in_stride]\n"
"add x22, x23, %x[in_stride]\n"
- "add x20, x22, %x[in_stride]\n"
- "cmp x24, #0x18\n"
+ "add x21, x22, %x[in_stride]\n"
+ "add x20, x21, %x[in_stride]\n"
"add %x[in], x20, %x[in_stride]\n"
- "mov x21, %x[out]\n"
- "sub %x[height], %x[height], #0x8\n"
"blt 3f\n"
"2:" // Main row loop: Column loop
- "ldr q15, [x9], #0x10\n"
- "ldr q1, [x28], #0x10\n"
- "sub x24, x24, #0x18\n"
- "cmp x24, #0x18\n"
- "ldr q0, [x27], #0x10\n"
- "ldr q27, [x26], #0x10\n"
- "zip1 v18.4s, v15.4s, v0.4s\n"
- "zip1 v20.4s, v1.4s, v27.4s\n"
- "ldr q13, [x25], #0x10\n"
- "ldr q14, [x23], #0x10\n"
- "zip2 v16.4s, v15.4s, v0.4s\n"
- "zip2 v3.4s, v1.4s, v27.4s\n"
- "ldr q12, [x22], #0x10\n"
- "ldr q11, [x20], #0x10\n"
- "zip1 v4.4s, v13.4s, v12.4s\n"
- "zip1 v28.4s, v14.4s, v11.4s\n"
- "ldr q5, [x9], #0x10\n"
- "ldr q30, [x28], #0x10\n"
- "zip2 v23.4s, v13.4s, v12.4s\n"
- "zip2 v19.4s, v14.4s, v11.4s\n"
- "ldr q25, [x27], #0x10\n"
- "ldr q11, [x26], #0x10\n"
- "zip1 v21.4s, v5.4s, v25.4s\n"
- "zip1 v14.4s, v30.4s, v11.4s\n"
- "ldr q6, [x25], #0x10\n"
- "ldr q27, [x23], #0x10\n"
- "zip2 v29.4s, v5.4s, v25.4s\n"
- "zip2 v17.4s, v30.4s, v11.4s\n"
- "ldr q2, [x22], #0x10\n"
- "ldr q10, [x20], #0x10\n"
- "zip1 v11.4s, v6.4s, v2.4s\n"
- "zip1 v1.4s, v27.4s, v10.4s\n"
- "ldr q8, [x9], #0x10\n"
- "ldr q5, [x28], #0x10\n"
- "zip2 v24.4s, v6.4s, v2.4s\n"
- "zip2 v0.4s, v27.4s, v10.4s\n"
- "ldr q6, [x27], #0x10\n"
- "ldr q31, [x26], #0x10\n"
- "zip1 v12.4s, v8.4s, v6.4s\n"
- "zip1 v10.4s, v5.4s, v31.4s\n"
- "ldr q30, [x25], #0x10\n"
- "ldr q2, [x23], #0x10\n"
- "zip2 v9.4s, v8.4s, v6.4s\n"
- "zip2 v13.4s, v5.4s, v31.4s\n"
- "ldr q7, [x22], #0x10\n"
- "ldr q8, [x20], #0x10\n"
- "zip1 v27.4s, v30.4s, v7.4s\n"
- "zip1 v31.4s, v2.4s, v8.4s\n"
- "ldr q5, [x9], #0x10\n"
- "ldr q26, [x28], #0x10\n"
- "zip2 v22.4s, v30.4s, v7.4s\n"
- "zip2 v8.4s, v2.4s, v8.4s\n"
- "ldr q2, [x27], #0x10\n"
- "ldr q6, [x26], #0x10\n"
- "zip1 v25.4s, v5.4s, v2.4s\n"
- "zip1 v15.4s, v26.4s, v6.4s\n"
- "ldr q7, [x25], #0x10\n"
- "ldr q30, [x23], #0x10\n"
- "zip2 v5.4s, v5.4s, v2.4s\n"
- "zip2 v26.4s, v26.4s, v6.4s\n"
- "ldr q2, [x22], #0x10\n"
- "zip1 v6.4s, v7.4s, v2.4s\n"
- "zip2 v7.4s, v7.4s, v2.4s\n"
- "zip1 v2.4s, v18.4s, v20.4s\n"
- ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
- "zip2 v20.4s, v18.4s, v20.4s\n"
- "ldr q18, [x20], #0x10\n"
- ".inst 0x4ea16a82 // bfcvtn2 v2.8h, v20.4s\n"
- "zip1 v20.4s, v30.4s, v18.4s\n"
- "zip2 v18.4s, v30.4s, v18.4s\n"
- "zip1 v30.4s, v16.4s, v3.4s\n"
- ".inst 0x0ea16bde // bfcvtn v30.4h, v30.4s\n"
- "zip2 v3.4s, v16.4s, v3.4s\n"
+ "ldr q28, [x9], #0x10\n"
+ "ldr q17, [x26], #0x10\n"
+ "sub x28, x28, #0x18\n"
+ "ldr q19, [x25], #0x10\n"
+ "ldr q15, [x24], #0x10\n"
+ "cmp x28, #0x18\n"
+ "ldr q20, [x23], #0x10\n"
+ "ldr q30, [x22], #0x10\n"
+ "ldr q9, [x21], #0x10\n"
+ "ldr q14, [x20], #0x10\n"
"ldr q16, [x9], #0x10\n"
- ".inst 0x4ea1687e // bfcvtn2 v30.8h, v3.4s\n"
- "zip1 v3.4s, v21.4s, v14.4s\n"
- ".inst 0x0ea16863 // bfcvtn v3.4h, v3.4s\n"
- "zip2 v21.4s, v21.4s, v14.4s\n"
- "ldr q14, [x28], #0x10\n"
- ".inst 0x4ea16aa3 // bfcvtn2 v3.8h, v21.4s\n"
- "zip1 v21.4s, v29.4s, v17.4s\n"
- ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n"
- "zip2 v29.4s, v29.4s, v17.4s\n"
- "ldr q17, [x27], #0x10\n"
- ".inst 0x4ea16bb5 // bfcvtn2 v21.8h, v29.4s\n"
- "zip1 v29.4s, v16.4s, v17.4s\n"
- "zip2 v16.4s, v16.4s, v17.4s\n"
- "zip1 v17.4s, v12.4s, v10.4s\n"
- ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
- "zip2 v10.4s, v12.4s, v10.4s\n"
- "ldr q12, [x26], #0x10\n"
- ".inst 0x4ea16951 // bfcvtn2 v17.8h, v10.4s\n"
- "zip1 v10.4s, v14.4s, v12.4s\n"
- "zip2 v14.4s, v14.4s, v12.4s\n"
- "zip1 v12.4s, v9.4s, v13.4s\n"
- ".inst 0x0ea1698c // bfcvtn v12.4h, v12.4s\n"
- "zip2 v13.4s, v9.4s, v13.4s\n"
- "ldr q9, [x25], #0x10\n"
- ".inst 0x4ea169ac // bfcvtn2 v12.8h, v13.4s\n"
- "zip1 v13.4s, v25.4s, v15.4s\n"
+ "ldr q22, [x26], #0x10\n"
+ "zip1 v27.4s, v28.4s, v19.4s\n"
+ "zip1 v7.4s, v17.4s, v15.4s\n"
+ "ldr q12, [x25], #0x10\n"
+ "ldr q24, [x24], #0x10\n"
+ "zip2 v6.4s, v28.4s, v19.4s\n"
+ "zip2 v26.4s, v17.4s, v15.4s\n"
+ "ldr q18, [x23], #0x10\n"
+ "ldr q23, [x22], #0x10\n"
+ "zip1 v25.4s, v20.4s, v9.4s\n"
+ "zip1 v11.4s, v30.4s, v14.4s\n"
+ "ldr q2, [x21], #0x10\n"
+ "ldr q0, [x20], #0x10\n"
+ "zip2 v28.4s, v20.4s, v9.4s\n"
+ "zip2 v31.4s, v30.4s, v14.4s\n"
+ "ldr q13, [x9], #0x10\n"
+ "ldr q10, [x26], #0x10\n"
+ "zip1 v14.4s, v16.4s, v12.4s\n"
+ "zip1 v15.4s, v22.4s, v24.4s\n"
+ "ldr q1, [x25], #0x10\n"
+ "ldr q30, [x24], #0x10\n"
+ "zip2 v3.4s, v16.4s, v12.4s\n"
+ "zip2 v20.4s, v22.4s, v24.4s\n"
+ "ldr q17, [x23], #0x10\n"
+ "ldr q9, [x22], #0x10\n"
+ "zip1 v21.4s, v18.4s, v2.4s\n"
+ "zip1 v22.4s, v23.4s, v0.4s\n"
+ "ldr q5, [x21], #0x10\n"
+ "ldr q4, [x20], #0x10\n"
+ "zip2 v16.4s, v18.4s, v2.4s\n"
+ "zip2 v12.4s, v23.4s, v0.4s\n"
+ "ldr q23, [x9], #0x10\n"
+ "ldr q24, [x26], #0x10\n"
+ "zip1 v2.4s, v13.4s, v1.4s\n"
+ "zip1 v18.4s, v10.4s, v30.4s\n"
+ "ldr q29, [x25], #0x10\n"
+ "ldr q0, [x24], #0x10\n"
+ "zip2 v1.4s, v13.4s, v1.4s\n"
+ "zip2 v30.4s, v10.4s, v30.4s\n"
+ "ldr q13, [x23], #0x10\n"
+ "ldr q10, [x22], #0x10\n"
+ "zip1 v8.4s, v17.4s, v5.4s\n"
+ "zip1 v19.4s, v9.4s, v4.4s\n"
+ "zip2 v5.4s, v17.4s, v5.4s\n"
+ "ldr q17, [x21], #0x10\n"
+ "zip2 v9.4s, v9.4s, v4.4s\n"
+ "zip1 v4.4s, v23.4s, v29.4s\n"
+ "zip2 v23.4s, v23.4s, v29.4s\n"
+ "zip1 v29.4s, v24.4s, v0.4s\n"
+ "zip2 v24.4s, v24.4s, v0.4s\n"
+ "zip1 v0.4s, v13.4s, v17.4s\n"
+ "zip2 v17.4s, v13.4s, v17.4s\n"
+ "zip1 v13.4s, v27.4s, v7.4s\n"
+ "zip2 v7.4s, v27.4s, v7.4s\n"
+ "ldr q27, [x20], #0x10\n"
".inst 0x0ea169ad // bfcvtn v13.4h, v13.4s\n"
- "zip2 v25.4s, v25.4s, v15.4s\n"
- "ldr q15, [x23], #0x10\n"
- ".inst 0x4ea16b2d // bfcvtn2 v13.8h, v25.4s\n"
- "zip1 v25.4s, v5.4s, v26.4s\n"
- ".inst 0x0ea16b39 // bfcvtn v25.4h, v25.4s\n"
- "zip2 v5.4s, v5.4s, v26.4s\n"
- "ldr q26, [x22], #0x10\n"
- ".inst 0x4ea168b9 // bfcvtn2 v25.8h, v5.4s\n"
- "zip1 v5.4s, v9.4s, v26.4s\n"
- "zip2 v9.4s, v9.4s, v26.4s\n"
- "zip1 v26.4s, v29.4s, v10.4s\n"
- ".inst 0x0ea16b5a // bfcvtn v26.4h, v26.4s\n"
- "zip2 v10.4s, v29.4s, v10.4s\n"
- "ldr q29, [x20], #0x10\n"
- ".inst 0x4ea1695a // bfcvtn2 v26.8h, v10.4s\n"
- "zip1 v10.4s, v15.4s, v29.4s\n"
- "zip2 v15.4s, v15.4s, v29.4s\n"
- "zip1 v29.4s, v16.4s, v14.4s\n"
- ".inst 0x0ea16bbd // bfcvtn v29.4h, v29.4s\n"
- "zip2 v14.4s, v16.4s, v14.4s\n"
- "ldr q16, [x9], #0x10\n"
- ".inst 0x4ea169dd // bfcvtn2 v29.8h, v14.4s\n"
- "zip1 v14.4s, v4.4s, v28.4s\n"
+ ".inst 0x4ea168ed // bfcvtn2 v13.8h, v7.4s\n"
+ "zip1 v7.4s, v6.4s, v26.4s\n"
+ "zip2 v26.4s, v6.4s, v26.4s\n"
+ "ldr q6, [x9], #0x10\n"
+ ".inst 0x0ea168e7 // bfcvtn v7.4h, v7.4s\n"
+ ".inst 0x4ea16b47 // bfcvtn2 v7.8h, v26.4s\n"
+ "zip1 v26.4s, v10.4s, v27.4s\n"
+ "zip2 v27.4s, v10.4s, v27.4s\n"
+ "zip1 v10.4s, v14.4s, v15.4s\n"
+ "zip2 v14.4s, v14.4s, v15.4s\n"
+ "ldr q15, [x26], #0x10\n"
+ ".inst 0x0ea1694a // bfcvtn v10.4h, v10.4s\n"
+ ".inst 0x4ea169ca // bfcvtn2 v10.8h, v14.4s\n"
+ "zip1 v14.4s, v3.4s, v20.4s\n"
+ "zip2 v20.4s, v3.4s, v20.4s\n"
+ "ldr q3, [x25], #0x10\n"
".inst 0x0ea169ce // bfcvtn v14.4h, v14.4s\n"
- "zip2 v4.4s, v4.4s, v28.4s\n"
- "ldr q28, [x28], #0x10\n"
- ".inst 0x4ea1688e // bfcvtn2 v14.8h, v4.4s\n"
- "zip1 v4.4s, v23.4s, v19.4s\n"
+ ".inst 0x4ea16a8e // bfcvtn2 v14.8h, v20.4s\n"
+ "zip1 v20.4s, v2.4s, v18.4s\n"
+ "zip2 v18.4s, v2.4s, v18.4s\n"
+ "ldr q2, [x24], #0x10\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ ".inst 0x4ea16a54 // bfcvtn2 v20.8h, v18.4s\n"
+ "zip1 v18.4s, v6.4s, v3.4s\n"
+ "zip2 v3.4s, v6.4s, v3.4s\n"
+ "zip1 v6.4s, v15.4s, v2.4s\n"
+ "zip2 v15.4s, v15.4s, v2.4s\n"
+ "zip1 v2.4s, v1.4s, v30.4s\n"
+ "zip2 v30.4s, v1.4s, v30.4s\n"
+ "ldr q1, [x23], #0x10\n"
+ ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n"
+ ".inst 0x4ea16bc2 // bfcvtn2 v2.8h, v30.4s\n"
+ "zip1 v30.4s, v4.4s, v29.4s\n"
+ "zip2 v4.4s, v4.4s, v29.4s\n"
+ "ldr q29, [x22], #0x10\n"
+ ".inst 0x0ea16bde // bfcvtn v30.4h, v30.4s\n"
+ ".inst 0x4ea1689e // bfcvtn2 v30.8h, v4.4s\n"
+ "zip1 v4.4s, v23.4s, v24.4s\n"
+ "zip2 v24.4s, v23.4s, v24.4s\n"
+ "ldr q23, [x21], #0x10\n"
".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\n"
- "zip2 v19.4s, v23.4s, v19.4s\n"
- "ldr q23, [x27], #0x10\n"
- ".inst 0x4ea16a64 // bfcvtn2 v4.8h, v19.4s\n"
- "zip1 v19.4s, v16.4s, v23.4s\n"
- "zip2 v16.4s, v16.4s, v23.4s\n"
- "zip1 v23.4s, v11.4s, v1.4s\n"
- ".inst 0x0ea16af7 // bfcvtn v23.4h, v23.4s\n"
- "zip2 v1.4s, v11.4s, v1.4s\n"
- "ldr q11, [x26], #0x10\n"
- ".inst 0x4ea16837 // bfcvtn2 v23.8h, v1.4s\n"
- "zip1 v1.4s, v28.4s, v11.4s\n"
- "zip2 v28.4s, v28.4s, v11.4s\n"
- "zip1 v11.4s, v19.4s, v1.4s\n"
+ ".inst 0x4ea16b04 // bfcvtn2 v4.8h, v24.4s\n"
+ "zip1 v24.4s, v18.4s, v6.4s\n"
+ "zip2 v18.4s, v18.4s, v6.4s\n"
+ "ldr q6, [x20], #0x10\n"
+ ".inst 0x0ea16b18 // bfcvtn v24.4h, v24.4s\n"
+ ".inst 0x4ea16a58 // bfcvtn2 v24.8h, v18.4s\n"
+ "zip1 v18.4s, v1.4s, v23.4s\n"
+ "zip2 v1.4s, v1.4s, v23.4s\n"
+ "zip1 v23.4s, v29.4s, v6.4s\n"
+ "zip2 v29.4s, v29.4s, v6.4s\n"
+ "zip1 v6.4s, v3.4s, v15.4s\n"
+ "zip2 v3.4s, v3.4s, v15.4s\n"
+ "ldr q15, [x9], #0x10\n"
+ ".inst 0x0ea168c6 // bfcvtn v6.4h, v6.4s\n"
+ ".inst 0x4ea16866 // bfcvtn2 v6.8h, v3.4s\n"
+ "zip1 v3.4s, v25.4s, v11.4s\n"
+ "zip2 v11.4s, v25.4s, v11.4s\n"
+ "ldr q25, [x26], #0x10\n"
+ ".inst 0x0ea16863 // bfcvtn v3.4h, v3.4s\n"
+ ".inst 0x4ea16963 // bfcvtn2 v3.8h, v11.4s\n"
+ "zip1 v11.4s, v28.4s, v31.4s\n"
+ "zip2 v28.4s, v28.4s, v31.4s\n"
+ "ldr q31, [x25], #0x10\n"
".inst 0x0ea1696b // bfcvtn v11.4h, v11.4s\n"
- "zip2 v19.4s, v19.4s, v1.4s\n"
- "ldr q1, [x25], #0x10\n"
- ".inst 0x4ea16a6b // bfcvtn2 v11.8h, v19.4s\n"
- "zip1 v19.4s, v16.4s, v28.4s\n"
- ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
- "zip2 v16.4s, v16.4s, v28.4s\n"
- "ldr q28, [x23], #0x10\n"
- ".inst 0x4ea16a13 // bfcvtn2 v19.8h, v16.4s\n"
- "zip1 v16.4s, v24.4s, v0.4s\n"
+ ".inst 0x4ea16b8b // bfcvtn2 v11.8h, v28.4s\n"
+ "zip1 v28.4s, v21.4s, v22.4s\n"
+ "zip2 v21.4s, v21.4s, v22.4s\n"
+ "ldr q22, [x24], #0x10\n"
+ ".inst 0x0ea16b9c // bfcvtn v28.4h, v28.4s\n"
+ ".inst 0x4ea16abc // bfcvtn2 v28.8h, v21.4s\n"
+ "zip1 v21.4s, v15.4s, v31.4s\n"
+ "zip2 v31.4s, v15.4s, v31.4s\n"
+ "zip1 v15.4s, v25.4s, v22.4s\n"
+ "zip2 v22.4s, v25.4s, v22.4s\n"
+ "zip1 v25.4s, v16.4s, v12.4s\n"
+ "zip2 v16.4s, v16.4s, v12.4s\n"
+ "ldr q12, [x23], #0x10\n"
+ ".inst 0x0ea16b39 // bfcvtn v25.4h, v25.4s\n"
+ ".inst 0x4ea16a19 // bfcvtn2 v25.8h, v16.4s\n"
+ "zip1 v16.4s, v21.4s, v15.4s\n"
+ "zip2 v21.4s, v21.4s, v15.4s\n"
+ "ldr q15, [x22], #0x10\n"
".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
- "zip2 v24.4s, v24.4s, v0.4s\n"
- "ldr q0, [x22], #0x10\n"
- ".inst 0x4ea16b10 // bfcvtn2 v16.8h, v24.4s\n"
- "ldr q24, [x20], #0x10\n"
- "str q2, [x21, #0x0]\n"
- "zip1 v2.4s, v1.4s, v0.4s\n"
- "zip2 v0.4s, v1.4s, v0.4s\n"
- "zip1 v1.4s, v28.4s, v24.4s\n"
- "zip2 v28.4s, v28.4s, v24.4s\n"
- "str q30, [x21, #0x10]\n"
- "zip1 v24.4s, v27.4s, v31.4s\n"
- "zip1 v30.4s, v22.4s, v8.4s\n"
- "str q3, [x21, #0x20]\n"
- "zip1 v3.4s, v6.4s, v20.4s\n"
- ".inst 0x0ea16b18 // bfcvtn v24.4h, v24.4s\n"
- "str q21, [x21, #0x30]\n"
- "zip1 v21.4s, v7.4s, v18.4s\n"
- "zip2 v31.4s, v27.4s, v31.4s\n"
- "str q17, [x21, #0x40]\n"
- "zip1 v17.4s, v5.4s, v10.4s\n"
- "zip1 v27.4s, v9.4s, v15.4s\n"
- "str q12, [x21, #0x50]\n"
- "zip1 v12.4s, v2.4s, v1.4s\n"
- ".inst 0x0ea16bde // bfcvtn v30.4h, v30.4s\n"
- "str q13, [x21, #0x60]\n"
- "zip1 v13.4s, v0.4s, v28.4s\n"
- "zip2 v22.4s, v22.4s, v8.4s\n"
- "str q25, [x21, #0x70]\n"
- ".inst 0x0ea16879 // bfcvtn v25.4h, v3.4s\n"
- "zip2 v8.4s, v6.4s, v20.4s\n"
- "str q26, [x21, #0x80]\n"
- ".inst 0x0ea16aa3 // bfcvtn v3.4h, v21.4s\n"
- "zip2 v18.4s, v7.4s, v18.4s\n"
- "str q29, [x21, #0x90]\n"
- ".inst 0x0ea16a27 // bfcvtn v7.4h, v17.4s\n"
- "zip2 v21.4s, v5.4s, v10.4s\n"
- "str q11, [x21, #0xa0]\n"
- ".inst 0x0ea16b65 // bfcvtn v5.4h, v27.4s\n"
- "zip2 v15.4s, v9.4s, v15.4s\n"
- "str q19, [x21, #0xb0]\n"
- ".inst 0x0ea16991 // bfcvtn v17.4h, v12.4s\n"
- "zip2 v20.4s, v2.4s, v1.4s\n"
- "str q14, [x21, #0xc0]\n"
- ".inst 0x0ea169bb // bfcvtn v27.4h, v13.4s\n"
- "zip2 v29.4s, v0.4s, v28.4s\n"
- "str q4, [x21, #0xd0]\n"
- ".inst 0x4ea16bf8 // bfcvtn2 v24.8h, v31.4s\n"
- ".inst 0x4ea16ade // bfcvtn2 v30.8h, v22.4s\n"
- "str q23, [x21, #0xe0]\n"
- ".inst 0x4ea16919 // bfcvtn2 v25.8h, v8.4s\n"
- ".inst 0x4ea16a43 // bfcvtn2 v3.8h, v18.4s\n"
- "str q16, [x21, #0xf0]\n"
- ".inst 0x4ea16aa7 // bfcvtn2 v7.8h, v21.4s\n"
- ".inst 0x4ea169e5 // bfcvtn2 v5.8h, v15.4s\n"
- "str q24, [x21, #0x100]\n"
- ".inst 0x4ea16a91 // bfcvtn2 v17.8h, v20.4s\n"
- ".inst 0x4ea16bbb // bfcvtn2 v27.8h, v29.4s\n"
- "str q30, [x21, #0x110]\n"
- "str q25, [x21, #0x120]\n"
- "str q3, [x21, #0x130]\n"
- "str q7, [x21, #0x140]\n"
- "str q5, [x21, #0x150]\n"
- "str q17, [x21, #0x160]\n"
- "str q27, [x21, #0x170]\n"
- "add x21, x21, %x[out_stride]\n"
+ ".inst 0x4ea16ab0 // bfcvtn2 v16.8h, v21.4s\n"
+ "zip1 v21.4s, v31.4s, v22.4s\n"
+ "zip2 v22.4s, v31.4s, v22.4s\n"
+ "ldr q31, [x21], #0x10\n"
+ ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n"
+ ".inst 0x4ea16ad5 // bfcvtn2 v21.8h, v22.4s\n"
+ "ldr q22, [x20], #0x10\n"
+ "str q13, [x27, #0x0]\n"
+ "zip1 v13.4s, v8.4s, v19.4s\n"
+ "zip2 v19.4s, v8.4s, v19.4s\n"
+ "str q7, [x27, #0x10]\n"
+ "zip1 v8.4s, v12.4s, v31.4s\n"
+ "zip2 v12.4s, v12.4s, v31.4s\n"
+ "str q10, [x27, #0x20]\n"
+ "zip1 v7.4s, v15.4s, v22.4s\n"
+ "zip2 v15.4s, v15.4s, v22.4s\n"
+ "str q14, [x27, #0x30]\n"
+ "zip1 v14.4s, v5.4s, v9.4s\n"
+ "zip1 v22.4s, v0.4s, v26.4s\n"
+ "str q20, [x27, #0x40]\n"
+ "zip1 v31.4s, v17.4s, v27.4s\n"
+ "zip1 v10.4s, v18.4s, v23.4s\n"
+ "str q2, [x27, #0x50]\n"
+ "zip1 v20.4s, v1.4s, v29.4s\n"
+ "zip1 v2.4s, v8.4s, v7.4s\n"
+ "str q30, [x27, #0x60]\n"
+ "zip1 v30.4s, v12.4s, v15.4s\n"
+ ".inst 0x0ea169ad // bfcvtn v13.4h, v13.4s\n"
+ "str q4, [x27, #0x70]\n"
+ ".inst 0x0ea169c4 // bfcvtn v4.4h, v14.4s\n"
+ "zip2 v9.4s, v5.4s, v9.4s\n"
+ "str q24, [x27, #0x80]\n"
+ ".inst 0x0ea16ad8 // bfcvtn v24.4h, v22.4s\n"
+ "zip2 v26.4s, v0.4s, v26.4s\n"
+ "str q6, [x27, #0x90]\n"
+ ".inst 0x0ea16bee // bfcvtn v14.4h, v31.4s\n"
+ "zip2 v5.4s, v17.4s, v27.4s\n"
+ "str q16, [x27, #0xa0]\n"
+ ".inst 0x0ea1695b // bfcvtn v27.4h, v10.4s\n"
+ "zip2 v17.4s, v18.4s, v23.4s\n"
+ "str q21, [x27, #0xb0]\n"
+ ".inst 0x0ea16a9f // bfcvtn v31.4h, v20.4s\n"
+ "zip2 v23.4s, v1.4s, v29.4s\n"
+ "str q3, [x27, #0xc0]\n"
+ ".inst 0x0ea16852 // bfcvtn v18.4h, v2.4s\n"
+ "zip2 v20.4s, v8.4s, v7.4s\n"
+ "str q11, [x27, #0xd0]\n"
+ ".inst 0x0ea16bd0 // bfcvtn v16.4h, v30.4s\n"
+ "zip2 v15.4s, v12.4s, v15.4s\n"
+ "str q28, [x27, #0xe0]\n"
+ ".inst 0x4ea16a6d // bfcvtn2 v13.8h, v19.4s\n"
+ ".inst 0x4ea16924 // bfcvtn2 v4.8h, v9.4s\n"
+ "str q25, [x27, #0xf0]\n"
+ ".inst 0x4ea16b58 // bfcvtn2 v24.8h, v26.4s\n"
+ ".inst 0x4ea168ae // bfcvtn2 v14.8h, v5.4s\n"
+ ".inst 0x4ea16a3b // bfcvtn2 v27.8h, v17.4s\n"
+ ".inst 0x4ea16aff // bfcvtn2 v31.8h, v23.4s\n"
+ ".inst 0x4ea16a92 // bfcvtn2 v18.8h, v20.4s\n"
+ ".inst 0x4ea169f0 // bfcvtn2 v16.8h, v15.4s\n"
+ "str q13, [x27, #0x100]\n"
+ "str q4, [x27, #0x110]\n"
+ "str q24, [x27, #0x120]\n"
+ "str q14, [x27, #0x130]\n"
+ "str q27, [x27, #0x140]\n"
+ "str q31, [x27, #0x150]\n"
+ "str q18, [x27, #0x160]\n"
+ "str q16, [x27, #0x170]\n"
+ "add x27, x27, %x[out_stride]\n"
"bge 2b\n"
"3:" // Main row loop: Column loop skip
- "cmp x24, #0x10\n"
+ "cbz x28, 10f\n"
+ "cmp x28, #0x10\n"
+ "movi v6.16b, #0x0\n"
+ "str q6, [x27, #0x0]\n"
+ "str q6, [x27, #0x10]\n"
+ "str q6, [x27, #0x20]\n"
+ "str q6, [x27, #0x30]\n"
+ "str q6, [x27, #0x40]\n"
+ "str q6, [x27, #0x50]\n"
+ "str q6, [x27, #0x60]\n"
+ "str q6, [x27, #0x70]\n"
+ "str q6, [x27, #0x80]\n"
+ "str q6, [x27, #0x90]\n"
+ "str q6, [x27, #0xa0]\n"
+ "str q6, [x27, #0xb0]\n"
+ "str q6, [x27, #0xc0]\n"
+ "str q6, [x27, #0xd0]\n"
+ "str q6, [x27, #0xe0]\n"
+ "str q6, [x27, #0xf0]\n"
+ "str q6, [x27, #0x100]\n"
+ "str q6, [x27, #0x110]\n"
+ "str q6, [x27, #0x120]\n"
+ "str q6, [x27, #0x130]\n"
+ "str q6, [x27, #0x140]\n"
+ "str q6, [x27, #0x150]\n"
+ "str q6, [x27, #0x160]\n"
+ "str q6, [x27, #0x170]\n"
"blt 5f\n"
"4:" // Main row loop: width 16 loop: loop
- "ldr q9, [x9], #0x10\n"
- "ldr q18, [x28], #0x10\n"
- "sub x24, x24, #0x10\n"
- "cmp x24, #0x10\n"
- "ldr q15, [x27], #0x10\n"
- "ldr q17, [x26], #0x10\n"
- "zip1 v14.4s, v9.4s, v15.4s\n"
- "zip1 v11.4s, v18.4s, v17.4s\n"
- "ldr q7, [x25], #0x10\n"
- "ldr q16, [x23], #0x10\n"
- "zip2 v12.4s, v9.4s, v15.4s\n"
- "zip2 v6.4s, v18.4s, v17.4s\n"
- "ldr q15, [x22], #0x10\n"
+ "ldr q19, [x9], #0x10\n"
+ "ldr q24, [x26], #0x10\n"
+ "sub x28, x28, #0x10\n"
+ "ldr q18, [x25], #0x10\n"
+ "ldr q21, [x24], #0x10\n"
+ "cmp x28, #0x10\n"
+ "ldr q15, [x23], #0x10\n"
+ "ldr q20, [x22], #0x10\n"
+ "ldr q30, [x21], #0x10\n"
+ "ldr q16, [x20], #0x10\n"
+ "ldr q6, [x9], #0x10\n"
+ "ldr q10, [x26], #0x10\n"
+ "zip1 v4.4s, v19.4s, v18.4s\n"
+ "zip1 v17.4s, v24.4s, v21.4s\n"
+ "ldr q13, [x25], #0x10\n"
+ "ldr q31, [x24], #0x10\n"
+ "zip2 v29.4s, v19.4s, v18.4s\n"
+ "zip2 v26.4s, v24.4s, v21.4s\n"
+ "ldr q21, [x23], #0x10\n"
+ "ldr q25, [x22], #0x10\n"
+ "zip1 v18.4s, v15.4s, v30.4s\n"
+ "zip1 v7.4s, v20.4s, v16.4s\n"
+ "ldr q19, [x21], #0x10\n"
"ldr q3, [x20], #0x10\n"
- "zip1 v30.4s, v7.4s, v15.4s\n"
- "zip1 v20.4s, v16.4s, v3.4s\n"
- "ldr q17, [x9], #0x10\n"
- "ldr q9, [x28], #0x10\n"
- "zip2 v1.4s, v7.4s, v15.4s\n"
- "zip2 v24.4s, v16.4s, v3.4s\n"
- "ldr q10, [x27], #0x10\n"
- "ldr q16, [x26], #0x10\n"
- "zip1 v0.4s, v17.4s, v10.4s\n"
- "zip1 v8.4s, v9.4s, v16.4s\n"
- "ldr q7, [x25], #0x10\n"
- "ldr q2, [x23], #0x10\n"
- "zip2 v17.4s, v17.4s, v10.4s\n"
- "zip2 v3.4s, v9.4s, v16.4s\n"
- "ldr q9, [x22], #0x10\n"
- "ldr q10, [x20], #0x10\n"
- "zip1 v25.4s, v7.4s, v9.4s\n"
- "zip1 v23.4s, v2.4s, v10.4s\n"
- "ldr q31, [x9], #0x10\n"
- "ldr q21, [x28], #0x10\n"
- "zip2 v16.4s, v7.4s, v9.4s\n"
- "zip2 v27.4s, v2.4s, v10.4s\n"
- "ldr q26, [x27], #0x10\n"
- "ldr q19, [x26], #0x10\n"
- "zip1 v2.4s, v31.4s, v26.4s\n"
- "zip1 v7.4s, v21.4s, v19.4s\n"
- "ldr q29, [x25], #0x10\n"
- "ldr q13, [x23], #0x10\n"
- "zip2 v31.4s, v31.4s, v26.4s\n"
- "zip2 v19.4s, v21.4s, v19.4s\n"
- "ldr q4, [x22], #0x10\n"
- "ldr q18, [x20], #0x10\n"
- "zip1 v26.4s, v29.4s, v4.4s\n"
- "zip1 v15.4s, v13.4s, v18.4s\n"
- "ldr q9, [x9], #0x10\n"
- "ldr q22, [x28], #0x10\n"
- "zip2 v4.4s, v29.4s, v4.4s\n"
- "zip2 v18.4s, v13.4s, v18.4s\n"
- "ldr q29, [x27], #0x10\n"
+ "zip2 v24.4s, v15.4s, v30.4s\n"
+ "zip2 v15.4s, v20.4s, v16.4s\n"
+ "ldr q27, [x9], #0x10\n"
+ "ldr q9, [x26], #0x10\n"
+ "zip1 v2.4s, v6.4s, v13.4s\n"
+ "zip1 v30.4s, v10.4s, v31.4s\n"
+ "ldr q0, [x25], #0x10\n"
+ "ldr q11, [x24], #0x10\n"
+ "zip2 v28.4s, v6.4s, v13.4s\n"
+ "zip2 v22.4s, v10.4s, v31.4s\n"
+ "ldr q8, [x23], #0x10\n"
+ "ldr q1, [x22], #0x10\n"
+ "zip1 v23.4s, v21.4s, v19.4s\n"
+ "zip1 v6.4s, v25.4s, v3.4s\n"
+ "ldr q31, [x21], #0x10\n"
+ "ldr q5, [x20], #0x10\n"
+ "zip2 v20.4s, v21.4s, v19.4s\n"
+ "zip2 v21.4s, v25.4s, v3.4s\n"
+ "ldr q16, [x9], #0x10\n"
"ldr q10, [x26], #0x10\n"
- "zip1 v21.4s, v9.4s, v29.4s\n"
- "zip1 v5.4s, v22.4s, v10.4s\n"
- "ldr q28, [x25], #0x10\n"
- "ldr q13, [x23], #0x10\n"
- "zip2 v29.4s, v9.4s, v29.4s\n"
- "zip2 v9.4s, v22.4s, v10.4s\n"
- "ldr q22, [x22], #0x10\n"
- "zip1 v10.4s, v28.4s, v22.4s\n"
- "zip2 v28.4s, v28.4s, v22.4s\n"
- "zip1 v22.4s, v14.4s, v11.4s\n"
- ".inst 0x0ea16ad6 // bfcvtn v22.4h, v22.4s\n"
- "zip2 v11.4s, v14.4s, v11.4s\n"
- "ldr q14, [x20], #0x10\n"
- ".inst 0x4ea16976 // bfcvtn2 v22.8h, v11.4s\n"
- "str q22, [x21, #0x0]\n"
- "zip1 v22.4s, v13.4s, v14.4s\n"
- "zip2 v14.4s, v13.4s, v14.4s\n"
- "zip1 v13.4s, v12.4s, v6.4s\n"
- "zip1 v11.4s, v0.4s, v8.4s\n"
- ".inst 0x0ea169ad // bfcvtn v13.4h, v13.4s\n"
- "zip2 v12.4s, v12.4s, v6.4s\n"
- "zip1 v6.4s, v17.4s, v3.4s\n"
- ".inst 0x0ea1696b // bfcvtn v11.4h, v11.4s\n"
- "zip2 v0.4s, v0.4s, v8.4s\n"
- "zip1 v8.4s, v2.4s, v7.4s\n"
- ".inst 0x0ea168c6 // bfcvtn v6.4h, v6.4s\n"
- "zip2 v3.4s, v17.4s, v3.4s\n"
- "zip1 v17.4s, v31.4s, v19.4s\n"
+ "zip1 v14.4s, v27.4s, v0.4s\n"
+ "zip1 v3.4s, v9.4s, v11.4s\n"
+ "ldr q19, [x25], #0x10\n"
+ "ldr q13, [x24], #0x10\n"
+ "zip2 v25.4s, v27.4s, v0.4s\n"
+ "zip2 v12.4s, v9.4s, v11.4s\n"
+ "ldr q11, [x23], #0x10\n"
+ "ldr q0, [x22], #0x10\n"
+ "zip1 v27.4s, v8.4s, v31.4s\n"
+ "zip1 v9.4s, v1.4s, v5.4s\n"
+ "zip2 v31.4s, v8.4s, v31.4s\n"
+ "ldr q8, [x21], #0x10\n"
+ "zip2 v5.4s, v1.4s, v5.4s\n"
+ "zip1 v1.4s, v16.4s, v19.4s\n"
+ "zip2 v19.4s, v16.4s, v19.4s\n"
+ "zip1 v16.4s, v10.4s, v13.4s\n"
+ "zip2 v13.4s, v10.4s, v13.4s\n"
+ "zip1 v10.4s, v11.4s, v8.4s\n"
+ "zip2 v11.4s, v11.4s, v8.4s\n"
+ "zip1 v8.4s, v4.4s, v17.4s\n"
+ "zip2 v17.4s, v4.4s, v17.4s\n"
+ "ldr q4, [x20], #0x10\n"
".inst 0x0ea16908 // bfcvtn v8.4h, v8.4s\n"
- "zip2 v2.4s, v2.4s, v7.4s\n"
- "zip1 v7.4s, v21.4s, v5.4s\n"
+ ".inst 0x4ea16a28 // bfcvtn2 v8.8h, v17.4s\n"
+ "zip1 v17.4s, v29.4s, v26.4s\n"
+ "zip2 v29.4s, v29.4s, v26.4s\n"
+ "zip1 v26.4s, v0.4s, v4.4s\n"
+ "zip2 v0.4s, v0.4s, v4.4s\n"
+ "zip1 v4.4s, v2.4s, v30.4s\n"
+ "zip2 v2.4s, v2.4s, v30.4s\n"
+ "str q8, [x27, #0x0]\n"
+ "zip1 v8.4s, v28.4s, v22.4s\n"
+ "zip1 v30.4s, v14.4s, v3.4s\n"
".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
- "zip2 v31.4s, v31.4s, v19.4s\n"
- "zip1 v19.4s, v29.4s, v9.4s\n"
- ".inst 0x0ea168e7 // bfcvtn v7.4h, v7.4s\n"
- "zip2 v21.4s, v21.4s, v5.4s\n"
- "zip1 v5.4s, v30.4s, v20.4s\n"
- ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
- "zip2 v29.4s, v29.4s, v9.4s\n"
- "zip1 v9.4s, v1.4s, v24.4s\n"
- ".inst 0x0ea168a5 // bfcvtn v5.4h, v5.4s\n"
- "zip2 v20.4s, v30.4s, v20.4s\n"
- "zip1 v30.4s, v25.4s, v23.4s\n"
- ".inst 0x0ea16929 // bfcvtn v9.4h, v9.4s\n"
- "zip2 v1.4s, v1.4s, v24.4s\n"
- "zip1 v24.4s, v16.4s, v27.4s\n"
+ "zip2 v22.4s, v28.4s, v22.4s\n"
+ "zip1 v28.4s, v25.4s, v12.4s\n"
+ ".inst 0x0ea16884 // bfcvtn v4.4h, v4.4s\n"
+ ".inst 0x0ea16908 // bfcvtn v8.4h, v8.4s\n"
".inst 0x0ea16bde // bfcvtn v30.4h, v30.4s\n"
- "zip2 v23.4s, v25.4s, v23.4s\n"
- "zip1 v25.4s, v26.4s, v15.4s\n"
+ "zip2 v14.4s, v14.4s, v3.4s\n"
+ "zip1 v3.4s, v1.4s, v16.4s\n"
+ ".inst 0x0ea16b9c // bfcvtn v28.4h, v28.4s\n"
+ "zip2 v25.4s, v25.4s, v12.4s\n"
+ "zip1 v12.4s, v19.4s, v13.4s\n"
+ "zip2 v16.4s, v1.4s, v16.4s\n"
+ "zip1 v1.4s, v18.4s, v7.4s\n"
+ ".inst 0x0ea16863 // bfcvtn v3.4h, v3.4s\n"
+ "zip2 v13.4s, v19.4s, v13.4s\n"
+ "zip1 v19.4s, v24.4s, v15.4s\n"
+ ".inst 0x0ea1698c // bfcvtn v12.4h, v12.4s\n"
+ "zip2 v7.4s, v18.4s, v7.4s\n"
+ "zip1 v18.4s, v23.4s, v6.4s\n"
+ ".inst 0x0ea16821 // bfcvtn v1.4h, v1.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ "zip2 v15.4s, v24.4s, v15.4s\n"
+ "zip1 v24.4s, v20.4s, v21.4s\n"
+ "zip2 v23.4s, v23.4s, v6.4s\n"
+ "zip1 v6.4s, v27.4s, v9.4s\n"
+ ".inst 0x0ea16a52 // bfcvtn v18.4h, v18.4s\n"
+ "zip2 v21.4s, v20.4s, v21.4s\n"
+ "zip1 v20.4s, v31.4s, v5.4s\n"
".inst 0x0ea16b18 // bfcvtn v24.4h, v24.4s\n"
- "zip2 v27.4s, v16.4s, v27.4s\n"
- "zip1 v16.4s, v4.4s, v18.4s\n"
- ".inst 0x0ea16b39 // bfcvtn v25.4h, v25.4s\n"
- "zip2 v15.4s, v26.4s, v15.4s\n"
- "zip1 v26.4s, v10.4s, v22.4s\n"
- ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
- "zip2 v18.4s, v4.4s, v18.4s\n"
- "zip1 v4.4s, v28.4s, v14.4s\n"
- ".inst 0x0ea16b5a // bfcvtn v26.4h, v26.4s\n"
- "zip2 v10.4s, v10.4s, v22.4s\n"
- ".inst 0x0ea16896 // bfcvtn v22.4h, v4.4s\n"
- "zip2 v4.4s, v28.4s, v14.4s\n"
- ".inst 0x4ea1698d // bfcvtn2 v13.8h, v12.4s\n"
- "str q13, [x21, #0x10]\n"
- ".inst 0x4ea1680b // bfcvtn2 v11.8h, v0.4s\n"
- ".inst 0x4ea16866 // bfcvtn2 v6.8h, v3.4s\n"
- "str q11, [x21, #0x20]\n"
- ".inst 0x4ea16848 // bfcvtn2 v8.8h, v2.4s\n"
- ".inst 0x4ea16bf1 // bfcvtn2 v17.8h, v31.4s\n"
- "str q6, [x21, #0x30]\n"
- ".inst 0x4ea16aa7 // bfcvtn2 v7.8h, v21.4s\n"
- ".inst 0x4ea16bb3 // bfcvtn2 v19.8h, v29.4s\n"
- "str q8, [x21, #0x40]\n"
- ".inst 0x4ea16a85 // bfcvtn2 v5.8h, v20.4s\n"
- ".inst 0x4ea16829 // bfcvtn2 v9.8h, v1.4s\n"
- "str q17, [x21, #0x50]\n"
- ".inst 0x4ea16afe // bfcvtn2 v30.8h, v23.4s\n"
- ".inst 0x4ea16b78 // bfcvtn2 v24.8h, v27.4s\n"
- "str q7, [x21, #0x60]\n"
- ".inst 0x4ea169f9 // bfcvtn2 v25.8h, v15.4s\n"
- ".inst 0x4ea16a50 // bfcvtn2 v16.8h, v18.4s\n"
- "str q19, [x21, #0x70]\n"
- ".inst 0x4ea1695a // bfcvtn2 v26.8h, v10.4s\n"
- ".inst 0x4ea16896 // bfcvtn2 v22.8h, v4.4s\n"
- "str q5, [x21, #0xc0]\n"
- "str q9, [x21, #0xd0]\n"
- "str q30, [x21, #0xe0]\n"
- "str q24, [x21, #0xf0]\n"
- "str q25, [x21, #0x100]\n"
- "str q16, [x21, #0x110]\n"
- "str q26, [x21, #0x120]\n"
- "str q22, [x21, #0x130]\n"
- "add x21, x21, #0x80\n"
+ "zip2 v27.4s, v27.4s, v9.4s\n"
+ "zip1 v9.4s, v10.4s, v26.4s\n"
+ ".inst 0x0ea168c6 // bfcvtn v6.4h, v6.4s\n"
+ ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n"
+ "zip2 v31.4s, v31.4s, v5.4s\n"
+ "zip1 v5.4s, v11.4s, v0.4s\n"
+ "zip2 v26.4s, v10.4s, v26.4s\n"
+ ".inst 0x0ea1692a // bfcvtn v10.4h, v9.4s\n"
+ "zip2 v0.4s, v11.4s, v0.4s\n"
+ ".inst 0x4ea16bb1 // bfcvtn2 v17.8h, v29.4s\n"
+ ".inst 0x4ea16844 // bfcvtn2 v4.8h, v2.4s\n"
+ ".inst 0x0ea168a9 // bfcvtn v9.4h, v5.4s\n"
+ ".inst 0x4ea16ac8 // bfcvtn2 v8.8h, v22.4s\n"
+ ".inst 0x4ea169de // bfcvtn2 v30.8h, v14.4s\n"
+ ".inst 0x4ea16b3c // bfcvtn2 v28.8h, v25.4s\n"
+ ".inst 0x4ea16a03 // bfcvtn2 v3.8h, v16.4s\n"
+ ".inst 0x4ea169ac // bfcvtn2 v12.8h, v13.4s\n"
+ "str q17, [x27, #0x10]\n"
+ ".inst 0x4ea168e1 // bfcvtn2 v1.8h, v7.4s\n"
+ ".inst 0x4ea169f3 // bfcvtn2 v19.8h, v15.4s\n"
+ "str q4, [x27, #0x20]\n"
+ ".inst 0x4ea16af2 // bfcvtn2 v18.8h, v23.4s\n"
+ ".inst 0x4ea16ab8 // bfcvtn2 v24.8h, v21.4s\n"
+ "str q8, [x27, #0x30]\n"
+ ".inst 0x4ea16b66 // bfcvtn2 v6.8h, v27.4s\n"
+ ".inst 0x4ea16bf4 // bfcvtn2 v20.8h, v31.4s\n"
+ "str q30, [x27, #0x40]\n"
+ ".inst 0x4ea16b4a // bfcvtn2 v10.8h, v26.4s\n"
+ ".inst 0x4ea16809 // bfcvtn2 v9.8h, v0.4s\n"
+ "str q28, [x27, #0x50]\n"
+ "str q3, [x27, #0x60]\n"
+ "str q12, [x27, #0x70]\n"
+ "str q1, [x27, #0xc0]\n"
+ "str q19, [x27, #0xd0]\n"
+ "str q18, [x27, #0xe0]\n"
+ "str q24, [x27, #0xf0]\n"
+ "str q6, [x27, #0x100]\n"
+ "str q20, [x27, #0x110]\n"
+ "str q10, [x27, #0x120]\n"
+ "str q9, [x27, #0x130]\n"
+ "add x27, x27, #0x80\n"
"bge 4b\n"
"5:" // Main row loop: width 16 loop: skip
- "cmp x24, #0x4\n"
+ "cmp x28, #0x4\n"
"blt 7f\n"
"6:" // Main row loop: width 4 loop: loop
- "ldr q23, [x9], #0x10\n"
- "ldr q20, [x28], #0x10\n"
- "sub x24, x24, #0x4\n"
- "cmp x24, #0x4\n"
- "ldr q17, [x27], #0x10\n"
- "ldr q16, [x26], #0x10\n"
- "zip1 v22.4s, v23.4s, v17.4s\n"
- "zip1 v21.4s, v20.4s, v16.4s\n"
- "ldr q19, [x25], #0x10\n"
- "ldr q18, [x23], #0x10\n"
- "zip2 v28.4s, v23.4s, v17.4s\n"
- "zip2 v20.4s, v20.4s, v16.4s\n"
- "ldr q17, [x22], #0x10\n"
- "ldr q16, [x20], #0x10\n"
- "zip1 v27.4s, v19.4s, v17.4s\n"
- "zip1 v26.4s, v18.4s, v16.4s\n"
- "zip2 v25.4s, v19.4s, v17.4s\n"
- "zip2 v24.4s, v18.4s, v16.4s\n"
- "zip1 v19.4s, v22.4s, v21.4s\n"
- "zip1 v18.4s, v28.4s, v20.4s\n"
+ "ldr q25, [x9], #0x10\n"
+ "ldr q24, [x26], #0x10\n"
+ "sub x28, x28, #0x4\n"
+ "ldr q21, [x25], #0x10\n"
+ "ldr q20, [x24], #0x10\n"
+ "cmp x28, #0x4\n"
+ "ldr q23, [x23], #0x10\n"
+ "ldr q19, [x22], #0x10\n"
+ "ldr q18, [x21], #0x10\n"
+ "ldr q17, [x20], #0x10\n"
+ "zip1 v22.4s, v25.4s, v21.4s\n"
+ "zip1 v16.4s, v24.4s, v20.4s\n"
+ "zip2 v21.4s, v25.4s, v21.4s\n"
+ "zip2 v20.4s, v24.4s, v20.4s\n"
+ "zip1 v27.4s, v23.4s, v18.4s\n"
+ "zip1 v26.4s, v19.4s, v17.4s\n"
+ "zip2 v25.4s, v23.4s, v18.4s\n"
+ "zip2 v24.4s, v19.4s, v17.4s\n"
+ "zip1 v19.4s, v22.4s, v16.4s\n"
+ "zip1 v18.4s, v21.4s, v20.4s\n"
"zip1 v17.4s, v27.4s, v26.4s\n"
+ "zip2 v23.4s, v22.4s, v16.4s\n"
"zip1 v16.4s, v25.4s, v24.4s\n"
- ".inst 0x0ea16a77 // bfcvtn v23.4h, v19.4s\n"
- "zip2 v22.4s, v22.4s, v21.4s\n"
- ".inst 0x0ea16a55 // bfcvtn v21.4h, v18.4s\n"
- "zip2 v20.4s, v28.4s, v20.4s\n"
+ "zip2 v22.4s, v21.4s, v20.4s\n"
+ ".inst 0x0ea16a75 // bfcvtn v21.4h, v19.4s\n"
+ ".inst 0x0ea16a54 // bfcvtn v20.4h, v18.4s\n"
".inst 0x0ea16a33 // bfcvtn v19.4h, v17.4s\n"
"zip2 v18.4s, v27.4s, v26.4s\n"
".inst 0x0ea16a11 // bfcvtn v17.4h, v16.4s\n"
"zip2 v16.4s, v25.4s, v24.4s\n"
- ".inst 0x4ea16ad7 // bfcvtn2 v23.8h, v22.4s\n"
- ".inst 0x4ea16a95 // bfcvtn2 v21.8h, v20.4s\n"
- "str q23, [x21, #0x0]\n"
+ ".inst 0x4ea16af5 // bfcvtn2 v21.8h, v23.4s\n"
+ ".inst 0x4ea16ad4 // bfcvtn2 v20.8h, v22.4s\n"
".inst 0x4ea16a53 // bfcvtn2 v19.8h, v18.4s\n"
".inst 0x4ea16a11 // bfcvtn2 v17.8h, v16.4s\n"
- "str q21, [x21, #0x10]\n"
- "str q19, [x21, #0xc0]\n"
- "str q17, [x21, #0xd0]\n"
- "add x21, x21, #0x20\n"
+ "str q21, [x27, #0x0]\n"
+ "str q20, [x27, #0x10]\n"
+ "str q19, [x27, #0xc0]\n"
+ "str q17, [x27, #0xd0]\n"
+ "add x27, x27, #0x20\n"
"bge 6b\n"
"7:" // Main row loop: width 4 loop: skip
- "cmp x24, #0x1\n"
+ "cmp x28, #0x1\n"
"blt 9f\n"
"8:" // Main row loop: width 1 loop: loop
- "ldr s19, [x9], #0x4\n"
- "ldr s18, [x28], #0x4\n"
- "sub x24, x24, #0x1\n"
- "cmp x24, #0x1\n"
- "ldr s17, [x27], #0x4\n"
- "ldr s16, [x26], #0x4\n"
+ "ldr s23, [x9], #0x4\n"
+ "ldr s22, [x26], #0x4\n"
+ "sub x28, x28, #0x1\n"
+ "ldr s19, [x25], #0x4\n"
+ "ldr s17, [x24], #0x4\n"
+ "cmp x28, #0x1\n"
+ "ldr s21, [x23], #0x4\n"
+ "ldr s20, [x22], #0x4\n"
+ "ldr s18, [x21], #0x4\n"
+ "ldr s16, [x20], #0x4\n"
+ "zip1 v19.4s, v23.4s, v19.4s\n"
+ "zip1 v17.4s, v22.4s, v17.4s\n"
+ "zip1 v18.4s, v21.4s, v18.4s\n"
+ "zip1 v16.4s, v20.4s, v16.4s\n"
"zip1 v17.4s, v19.4s, v17.4s\n"
"zip1 v16.4s, v18.4s, v16.4s\n"
- "ldr s20, [x25], #0x4\n"
- "ldr s19, [x23], #0x4\n"
- "zip1 v16.4s, v17.4s, v16.4s\n"
- ".inst 0x0ea16a12 // bfcvtn v18.4h, v16.4s\n"
- "ldr s17, [x22], #0x4\n"
- "ldr s16, [x20], #0x4\n"
- "zip1 v17.4s, v20.4s, v17.4s\n"
- "zip1 v16.4s, v19.4s, v16.4s\n"
- "zip1 v16.4s, v17.4s, v16.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
- "str d18, [x21, #0x0]\n"
- "str d16, [x21, #0xc0]\n"
- "add x21, x21, #0x8\n"
+ "str d17, [x27, #0x0]\n"
+ "str d16, [x27, #0xc0]\n"
+ "add x27, x27, #0x8\n"
"bge 8b\n"
"9:" // Main row loop: width 1 loop: skip
+ "10:" // Main row loop: odd col skip
"cmp %x[height], #0x8\n"
"add %x[out], %x[out], #0x180\n"
"bge 1b\n"
- "cbz %x[height], 20f\n"
- "10:" // Main loop skip
- "11:" // Tail row loop: Head
+ "cbz %x[height], 22f\n"
+ "11:" // Main loop skip
+ "12:" // Tail row loop: Head
"mov x9, %x[in]\n"
- "add x28, x9, %x[in_stride]\n"
- "add x27, x28, %x[in_stride]\n"
"mov x20, %x[width]\n"
- "add x26, x27, %x[in_stride]\n"
"cmp %x[height], #0x3\n"
- "add %x[in], x26, %x[in_stride]\n"
- "csel x26, x26, %x[pad_row], GT\n"
- "csel x27, x27, %x[pad_row], GE\n"
+ "mov x27, %x[out]\n"
+ "add x26, x9, %x[in_stride]\n"
+ "add x25, x26, %x[in_stride]\n"
+ "add x24, x25, %x[in_stride]\n"
+ "csel x25, x25, %x[pad_row], GE\n"
+ "add %x[in], x24, %x[in_stride]\n"
+ "csel x24, x24, %x[pad_row], GT\n"
"cmp %x[height], #0x1\n"
- "csel x28, x28, %x[pad_row], GT\n"
- "cmp x20, #0x18\n"
- "mov x21, %x[out]\n"
"sub %x[height], %x[height], #0x4\n"
- "blt 13f\n"
- "12:" // Tail row loop: Column loop
- "ldr q20, [x9], #0x10\n"
- "ldr q18, [x28], #0x10\n"
+ "csel x26, x26, %x[pad_row], GT\n"
+ "cmp x20, #0x18\n"
+ "blt 14f\n"
+ "13:" // Tail row loop: Column loop
+ "ldr q24, [x9], #0x10\n"
+ "ldr q22, [x26], #0x10\n"
"sub x20, x20, #0x18\n"
+ "ldr q19, [x25], #0x10\n"
+ "ldr q17, [x24], #0x10\n"
"cmp x20, #0x18\n"
- "ldr q17, [x27], #0x10\n"
- "ldr q16, [x26], #0x10\n"
- "zip1 v9.4s, v20.4s, v17.4s\n"
- "zip1 v30.4s, v18.4s, v16.4s\n"
- "ldr q21, [x9], #0x10\n"
- "ldr q19, [x28], #0x10\n"
- "zip2 v17.4s, v20.4s, v17.4s\n"
- "zip2 v5.4s, v18.4s, v16.4s\n"
- "ldr q18, [x27], #0x10\n"
- "ldr q16, [x26], #0x10\n"
- "zip1 v0.4s, v21.4s, v18.4s\n"
- "zip1 v3.4s, v19.4s, v16.4s\n"
- "ldr q23, [x9], #0x10\n"
- "ldr q20, [x28], #0x10\n"
- "zip2 v1.4s, v21.4s, v18.4s\n"
- "zip2 v16.4s, v19.4s, v16.4s\n"
- "ldr q19, [x27], #0x10\n"
- "ldr q18, [x26], #0x10\n"
- "zip1 v4.4s, v23.4s, v19.4s\n"
- "zip1 v2.4s, v20.4s, v18.4s\n"
- "ldr q22, [x9], #0x10\n"
- "ldr q21, [x28], #0x10\n"
- "zip2 v27.4s, v23.4s, v19.4s\n"
- "zip2 v28.4s, v20.4s, v18.4s\n"
- "ldr q20, [x27], #0x10\n"
- "ldr q24, [x26], #0x10\n"
- "zip1 v25.4s, v22.4s, v20.4s\n"
- "zip1 v26.4s, v21.4s, v24.4s\n"
- "ldr q19, [x9], #0x10\n"
- "ldr q18, [x28], #0x10\n"
- "zip2 v14.4s, v22.4s, v20.4s\n"
- "zip2 v12.4s, v21.4s, v24.4s\n"
- "ldr q31, [x27], #0x10\n"
- "ldr q24, [x26], #0x10\n"
- "zip1 v15.4s, v19.4s, v31.4s\n"
- "zip1 v13.4s, v18.4s, v24.4s\n"
- "ldr q21, [x9], #0x10\n"
- "ldr q11, [x28], #0x10\n"
- "zip2 v20.4s, v19.4s, v31.4s\n"
- "zip2 v10.4s, v18.4s, v24.4s\n"
- "ldr q22, [x27], #0x10\n"
+ "ldr q26, [x9], #0x10\n"
"ldr q23, [x26], #0x10\n"
- "zip1 v19.4s, v21.4s, v22.4s\n"
- "zip1 v18.4s, v11.4s, v23.4s\n"
- "zip2 v6.4s, v21.4s, v22.4s\n"
- "zip2 v11.4s, v11.4s, v23.4s\n"
- "zip1 v8.4s, v9.4s, v30.4s\n"
- "zip1 v21.4s, v17.4s, v5.4s\n"
- "zip1 v7.4s, v0.4s, v3.4s\n"
- "zip1 v31.4s, v1.4s, v16.4s\n"
- "zip1 v29.4s, v4.4s, v2.4s\n"
- "zip1 v22.4s, v27.4s, v28.4s\n"
- "zip1 v24.4s, v25.4s, v26.4s\n"
- "zip1 v23.4s, v14.4s, v12.4s\n"
- ".inst 0x0ea16908 // bfcvtn v8.4h, v8.4s\n"
- "zip2 v9.4s, v9.4s, v30.4s\n"
- "zip1 v30.4s, v15.4s, v13.4s\n"
- ".inst 0x0ea16ab5 // bfcvtn v21.4h, v21.4s\n"
- "zip2 v5.4s, v17.4s, v5.4s\n"
- "zip1 v17.4s, v20.4s, v10.4s\n"
+ "ldr q18, [x25], #0x10\n"
+ "ldr q16, [x24], #0x10\n"
+ "ldr q25, [x9], #0x10\n"
+ "zip1 v6.4s, v24.4s, v19.4s\n"
+ "zip1 v11.4s, v22.4s, v17.4s\n"
+ "ldr q21, [x26], #0x10\n"
+ "ldr q20, [x25], #0x10\n"
+ "zip2 v2.4s, v24.4s, v19.4s\n"
+ "zip2 v22.4s, v22.4s, v17.4s\n"
+ "ldr q17, [x24], #0x10\n"
+ "ldr q7, [x9], #0x10\n"
+ "zip1 v19.4s, v26.4s, v18.4s\n"
+ "zip1 v5.4s, v23.4s, v16.4s\n"
+ "ldr q3, [x26], #0x10\n"
+ "ldr q24, [x25], #0x10\n"
+ "zip2 v0.4s, v26.4s, v18.4s\n"
+ "zip2 v4.4s, v23.4s, v16.4s\n"
+ "ldr q16, [x24], #0x10\n"
+ "ldr q1, [x9], #0x10\n"
+ "zip1 v31.4s, v25.4s, v20.4s\n"
+ "zip1 v23.4s, v21.4s, v17.4s\n"
+ "ldr q30, [x26], #0x10\n"
+ "ldr q18, [x25], #0x10\n"
+ "zip2 v29.4s, v25.4s, v20.4s\n"
+ "zip2 v28.4s, v21.4s, v17.4s\n"
+ "ldr q25, [x24], #0x10\n"
+ "ldr q20, [x9], #0x10\n"
+ "zip1 v21.4s, v7.4s, v24.4s\n"
+ "zip1 v26.4s, v3.4s, v16.4s\n"
+ "ldr q17, [x26], #0x10\n"
+ "ldr q27, [x25], #0x10\n"
+ "zip2 v24.4s, v7.4s, v24.4s\n"
+ "zip2 v14.4s, v3.4s, v16.4s\n"
+ "ldr q10, [x24], #0x10\n"
+ "zip1 v16.4s, v1.4s, v18.4s\n"
+ "zip1 v13.4s, v30.4s, v25.4s\n"
+ "zip2 v18.4s, v1.4s, v18.4s\n"
+ "zip2 v12.4s, v30.4s, v25.4s\n"
+ "zip1 v15.4s, v20.4s, v27.4s\n"
+ "zip1 v9.4s, v17.4s, v10.4s\n"
+ "zip2 v8.4s, v20.4s, v27.4s\n"
+ "zip2 v10.4s, v17.4s, v10.4s\n"
+ "zip1 v17.4s, v6.4s, v11.4s\n"
+ "zip1 v7.4s, v2.4s, v22.4s\n"
+ "zip1 v3.4s, v19.4s, v5.4s\n"
+ "zip1 v1.4s, v0.4s, v4.4s\n"
+ "zip1 v30.4s, v31.4s, v23.4s\n"
+ "zip1 v25.4s, v29.4s, v28.4s\n"
+ "zip1 v27.4s, v21.4s, v26.4s\n"
+ "zip1 v20.4s, v24.4s, v14.4s\n"
+ ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n"
+ "zip2 v6.4s, v6.4s, v11.4s\n"
+ "zip1 v11.4s, v16.4s, v13.4s\n"
".inst 0x0ea168e7 // bfcvtn v7.4h, v7.4s\n"
- "zip2 v0.4s, v0.4s, v3.4s\n"
- "zip1 v3.4s, v19.4s, v18.4s\n"
- ".inst 0x0ea16bff // bfcvtn v31.4h, v31.4s\n"
- "zip2 v16.4s, v1.4s, v16.4s\n"
- "zip1 v1.4s, v6.4s, v11.4s\n"
- ".inst 0x0ea16bbd // bfcvtn v29.4h, v29.4s\n"
- "zip2 v2.4s, v4.4s, v2.4s\n"
- ".inst 0x0ea16ac4 // bfcvtn v4.4h, v22.4s\n"
- "zip2 v27.4s, v27.4s, v28.4s\n"
- ".inst 0x0ea16b1c // bfcvtn v28.4h, v24.4s\n"
- "zip2 v25.4s, v25.4s, v26.4s\n"
- ".inst 0x0ea16afa // bfcvtn v26.4h, v23.4s\n"
- "zip2 v14.4s, v14.4s, v12.4s\n"
- ".inst 0x0ea16bd8 // bfcvtn v24.4h, v30.4s\n"
- "zip2 v13.4s, v15.4s, v13.4s\n"
- ".inst 0x0ea16a2f // bfcvtn v15.4h, v17.4s\n"
- "zip2 v12.4s, v20.4s, v10.4s\n"
- ".inst 0x0ea16874 // bfcvtn v20.4h, v3.4s\n"
- "zip2 v10.4s, v19.4s, v18.4s\n"
- ".inst 0x0ea16831 // bfcvtn v17.4h, v1.4s\n"
- "zip2 v18.4s, v6.4s, v11.4s\n"
- ".inst 0x4ea16928 // bfcvtn2 v8.8h, v9.4s\n"
- ".inst 0x4ea168b5 // bfcvtn2 v21.8h, v5.4s\n"
- "str q8, [x21, #0x0]\n"
- ".inst 0x4ea16807 // bfcvtn2 v7.8h, v0.4s\n"
- ".inst 0x4ea16a1f // bfcvtn2 v31.8h, v16.4s\n"
- "str q21, [x21, #0x10]\n"
- ".inst 0x4ea1685d // bfcvtn2 v29.8h, v2.4s\n"
- ".inst 0x4ea16b64 // bfcvtn2 v4.8h, v27.4s\n"
- "str q7, [x21, #0x20]\n"
- ".inst 0x4ea16b3c // bfcvtn2 v28.8h, v25.4s\n"
- ".inst 0x4ea169da // bfcvtn2 v26.8h, v14.4s\n"
- "str q31, [x21, #0x30]\n"
- ".inst 0x4ea169b8 // bfcvtn2 v24.8h, v13.4s\n"
- ".inst 0x4ea1698f // bfcvtn2 v15.8h, v12.4s\n"
- "str q29, [x21, #0x40]\n"
- ".inst 0x4ea16954 // bfcvtn2 v20.8h, v10.4s\n"
- ".inst 0x4ea16a51 // bfcvtn2 v17.8h, v18.4s\n"
- "str q4, [x21, #0x50]\n"
- "str q28, [x21, #0x60]\n"
- "str q26, [x21, #0x70]\n"
- "str q24, [x21, #0x80]\n"
- "str q15, [x21, #0x90]\n"
- "str q20, [x21, #0xa0]\n"
- "str q17, [x21, #0xb0]\n"
- "add x21, x21, %x[out_stride]\n"
- "bge 12b\n"
- "13:" // Tail row loop: Column loop skip
+ "zip2 v22.4s, v2.4s, v22.4s\n"
+ "zip1 v2.4s, v18.4s, v12.4s\n"
+ ".inst 0x0ea16863 // bfcvtn v3.4h, v3.4s\n"
+ "zip2 v5.4s, v19.4s, v5.4s\n"
+ "zip1 v19.4s, v15.4s, v9.4s\n"
+ ".inst 0x0ea16821 // bfcvtn v1.4h, v1.4s\n"
+ "zip2 v4.4s, v0.4s, v4.4s\n"
+ "zip1 v0.4s, v8.4s, v10.4s\n"
+ ".inst 0x0ea16bde // bfcvtn v30.4h, v30.4s\n"
+ "zip2 v31.4s, v31.4s, v23.4s\n"
+ ".inst 0x0ea16b37 // bfcvtn v23.4h, v25.4s\n"
+ "zip2 v29.4s, v29.4s, v28.4s\n"
+ ".inst 0x0ea16b7c // bfcvtn v28.4h, v27.4s\n"
+ "zip2 v27.4s, v21.4s, v26.4s\n"
+ ".inst 0x0ea16a9a // bfcvtn v26.4h, v20.4s\n"
+ "zip2 v25.4s, v24.4s, v14.4s\n"
+ ".inst 0x0ea16978 // bfcvtn v24.4h, v11.4s\n"
+ "zip2 v14.4s, v16.4s, v13.4s\n"
+ ".inst 0x0ea16855 // bfcvtn v21.4h, v2.4s\n"
+ "zip2 v20.4s, v18.4s, v12.4s\n"
+ ".inst 0x0ea16a73 // bfcvtn v19.4h, v19.4s\n"
+ "zip2 v18.4s, v15.4s, v9.4s\n"
+ ".inst 0x0ea1680b // bfcvtn v11.4h, v0.4s\n"
+ "zip2 v16.4s, v8.4s, v10.4s\n"
+ ".inst 0x4ea168d1 // bfcvtn2 v17.8h, v6.4s\n"
+ ".inst 0x4ea16ac7 // bfcvtn2 v7.8h, v22.4s\n"
+ ".inst 0x4ea168a3 // bfcvtn2 v3.8h, v5.4s\n"
+ ".inst 0x4ea16881 // bfcvtn2 v1.8h, v4.4s\n"
+ ".inst 0x4ea16bfe // bfcvtn2 v30.8h, v31.4s\n"
+ ".inst 0x4ea16bb7 // bfcvtn2 v23.8h, v29.4s\n"
+ ".inst 0x4ea16b7c // bfcvtn2 v28.8h, v27.4s\n"
+ ".inst 0x4ea16b3a // bfcvtn2 v26.8h, v25.4s\n"
+ "str q17, [x27, #0x0]\n"
+ ".inst 0x4ea169d8 // bfcvtn2 v24.8h, v14.4s\n"
+ ".inst 0x4ea16a95 // bfcvtn2 v21.8h, v20.4s\n"
+ "str q7, [x27, #0x10]\n"
+ ".inst 0x4ea16a53 // bfcvtn2 v19.8h, v18.4s\n"
+ ".inst 0x4ea16a0b // bfcvtn2 v11.8h, v16.4s\n"
+ "str q3, [x27, #0x20]\n"
+ "str q1, [x27, #0x30]\n"
+ "str q30, [x27, #0x40]\n"
+ "str q23, [x27, #0x50]\n"
+ "str q28, [x27, #0x60]\n"
+ "str q26, [x27, #0x70]\n"
+ "str q24, [x27, #0x80]\n"
+ "str q21, [x27, #0x90]\n"
+ "str q19, [x27, #0xa0]\n"
+ "str q11, [x27, #0xb0]\n"
+ "add x27, x27, %x[out_stride]\n"
+ "bge 13b\n"
+ "14:" // Tail row loop: Column loop skip
+ "cbz x20, 21f\n"
"cmp x20, #0x10\n"
- "blt 15f\n"
- "14:" // Tail row loop: width 16 loop: loop
- "ldr q21, [x9], #0x10\n"
- "ldr q20, [x28], #0x10\n"
+ "movi v16.16b, #0x0\n"
+ "str q16, [x27, #0x0]\n"
+ "str q16, [x27, #0x10]\n"
+ "str q16, [x27, #0x20]\n"
+ "str q16, [x27, #0x30]\n"
+ "str q16, [x27, #0x40]\n"
+ "str q16, [x27, #0x50]\n"
+ "str q16, [x27, #0x60]\n"
+ "str q16, [x27, #0x70]\n"
+ "str q16, [x27, #0x80]\n"
+ "str q16, [x27, #0x90]\n"
+ "str q16, [x27, #0xa0]\n"
+ "str q16, [x27, #0xb0]\n"
+ "blt 16f\n"
+ "15:" // Tail row loop: width 16 loop: loop
+ "ldr q20, [x9], #0x10\n"
+ "ldr q19, [x26], #0x10\n"
"sub x20, x20, #0x10\n"
+ "ldr q18, [x25], #0x10\n"
+ "ldr q17, [x24], #0x10\n"
"cmp x20, #0x10\n"
- "ldr q17, [x27], #0x10\n"
- "ldr q16, [x26], #0x10\n"
- "zip1 v30.4s, v21.4s, v17.4s\n"
- "zip1 v29.4s, v20.4s, v16.4s\n"
- "ldr q19, [x9], #0x10\n"
- "ldr q18, [x28], #0x10\n"
- "zip2 v28.4s, v21.4s, v17.4s\n"
- "zip2 v27.4s, v20.4s, v16.4s\n"
- "ldr q17, [x27], #0x10\n"
- "ldr q16, [x26], #0x10\n"
- "zip1 v26.4s, v19.4s, v17.4s\n"
- "zip1 v25.4s, v18.4s, v16.4s\n"
- "ldr q21, [x9], #0x10\n"
- "ldr q20, [x28], #0x10\n"
- "zip2 v8.4s, v19.4s, v17.4s\n"
- "zip2 v24.4s, v18.4s, v16.4s\n"
- "ldr q17, [x27], #0x10\n"
- "ldr q16, [x26], #0x10\n"
- "zip1 v7.4s, v21.4s, v17.4s\n"
- "zip1 v6.4s, v20.4s, v16.4s\n"
+ "ldr q0, [x9], #0x10\n"
+ "ldr q31, [x26], #0x10\n"
+ "ldr q24, [x25], #0x10\n"
+ "ldr q16, [x24], #0x10\n"
+ "ldr q23, [x9], #0x10\n"
+ "zip1 v30.4s, v20.4s, v18.4s\n"
+ "zip1 v29.4s, v19.4s, v17.4s\n"
+ "ldr q22, [x26], #0x10\n"
+ "ldr q21, [x25], #0x10\n"
+ "zip2 v28.4s, v20.4s, v18.4s\n"
+ "zip2 v27.4s, v19.4s, v17.4s\n"
+ "ldr q20, [x24], #0x10\n"
"ldr q19, [x9], #0x10\n"
- "ldr q18, [x28], #0x10\n"
- "zip2 v5.4s, v21.4s, v17.4s\n"
- "zip2 v4.4s, v20.4s, v16.4s\n"
- "ldr q17, [x27], #0x10\n"
- "ldr q16, [x26], #0x10\n"
+ "zip1 v26.4s, v0.4s, v24.4s\n"
+ "zip1 v25.4s, v31.4s, v16.4s\n"
+ "ldr q18, [x26], #0x10\n"
+ "ldr q17, [x25], #0x10\n"
+ "zip2 v8.4s, v0.4s, v24.4s\n"
+ "zip2 v24.4s, v31.4s, v16.4s\n"
+ "ldr q16, [x24], #0x10\n"
+ "zip1 v7.4s, v23.4s, v21.4s\n"
+ "zip1 v6.4s, v22.4s, v20.4s\n"
+ "zip2 v5.4s, v23.4s, v21.4s\n"
+ "zip2 v4.4s, v22.4s, v20.4s\n"
"zip1 v3.4s, v19.4s, v17.4s\n"
"zip1 v2.4s, v18.4s, v16.4s\n"
"zip2 v1.4s, v19.4s, v17.4s\n"
@@ -698,70 +739,71 @@ void a64_transpose_interleave_24_2x4_fp32bf16(bfloat16 *out, const float *in, si
"zip2 v16.4s, v1.4s, v0.4s\n"
".inst 0x4ea16bdf // bfcvtn2 v31.8h, v30.4s\n"
".inst 0x4ea16b9d // bfcvtn2 v29.8h, v28.4s\n"
- "str q31, [x21, #0x0]\n"
".inst 0x4ea16b5b // bfcvtn2 v27.8h, v26.4s\n"
".inst 0x4ea16b19 // bfcvtn2 v25.8h, v24.4s\n"
- "str q29, [x21, #0x10]\n"
".inst 0x4ea16ad7 // bfcvtn2 v23.8h, v22.4s\n"
".inst 0x4ea16a95 // bfcvtn2 v21.8h, v20.4s\n"
- "str q27, [x21, #0x20]\n"
".inst 0x4ea16a53 // bfcvtn2 v19.8h, v18.4s\n"
".inst 0x4ea16a11 // bfcvtn2 v17.8h, v16.4s\n"
- "str q25, [x21, #0x30]\n"
- "str q23, [x21, #0x40]\n"
- "str q21, [x21, #0x50]\n"
- "str q19, [x21, #0x60]\n"
- "str q17, [x21, #0x70]\n"
- "add x21, x21, #0x80\n"
- "bge 14b\n"
- "15:" // Tail row loop: width 16 loop: skip
+ "str q31, [x27, #0x0]\n"
+ "str q29, [x27, #0x10]\n"
+ "str q27, [x27, #0x20]\n"
+ "str q25, [x27, #0x30]\n"
+ "str q23, [x27, #0x40]\n"
+ "str q21, [x27, #0x50]\n"
+ "str q19, [x27, #0x60]\n"
+ "str q17, [x27, #0x70]\n"
+ "add x27, x27, #0x80\n"
+ "bge 15b\n"
+ "16:" // Tail row loop: width 16 loop: skip
"cmp x20, #0x4\n"
- "blt 17f\n"
- "16:" // Tail row loop: width 4 loop: loop
- "ldr q20, [x9], #0x10\n"
- "ldr q19, [x28], #0x10\n"
+ "blt 18f\n"
+ "17:" // Tail row loop: width 4 loop: loop
+ "ldr q21, [x9], #0x10\n"
+ "ldr q20, [x26], #0x10\n"
"sub x20, x20, #0x4\n"
+ "ldr q19, [x25], #0x10\n"
+ "ldr q17, [x24], #0x10\n"
"cmp x20, #0x4\n"
- "ldr q17, [x27], #0x10\n"
- "ldr q16, [x26], #0x10\n"
- "zip1 v22.4s, v20.4s, v17.4s\n"
- "zip1 v18.4s, v19.4s, v16.4s\n"
- "zip2 v21.4s, v20.4s, v17.4s\n"
- "zip2 v20.4s, v19.4s, v16.4s\n"
- "zip1 v17.4s, v22.4s, v18.4s\n"
+ "zip1 v18.4s, v21.4s, v19.4s\n"
+ "zip1 v16.4s, v20.4s, v17.4s\n"
+ "zip2 v21.4s, v21.4s, v19.4s\n"
+ "zip2 v20.4s, v20.4s, v17.4s\n"
+ "zip1 v17.4s, v18.4s, v16.4s\n"
+ "zip2 v19.4s, v18.4s, v16.4s\n"
"zip1 v16.4s, v21.4s, v20.4s\n"
- ".inst 0x0ea16a33 // bfcvtn v19.4h, v17.4s\n"
- "zip2 v18.4s, v22.4s, v18.4s\n"
- ".inst 0x0ea16a11 // bfcvtn v17.4h, v16.4s\n"
- "zip2 v16.4s, v21.4s, v20.4s\n"
- ".inst 0x4ea16a53 // bfcvtn2 v19.8h, v18.4s\n"
- ".inst 0x4ea16a11 // bfcvtn2 v17.8h, v16.4s\n"
- "str q19, [x21, #0x0]\n"
- "str q17, [x21, #0x10]\n"
- "add x21, x21, #0x20\n"
- "bge 16b\n"
- "17:" // Tail row loop: width 4 loop: skip
+ ".inst 0x0ea16a32 // bfcvtn v18.4h, v17.4s\n"
+ "zip2 v17.4s, v21.4s, v20.4s\n"
+ ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
+ ".inst 0x4ea16a72 // bfcvtn2 v18.8h, v19.4s\n"
+ ".inst 0x4ea16a30 // bfcvtn2 v16.8h, v17.4s\n"
+ "str q18, [x27, #0x0]\n"
+ "str q16, [x27, #0x10]\n"
+ "add x27, x27, #0x20\n"
+ "bge 17b\n"
+ "18:" // Tail row loop: width 4 loop: skip
"cmp x20, #0x1\n"
- "blt 19f\n"
- "18:" // Tail row loop: width 1 loop: loop
+ "blt 20f\n"
+ "19:" // Tail row loop: width 1 loop: loop
"ldr s19, [x9], #0x4\n"
- "ldr s18, [x28], #0x4\n"
+ "ldr s18, [x26], #0x4\n"
"sub x20, x20, #0x1\n"
+ "ldr s17, [x25], #0x4\n"
+ "ldr s16, [x24], #0x4\n"
"cmp x20, #0x1\n"
- "ldr s17, [x27], #0x4\n"
- "ldr s16, [x26], #0x4\n"
"zip1 v17.4s, v19.4s, v17.4s\n"
"zip1 v16.4s, v18.4s, v16.4s\n"
"zip1 v16.4s, v17.4s, v16.4s\n"
".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n"
- "str d16, [x21, #0x0]\n"
- "add x21, x21, #0x8\n"
- "bge 18b\n"
- "19:" // Tail row loop: width 1 loop: skip
+ "str d16, [x27, #0x0]\n"
+ "add x27, x27, #0x8\n"
+ "bge 19b\n"
+ "20:" // Tail row loop: width 1 loop: skip
+ "21:" // Tail row loop: odd col skip
"cmp %x[height], #0x1\n"
"add %x[out], %x[out], #0xc0\n"
- "bge 11b\n"
- "20:" // Done
+ "bge 12b\n"
+ "22:" // Done
: [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
: [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [pad_row] "r" (pad_row), [width] "r" (width)
: "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", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"