From 0eb9cfbe6e78f80008164cb0ee18afa09a7fe4eb Mon Sep 17 00:00:00 2001 From: Renato Arantes Date: Thu, 23 Nov 2023 11:12:51 +0000 Subject: [ONCPUML-1387] Add ACL based reorder for f32 to bf16 data type conversion. The reorders supported at the moment are: ab->BA4b4a ab->BA8b4a Co-Authored-By: David Mansell Change-Id: Ic466465629ce3bcdcee0089e251485b79b60e1f3 Signed-off-by: Renato Arantes Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10775 Comments-Addressed: Arm Jenkins Reviewed-by: Jakub Sujak Tested-by: Arm Jenkins Benchmark: Arm Jenkins --- .../a64_transpose_interleave_4_2x4_fp32bf16.hpp | 346 +++++++++++++++++++++ 1 file changed, 346 insertions(+) create mode 100644 src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_2x4_fp32bf16.hpp (limited to 'src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_2x4_fp32bf16.hpp') diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_2x4_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_2x4_fp32bf16.hpp new file mode 100644 index 0000000000..98200c50c5 --- /dev/null +++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_4_2x4_fp32bf16.hpp @@ -0,0 +1,346 @@ +/* + * Copyright (c) 2024 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 + +#if defined(__aarch64__) + +namespace { + +void a64_transpose_interleave_4_2x4_fp32bf16(bfloat16 *out, const float *in, size_t width, size_t in_stride, size_t height) +{ + float *pad_row = reinterpret_cast(alloca(width * sizeof(float))); + + if (height % 4) { + memset(pad_row, 0, width * sizeof(float)); + } + + size_t out_stride = 4 * roundup(height, 4) * sizeof(bfloat16); + + __asm__ __volatile__( + "cmp %x[height], #0x8\n" + "blt 8f\n" + "1:" // Main row loop: Head + "mov x9, %x[in]\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" + "add x24, x25, %x[in_stride]\n" + "cmp x28, #0x8\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" + "blt 3f\n" + "2:" // Main row loop: Unroll column loop + "ldr q19, [x9], #0x10\n" + "ldr q18, [x26], #0x10\n" + "sub x28, x28, #0x8\n" + "ldr q17, [x25], #0x10\n" + "ldr q16, [x24], #0x10\n" + "cmp x28, #0x8\n" + "ldr q1, [x23], #0x10\n" + "ldr q0, [x22], #0x10\n" + "ldr q31, [x21], #0x10\n" + "ldr q24, [x20], #0x10\n" + "ldr q23, [x9], #0x10\n" + "ldr q22, [x26], #0x10\n" + "zip1 v30.4s, v19.4s, v17.4s\n" + "zip1 v29.4s, v18.4s, v16.4s\n" + "ldr q21, [x25], #0x10\n" + "ldr q20, [x24], #0x10\n" + "zip2 v28.4s, v19.4s, v17.4s\n" + "zip2 v27.4s, v18.4s, v16.4s\n" + "ldr q19, [x23], #0x10\n" + "ldr q18, [x22], #0x10\n" + "zip1 v26.4s, v1.4s, v31.4s\n" + "zip1 v25.4s, v0.4s, v24.4s\n" + "ldr q17, [x21], #0x10\n" + "ldr q16, [x20], #0x10\n" + "zip2 v8.4s, v1.4s, v31.4s\n" + "zip2 v24.4s, v0.4s, v24.4s\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" + "zip2 v0.4s, v18.4s, v16.4s\n" + "zip1 v23.4s, v30.4s, v29.4s\n" + "zip1 v22.4s, v28.4s, v27.4s\n" + "zip1 v21.4s, v26.4s, v25.4s\n" + "zip1 v20.4s, v8.4s, v24.4s\n" + "zip1 v19.4s, v7.4s, v6.4s\n" + "zip1 v18.4s, v5.4s, v4.4s\n" + "zip1 v17.4s, v3.4s, v2.4s\n" + "zip1 v16.4s, v1.4s, v0.4s\n" + ".inst 0x0ea16aff // bfcvtn v31.4h, v23.4s\n" + "zip2 v30.4s, v30.4s, v29.4s\n" + ".inst 0x0ea16add // bfcvtn v29.4h, v22.4s\n" + "zip2 v28.4s, v28.4s, v27.4s\n" + ".inst 0x0ea16abb // bfcvtn v27.4h, v21.4s\n" + "zip2 v26.4s, v26.4s, v25.4s\n" + ".inst 0x0ea16a99 // bfcvtn v25.4h, v20.4s\n" + "zip2 v24.4s, v8.4s, v24.4s\n" + ".inst 0x0ea16a77 // bfcvtn v23.4h, v19.4s\n" + "zip2 v22.4s, v7.4s, v6.4s\n" + ".inst 0x0ea16a55 // bfcvtn v21.4h, v18.4s\n" + "zip2 v20.4s, v5.4s, v4.4s\n" + ".inst 0x0ea16a33 // bfcvtn v19.4h, v17.4s\n" + "zip2 v18.4s, v3.4s, v2.4s\n" + ".inst 0x0ea16a11 // bfcvtn v17.4h, v16.4s\n" + "zip2 v16.4s, v1.4s, v0.4s\n" + ".inst 0x4ea16bdf // bfcvtn2 v31.8h, v30.4s\n" + ".inst 0x4ea16b9d // bfcvtn2 v29.8h, v28.4s\n" + ".inst 0x4ea16b5b // bfcvtn2 v27.8h, v26.4s\n" + ".inst 0x4ea16b19 // bfcvtn2 v25.8h, v24.4s\n" + ".inst 0x4ea16ad7 // bfcvtn2 v23.8h, v22.4s\n" + ".inst 0x4ea16a95 // bfcvtn2 v21.8h, v20.4s\n" + "str q31, [x27, #0x0]\n" + "str q29, [x27, #0x10]\n" + ".inst 0x4ea16a53 // bfcvtn2 v19.8h, v18.4s\n" + ".inst 0x4ea16a11 // bfcvtn2 v17.8h, v16.4s\n" + "str q27, [x27, #0x20]\n" + "str q25, [x27, #0x30]\n" + "add x27, x27, %x[out_stride]\n" + "str q23, [x27, #0x0]\n" + "str q21, [x27, #0x10]\n" + "str q19, [x27, #0x20]\n" + "str q17, [x27, #0x30]\n" + "add x27, x27, %x[out_stride]\n" + "bge 2b\n" + "3:" // Main row loop: Unroll column loop skip + "cmp x28, #0x4\n" + "blt 5f\n" + "4:" // Main row loop: Column loop + "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" + "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 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, [x27, #0x0]\n" + "str q20, [x27, #0x10]\n" + "str q19, [x27, #0x20]\n" + "str q17, [x27, #0x30]\n" + "add x27, x27, %x[out_stride]\n" + "bge 4b\n" + "5:" // Main row loop: Column loop skip + "cbz x28, 7f\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" + "6:" // Main row loop: width 1 loop: loop + "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" + ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n" + ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n" + "str d17, [x27, #0x0]\n" + "str d16, [x27, #0x20]\n" + "add x27, x27, #0x8\n" + "bge 6b\n" + "7:" // Main row loop: odd col skip + "cmp %x[height], #0x8\n" + "add %x[out], %x[out], #0x40\n" + "bge 1b\n" + "cbz %x[height], 16f\n" + "8:" // Main loop skip + "9:" // Tail row loop: Head + "mov x9, %x[in]\n" + "mov x20, %x[width]\n" + "cmp %x[height], #0x3\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" + "sub %x[height], %x[height], #0x4\n" + "csel x26, x26, %x[pad_row], GT\n" + "cmp x20, #0x8\n" + "blt 11f\n" + "10:" // Tail row loop: Unroll column loop + "ldr q25, [x9], #0x10\n" + "ldr q24, [x26], #0x10\n" + "sub x20, x20, #0x8\n" + "ldr q21, [x25], #0x10\n" + "ldr q20, [x24], #0x10\n" + "cmp x20, #0x8\n" + "ldr q23, [x9], #0x10\n" + "ldr q19, [x26], #0x10\n" + "ldr q18, [x25], #0x10\n" + "ldr q17, [x24], #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" + "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 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, [x27, #0x0]\n" + "str q20, [x27, #0x10]\n" + "add x27, x27, %x[out_stride]\n" + "str q19, [x27, #0x0]\n" + "str q17, [x27, #0x10]\n" + "add x27, x27, %x[out_stride]\n" + "bge 10b\n" + "11:" // Tail row loop: Unroll column loop skip + "cmp x20, #0x4\n" + "blt 13f\n" + "12:" // Tail row loop: Column 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" + "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 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, %x[out_stride]\n" + "bge 12b\n" + "13:" // Tail row loop: Column loop skip + "cbz x20, 15f\n" + "movi v16.16b, #0x0\n" + "str q16, [x27, #0x0]\n" + "str q16, [x27, #0x10]\n" + "14:" // Tail row loop: width 1 loop: loop + "ldr s19, [x9], #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" + "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, [x27, #0x0]\n" + "add x27, x27, #0x8\n" + "bge 14b\n" + "15:" // Tail row loop: odd col skip + "cmp %x[height], #0x1\n" + "add %x[out], %x[out], #0x20\n" + "bge 9b\n" + "16:" // 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", "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" + ); +} + +} // anonymous namespace +template<> +void Transform<4, 4, true, VLType::None>( + bfloat16 *out, const float *in, int stride, int x0, int xmax, int k0, int kmax) +{ + a64_transpose_interleave_4_2x4_fp32bf16( + out, + in + k0 * stride + x0, + (xmax-x0), + stride * sizeof(float), + (kmax-k0) + ); +} + + +#endif // defined(__aarch64__) -- cgit v1.2.1