From 4ee8b1599dbaf7634d25607fa5ac96ba3dc6b0f2 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Fri, 16 Jul 2021 16:16:43 +0100 Subject: Update GEMM assembly kernels - Introduce Fp32 kernels with internal calculations in Bfloat16 when fast_mode is enabled - Improve kernel selection heuristics Signed-off-by: Georgios Pinitas Change-Id: I68a9e7e862b6fd2721b46e0d7cc791091c4ab279 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5965 Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins --- .../transforms/a64_transpose_interleave_16_2x4.hpp | 511 +++++++++++++++++++++ 1 file changed, 511 insertions(+) create mode 100644 src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4.hpp (limited to 'src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4.hpp') diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4.hpp new file mode 100644 index 0000000000..8c8dfd1d0d --- /dev/null +++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_16_2x4.hpp @@ -0,0 +1,511 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING + * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS + * IN THE SOFTWARE. + */ + +#pragma once + +#ifdef __aarch64__ + +namespace { + +void a64_transpose_interleave_16_2x4(uint16_t *out, const uint16_t *in, size_t width, size_t in_stride, size_t height) +{ + uint16_t *pad_row = reinterpret_cast(alloca(width * sizeof(uint16_t))); + + if (height % 4) { + memset(pad_row, 0, width * sizeof(uint16_t)); + } + + size_t out_stride = 16 * roundup(height, 4) * sizeof(uint16_t); + + __asm__ __volatile__( + "cmp %x[height], #0x8\n" + "blt 10f\n" + "1:" // Main row loop: Head + "mov x28, %x[in]\n" + "mov x27, %x[out]\n" + "add x26, x28, %x[in_stride]\n" + "add x25, x26, %x[in_stride]\n" + "add x24, x25, %x[in_stride]\n" + "add x23, x24, %x[in_stride]\n" + "add x22, x23, %x[in_stride]\n" + "add x21, x22, %x[in_stride]\n" + "add x20, x21, %x[in_stride]\n" + "add %x[in], x20, %x[in_stride]\n" + "sub %x[height], %x[height], #0x8\n" + "mov x19, %x[width]\n" + "cmp x19, #0x20\n" + "blt 3f\n" + "2:" // Main row loop: Unroll column loop + "ldr q24, [x28], #0x10\n" + "sub x19, x19, #0x20\n" + "ldr q4, [x26], #0x10\n" + "cmp x19, #0x20\n" + "ldr q26, [x25], #0x10\n" + "zip1 v2.8h, v24.8h, v26.8h\n" + "ldr q3, [x28], #0x10\n" + "zip2 v9.8h, v24.8h, v26.8h\n" + "ldr q0, [x26], #0x10\n" + "ldr q22, [x25], #0x10\n" + "zip1 v31.8h, v3.8h, v22.8h\n" + "ldr q23, [x28], #0x10\n" + "zip2 v25.8h, v3.8h, v22.8h\n" + "ldr q22, [x26], #0x10\n" + "ldr q5, [x25], #0x10\n" + "zip1 v17.8h, v23.8h, v5.8h\n" + "ldr q19, [x28], #0x10\n" + "zip2 v20.8h, v23.8h, v5.8h\n" + "ldr q1, [x26], #0x10\n" + "ldr q16, [x25], #0x10\n" + "zip1 v15.8h, v19.8h, v16.8h\n" + "ldr q8, [x24], #0x10\n" + "zip2 v11.8h, v19.8h, v16.8h\n" + "ldr q26, [x23], #0x10\n" + "ldr q19, [x22], #0x10\n" + "zip1 v5.8h, v4.8h, v8.8h\n" + "ldr q18, [x24], #0x10\n" + "zip2 v6.8h, v4.8h, v8.8h\n" + "ldr q7, [x23], #0x10\n" + "zip1 v27.8h, v2.8h, v5.8h\n" + "ldr q23, [x22], #0x10\n" + "zip2 v8.8h, v2.8h, v5.8h\n" + "ldr q24, [x21], #0x10\n" + "zip1 v12.8h, v9.8h, v6.8h\n" + "ldr q13, [x24], #0x10\n" + "zip2 v16.8h, v9.8h, v6.8h\n" + "ldr q9, [x23], #0x10\n" + "zip1 v29.8h, v0.8h, v18.8h\n" + "ldr q10, [x22], #0x10\n" + "zip1 v14.8h, v31.8h, v29.8h\n" + "ldr q4, [x21], #0x10\n" + "zip2 v21.8h, v31.8h, v29.8h\n" + "ldr q6, [x24], #0x10\n" + "zip2 v18.8h, v0.8h, v18.8h\n" + "ldr q3, [x23], #0x10\n" + "zip1 v0.8h, v25.8h, v18.8h\n" + "ldr q31, [x22], #0x10\n" + "zip2 v29.8h, v25.8h, v18.8h\n" + "ldr q5, [x21], #0x10\n" + "zip1 v28.8h, v26.8h, v24.8h\n" + "ldr q25, [x20], #0x10\n" + "zip2 v26.8h, v26.8h, v24.8h\n" + "ldr q30, [x21], #0x10\n" + "zip1 v24.8h, v7.8h, v4.8h\n" + "zip2 v4.8h, v7.8h, v4.8h\n" + "ldr q18, [x20], #0x10\n" + "zip1 v7.8h, v19.8h, v25.8h\n" + "ldr q2, [x20], #0x10\n" + "zip2 v25.8h, v19.8h, v25.8h\n" + "zip1 v19.8h, v28.8h, v7.8h\n" + "zip2 v7.8h, v28.8h, v7.8h\n" + "ldr q28, [x20], #0x10\n" + "str q27, [x27, #0x0]\n" + "zip1 v27.8h, v26.8h, v25.8h\n" + "zip2 v26.8h, v26.8h, v25.8h\n" + "str q8, [x27, #0x10]\n" + "zip1 v25.8h, v23.8h, v18.8h\n" + "str q12, [x27, #0x20]\n" + "zip1 v8.8h, v24.8h, v25.8h\n" + "str q16, [x27, #0x30]\n" + "zip2 v25.8h, v24.8h, v25.8h\n" + "str q14, [x27, #0x40]\n" + "zip2 v12.8h, v23.8h, v18.8h\n" + "str q21, [x27, #0x50]\n" + "zip1 v21.8h, v4.8h, v12.8h\n" + "str q0, [x27, #0x60]\n" + "zip2 v14.8h, v4.8h, v12.8h\n" + "str q29, [x27, #0x70]\n" + "zip1 v12.8h, v22.8h, v13.8h\n" + "str q19, [x27, #0x80]\n" + "zip1 v24.8h, v17.8h, v12.8h\n" + "str q7, [x27, #0x90]\n" + "zip2 v23.8h, v17.8h, v12.8h\n" + "str q27, [x27, #0xa0]\n" + "zip2 v16.8h, v22.8h, v13.8h\n" + "str q26, [x27, #0xb0]\n" + "zip1 v19.8h, v20.8h, v16.8h\n" + "str q8, [x27, #0xc0]\n" + "zip2 v18.8h, v20.8h, v16.8h\n" + "str q25, [x27, #0xd0]\n" + "zip1 v16.8h, v1.8h, v6.8h\n" + "str q21, [x27, #0xe0]\n" + "zip1 v21.8h, v15.8h, v16.8h\n" + "str q14, [x27, #0xf0]\n" + "add x27, x27, %x[out_stride]\n" + "zip2 v17.8h, v15.8h, v16.8h\n" + "str q24, [x27, #0x0]\n" + "zip2 v16.8h, v1.8h, v6.8h\n" + "str q23, [x27, #0x10]\n" + "zip1 v20.8h, v11.8h, v16.8h\n" + "str q19, [x27, #0x20]\n" + "zip2 v19.8h, v11.8h, v16.8h\n" + "str q18, [x27, #0x30]\n" + "zip1 v18.8h, v9.8h, v5.8h\n" + "str q21, [x27, #0x40]\n" + "zip1 v16.8h, v10.8h, v2.8h\n" + "str q17, [x27, #0x50]\n" + "zip1 v17.8h, v18.8h, v16.8h\n" + "str q20, [x27, #0x60]\n" + "zip2 v16.8h, v18.8h, v16.8h\n" + "str q19, [x27, #0x70]\n" + "zip2 v18.8h, v9.8h, v5.8h\n" + "str q17, [x27, #0x80]\n" + "zip2 v17.8h, v10.8h, v2.8h\n" + "str q16, [x27, #0x90]\n" + "zip1 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xa0]\n" + "zip2 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xb0]\n" + "zip1 v18.8h, v3.8h, v30.8h\n" + "zip1 v17.8h, v31.8h, v28.8h\n" + "zip1 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xc0]\n" + "zip2 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xd0]\n" + "zip2 v18.8h, v3.8h, v30.8h\n" + "zip2 v17.8h, v31.8h, v28.8h\n" + "zip1 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xe0]\n" + "zip2 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xf0]\n" + "add x27, x27, %x[out_stride]\n" + "bge 2b\n" + "3:" // Main row loop: Unroll column loop skip + "cmp x19, #0x10\n" + "blt 5f\n" + "4:" // Main row loop: Column loop + "ldr q17, [x28], #0x10\n" + "sub x19, x19, #0x10\n" + "ldr q20, [x26], #0x10\n" + "cmp x19, #0x10\n" + "ldr q16, [x25], #0x10\n" + "zip1 v19.8h, v17.8h, v16.8h\n" + "ldr q18, [x28], #0x10\n" + "zip2 v22.8h, v17.8h, v16.8h\n" + "ldr q21, [x26], #0x10\n" + "ldr q17, [x25], #0x10\n" + "zip1 v0.8h, v18.8h, v17.8h\n" + "ldr q16, [x24], #0x10\n" + "zip2 v31.8h, v18.8h, v17.8h\n" + "ldr q30, [x23], #0x10\n" + "ldr q29, [x22], #0x10\n" + "zip1 v17.8h, v20.8h, v16.8h\n" + "ldr q18, [x24], #0x10\n" + "zip2 v16.8h, v20.8h, v16.8h\n" + "ldr q28, [x23], #0x10\n" + "zip1 v27.8h, v19.8h, v17.8h\n" + "ldr q26, [x22], #0x10\n" + "zip2 v20.8h, v19.8h, v17.8h\n" + "ldr q25, [x21], #0x10\n" + "zip1 v19.8h, v22.8h, v16.8h\n" + "ldr q24, [x20], #0x10\n" + "zip2 v23.8h, v22.8h, v16.8h\n" + "ldr q22, [x21], #0x10\n" + "zip1 v17.8h, v21.8h, v18.8h\n" + "zip2 v18.8h, v21.8h, v18.8h\n" + "ldr q21, [x20], #0x10\n" + "zip1 v16.8h, v0.8h, v17.8h\n" + "str q27, [x27, #0x0]\n" + "zip2 v17.8h, v0.8h, v17.8h\n" + "str q20, [x27, #0x10]\n" + "zip1 v20.8h, v31.8h, v18.8h\n" + "str q19, [x27, #0x20]\n" + "zip2 v19.8h, v31.8h, v18.8h\n" + "str q23, [x27, #0x30]\n" + "zip1 v18.8h, v30.8h, v25.8h\n" + "str q16, [x27, #0x40]\n" + "zip1 v16.8h, v29.8h, v24.8h\n" + "str q17, [x27, #0x50]\n" + "zip1 v17.8h, v18.8h, v16.8h\n" + "str q20, [x27, #0x60]\n" + "zip2 v16.8h, v18.8h, v16.8h\n" + "str q19, [x27, #0x70]\n" + "zip2 v18.8h, v30.8h, v25.8h\n" + "str q17, [x27, #0x80]\n" + "zip2 v17.8h, v29.8h, v24.8h\n" + "str q16, [x27, #0x90]\n" + "zip1 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xa0]\n" + "zip2 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xb0]\n" + "zip1 v18.8h, v28.8h, v22.8h\n" + "zip1 v17.8h, v26.8h, v21.8h\n" + "zip1 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xc0]\n" + "zip2 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xd0]\n" + "zip2 v18.8h, v28.8h, v22.8h\n" + "zip2 v17.8h, v26.8h, v21.8h\n" + "zip1 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xe0]\n" + "zip2 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0xf0]\n" + "add x27, x27, %x[out_stride]\n" + "bge 4b\n" + "5:" // Main row loop: Column loop skip + "cmp x19, #0x4\n" + "blt 7f\n" + "6:" // Main row loop: width 4 loop: loop + "ldr d17, [x28], #0x8\n" + "sub x19, x19, #0x4\n" + "ldr d18, [x26], #0x8\n" + "cmp x19, #0x4\n" + "ldr d16, [x25], #0x8\n" + "zip1 v17.8h, v17.8h, v16.8h\n" + "ldr d16, [x24], #0x8\n" + "ldr d21, [x23], #0x8\n" + "zip1 v16.8h, v18.8h, v16.8h\n" + "ldr d20, [x22], #0x8\n" + "ldr d19, [x21], #0x8\n" + "zip1 v18.8h, v17.8h, v16.8h\n" + "zip2 v17.8h, v17.8h, v16.8h\n" + "ldr d16, [x20], #0x8\n" + "str q18, [x27, #0x0]\n" + "zip1 v18.8h, v21.8h, v19.8h\n" + "str q17, [x27, #0x10]\n" + "zip1 v17.8h, v20.8h, v16.8h\n" + "zip1 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0x80]\n" + "zip2 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0x90]\n" + "add x27, x27, #0x20\n" + "bge 6b\n" + "7:" // Main row loop: width 4 loop: skip + "cmp x19, #0x1\n" + "blt 9f\n" + "8:" // Main row loop: width 1 loop: loop + "ldr h18, [x28], #0x2\n" + "sub x19, x19, #0x1\n" + "ldr h17, [x26], #0x2\n" + "cmp x19, #0x1\n" + "ldr h16, [x25], #0x2\n" + "zip1 v18.8h, v18.8h, v16.8h\n" + "ldr h16, [x24], #0x2\n" + "ldr h20, [x23], #0x2\n" + "zip1 v16.8h, v17.8h, v16.8h\n" + "ldr h19, [x22], #0x2\n" + "ldr h17, [x21], #0x2\n" + "zip1 v18.8h, v18.8h, v16.8h\n" + "ldr h16, [x20], #0x2\n" + "zip1 v17.8h, v20.8h, v17.8h\n" + "str d18, [x27, #0x0]\n" + "zip1 v16.8h, v19.8h, v16.8h\n" + "zip1 v16.8h, v17.8h, v16.8h\n" + "str d16, [x27, #0x80]\n" + "add x27, x27, #0x8\n" + "bge 8b\n" + "9:" // Main row loop: width 1 loop: skip + "add %x[out], %x[out], #0x100\n" + "cmp %x[height], #0x8\n" + "bge 1b\n" + "cbz %x[height], 20f\n" + "10:" // Main loop skip + + "11:" // Tail row loop: Head + "mov x28, %x[in]\n" + "mov x27, %x[out]\n" + "add x26, x28, %x[in_stride]\n" + "add x25, x26, %x[in_stride]\n" + "add x24, x25, %x[in_stride]\n" + "add %x[in], x24, %x[in_stride]\n" + "cmp %x[height], #0x3\n" + "csel x24, x24, %x[pad_row], GT\n" + "csel x25, x25, %x[pad_row], GE\n" + "cmp %x[height], #0x1\n" + "csel x26, x26, %x[pad_row], GT\n" + "sub %x[height], %x[height], #0x4\n" + "mov x19, %x[width]\n" + "cmp x19, #0x20\n" + "blt 13f\n" + "12:" // Tail row loop: Unroll column loop + "ldr q17, [x28], #0x10\n" + "sub x19, x19, #0x20\n" + "ldr q19, [x26], #0x10\n" + "cmp x19, #0x20\n" + "ldr q16, [x25], #0x10\n" + "zip1 v20.8h, v17.8h, v16.8h\n" + "ldr q18, [x28], #0x10\n" + "zip2 v0.8h, v17.8h, v16.8h\n" + "ldr q31, [x26], #0x10\n" + "ldr q16, [x25], #0x10\n" + "zip1 v30.8h, v18.8h, v16.8h\n" + "ldr q17, [x28], #0x10\n" + "zip2 v29.8h, v18.8h, v16.8h\n" + "ldr q28, [x26], #0x10\n" + "ldr q16, [x25], #0x10\n" + "zip1 v27.8h, v17.8h, v16.8h\n" + "ldr q18, [x28], #0x10\n" + "zip2 v26.8h, v17.8h, v16.8h\n" + "ldr q25, [x26], #0x10\n" + "ldr q17, [x25], #0x10\n" + "zip1 v24.8h, v18.8h, v17.8h\n" + "ldr q16, [x24], #0x10\n" + "zip2 v23.8h, v18.8h, v17.8h\n" + "ldr q22, [x24], #0x10\n" + "zip1 v17.8h, v19.8h, v16.8h\n" + "zip2 v19.8h, v19.8h, v16.8h\n" + "ldr q21, [x24], #0x10\n" + "zip1 v16.8h, v20.8h, v17.8h\n" + "zip2 v17.8h, v20.8h, v17.8h\n" + "ldr q20, [x24], #0x10\n" + "zip1 v18.8h, v0.8h, v19.8h\n" + "zip2 v19.8h, v0.8h, v19.8h\n" + "str q16, [x27, #0x0]\n" + "zip1 v16.8h, v31.8h, v22.8h\n" + "str q17, [x27, #0x10]\n" + "zip1 v17.8h, v30.8h, v16.8h\n" + "str q18, [x27, #0x20]\n" + "zip2 v18.8h, v30.8h, v16.8h\n" + "str q19, [x27, #0x30]\n" + "zip2 v16.8h, v31.8h, v22.8h\n" + "str q17, [x27, #0x40]\n" + "zip1 v17.8h, v29.8h, v16.8h\n" + "str q18, [x27, #0x50]\n" + "zip2 v16.8h, v29.8h, v16.8h\n" + "str q17, [x27, #0x60]\n" + "zip1 v17.8h, v28.8h, v21.8h\n" + "str q16, [x27, #0x70]\n" + "add x27, x27, %x[out_stride]\n" + "zip1 v16.8h, v27.8h, v17.8h\n" + "str q16, [x27, #0x0]\n" + "zip2 v16.8h, v27.8h, v17.8h\n" + "zip2 v17.8h, v28.8h, v21.8h\n" + "str q16, [x27, #0x10]\n" + "zip1 v16.8h, v26.8h, v17.8h\n" + "str q16, [x27, #0x20]\n" + "zip2 v16.8h, v26.8h, v17.8h\n" + "str q16, [x27, #0x30]\n" + "zip1 v17.8h, v25.8h, v20.8h\n" + "zip1 v16.8h, v24.8h, v17.8h\n" + "str q16, [x27, #0x40]\n" + "zip2 v16.8h, v24.8h, v17.8h\n" + "str q16, [x27, #0x50]\n" + "zip2 v17.8h, v25.8h, v20.8h\n" + "zip1 v16.8h, v23.8h, v17.8h\n" + "str q16, [x27, #0x60]\n" + "zip2 v16.8h, v23.8h, v17.8h\n" + "str q16, [x27, #0x70]\n" + "add x27, x27, %x[out_stride]\n" + "bge 12b\n" + "13:" // Tail row loop: Unroll column loop skip + "cmp x19, #0x10\n" + "blt 15f\n" + "14:" // Tail row loop: Column loop + "ldr q17, [x28], #0x10\n" + "sub x19, x19, #0x10\n" + "ldr q25, [x26], #0x10\n" + "cmp x19, #0x10\n" + "ldr q16, [x25], #0x10\n" + "zip1 v24.8h, v17.8h, v16.8h\n" + "ldr q18, [x28], #0x10\n" + "zip2 v23.8h, v17.8h, v16.8h\n" + "ldr q22, [x26], #0x10\n" + "ldr q17, [x25], #0x10\n" + "zip1 v21.8h, v18.8h, v17.8h\n" + "ldr q16, [x24], #0x10\n" + "zip2 v20.8h, v18.8h, v17.8h\n" + "ldr q19, [x24], #0x10\n" + "zip1 v18.8h, v25.8h, v16.8h\n" + "zip2 v17.8h, v25.8h, v16.8h\n" + "zip1 v16.8h, v24.8h, v18.8h\n" + "str q16, [x27, #0x0]\n" + "zip2 v16.8h, v24.8h, v18.8h\n" + "str q16, [x27, #0x10]\n" + "zip1 v16.8h, v23.8h, v17.8h\n" + "str q16, [x27, #0x20]\n" + "zip2 v16.8h, v23.8h, v17.8h\n" + "str q16, [x27, #0x30]\n" + "zip1 v17.8h, v22.8h, v19.8h\n" + "zip1 v16.8h, v21.8h, v17.8h\n" + "str q16, [x27, #0x40]\n" + "zip2 v16.8h, v21.8h, v17.8h\n" + "str q16, [x27, #0x50]\n" + "zip2 v17.8h, v22.8h, v19.8h\n" + "zip1 v16.8h, v20.8h, v17.8h\n" + "str q16, [x27, #0x60]\n" + "zip2 v16.8h, v20.8h, v17.8h\n" + "str q16, [x27, #0x70]\n" + "add x27, x27, %x[out_stride]\n" + "bge 14b\n" + "15:" // Tail row loop: Column loop skip + "cmp x19, #0x4\n" + "blt 17f\n" + "16:" // Tail row loop: width 4 loop: loop + "ldr d18, [x28], #0x8\n" + "sub x19, x19, #0x4\n" + "ldr d17, [x26], #0x8\n" + "cmp x19, #0x4\n" + "ldr d16, [x25], #0x8\n" + "zip1 v18.8h, v18.8h, v16.8h\n" + "ldr d16, [x24], #0x8\n" + "zip1 v17.8h, v17.8h, v16.8h\n" + "zip1 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0x0]\n" + "zip2 v16.8h, v18.8h, v17.8h\n" + "str q16, [x27, #0x10]\n" + "add x27, x27, #0x20\n" + "bge 16b\n" + "17:" // Tail row loop: width 4 loop: skip + "cmp x19, #0x1\n" + "blt 19f\n" + "18:" // Tail row loop: width 1 loop: loop + "ldr h17, [x28], #0x2\n" + "sub x19, x19, #0x1\n" + "ldr h18, [x26], #0x2\n" + "cmp x19, #0x1\n" + "ldr h16, [x25], #0x2\n" + "zip1 v17.8h, v17.8h, v16.8h\n" + "ldr h16, [x24], #0x2\n" + "zip1 v16.8h, v18.8h, v16.8h\n" + "zip1 v16.8h, v17.8h, v16.8h\n" + "str d16, [x27, #0x0]\n" + "add x27, x27, #0x8\n" + "bge 18b\n" + "19:" // Tail row loop: width 1 loop: skip + "add %x[out], %x[out], #0x80\n" + "cmp %x[height], #0x1\n" + "bge 11b\n" + "20:" // 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", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28" + ); +} + +} // anonymous namespace + +template<> +void Transform<16, 4, true, VLType::None>( + bfloat16 *out, const bfloat16 *in, int stride, int x0, int xmax, int k0, int kmax) +{ + a64_transpose_interleave_16_2x4( + reinterpret_cast(out), + reinterpret_cast(in + k0 * stride + x0), + (xmax-x0) * sizeof(bfloat16) / 2, + stride * sizeof(bfloat16), + (kmax-k0) + ); +} + +#endif -- cgit v1.2.1