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_64.hpp | 255 +++++++++++++++++++++ 1 file changed, 255 insertions(+) create mode 100644 src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_64.hpp (limited to 'src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_64.hpp') diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_64.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_64.hpp new file mode 100644 index 0000000000..c341b315aa --- /dev/null +++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_64.hpp @@ -0,0 +1,255 @@ +/* + * 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_64(uint16_t *out, const uint16_t *in, size_t width, size_t in_stride, size_t height) +{ + size_t out_stride = 32 * height * sizeof(uint16_t); + + __asm__ __volatile__( + "cmp %x[height], #0x4\n" + "blt 10f\n" + "1:" // Main row loop: Head + "mov x24, %x[in]\n" + "mov x23, %x[out]\n" + "add x22, x24, %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], #0x4\n" + "mov x19, %x[width]\n" + "cmp x19, #0x20\n" + "blt 3f\n" + "2:" // Main row loop: Column loop + "ldr q31, [x24], #0x10\n" + "sub x19, x19, #0x20\n" + "ldr q30, [x22], #0x10\n" + "cmp x19, #0x20\n" + "ldr q29, [x21], #0x10\n" + "ldr q28, [x20], #0x10\n" + "ldr q27, [x24], #0x10\n" + "ldr q26, [x22], #0x10\n" + "ldr q25, [x21], #0x10\n" + "ldr q24, [x20], #0x10\n" + "ldr q23, [x24], #0x10\n" + "ldr q22, [x22], #0x10\n" + "ldr q21, [x21], #0x10\n" + "ldr q20, [x20], #0x10\n" + "ldr q19, [x24], #0x10\n" + "ldr q18, [x22], #0x10\n" + "ldr q17, [x21], #0x10\n" + "ldr q16, [x20], #0x10\n" + "str q31, [x23, #0x0]\n" + "str q27, [x23, #0x10]\n" + "str q23, [x23, #0x20]\n" + "str q19, [x23, #0x30]\n" + "str q30, [x23, #0x40]\n" + "str q26, [x23, #0x50]\n" + "str q22, [x23, #0x60]\n" + "str q18, [x23, #0x70]\n" + "str q29, [x23, #0x80]\n" + "str q25, [x23, #0x90]\n" + "str q21, [x23, #0xa0]\n" + "str q17, [x23, #0xb0]\n" + "str q28, [x23, #0xc0]\n" + "str q24, [x23, #0xd0]\n" + "str q20, [x23, #0xe0]\n" + "str q16, [x23, #0xf0]\n" + "add x23, x23, %x[out_stride]\n" + "bge 2b\n" + "3:" // Main row loop: Column loop skip + "cmp x19, #0x10\n" + "blt 5f\n" + "4:" // Main row loop: width 16 loop: loop + "ldr q23, [x24], #0x10\n" + "sub x19, x19, #0x10\n" + "ldr q22, [x22], #0x10\n" + "cmp x19, #0x10\n" + "ldr q21, [x21], #0x10\n" + "ldr q20, [x20], #0x10\n" + "ldr q19, [x24], #0x10\n" + "ldr q18, [x22], #0x10\n" + "ldr q17, [x21], #0x10\n" + "ldr q16, [x20], #0x10\n" + "str q23, [x23, #0x0]\n" + "str q19, [x23, #0x10]\n" + "str q22, [x23, #0x40]\n" + "str q18, [x23, #0x50]\n" + "str q21, [x23, #0x80]\n" + "str q17, [x23, #0x90]\n" + "str q20, [x23, #0xc0]\n" + "str q16, [x23, #0xd0]\n" + "add x23, x23, #0x20\n" + "bge 4b\n" + "5:" // Main row loop: width 16 loop: skip + "cmp x19, #0x4\n" + "blt 7f\n" + "6:" // Main row loop: width 4 loop: loop + "ldr d19, [x24], #0x8\n" + "sub x19, x19, #0x4\n" + "ldr d18, [x22], #0x8\n" + "cmp x19, #0x4\n" + "ldr d17, [x21], #0x8\n" + "ldr d16, [x20], #0x8\n" + "str d19, [x23, #0x0]\n" + "str d18, [x23, #0x40]\n" + "str d17, [x23, #0x80]\n" + "str d16, [x23, #0xc0]\n" + "add x23, x23, #0x8\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 h19, [x24], #0x2\n" + "sub x19, x19, #0x1\n" + "ldr h18, [x22], #0x2\n" + "cmp x19, #0x1\n" + "ldr h17, [x21], #0x2\n" + "ldr h16, [x20], #0x2\n" + "str h19, [x23, #0x0]\n" + "str h18, [x23, #0x40]\n" + "str h17, [x23, #0x80]\n" + "str h16, [x23, #0xc0]\n" + "add x23, x23, #0x2\n" + "bge 8b\n" + "9:" // Main row loop: width 1 loop: skip + "add %x[out], %x[out], #0x100\n" + "cmp %x[height], #0x4\n" + "bge 1b\n" + "cbz %x[height], 20f\n" + "10:" // Main loop skip + + "11:" // Tail row loop: Head + "mov x24, %x[in]\n" + "mov x23, %x[out]\n" + "add %x[in], x24, %x[in_stride]\n" + "sub %x[height], %x[height], #0x1\n" + "mov x19, %x[width]\n" + "cmp x19, #0x20\n" + "blt 13f\n" + "12:" // Tail row loop: Column loop + "ldr q19, [x24], #0x10\n" + "sub x19, x19, #0x20\n" + "cmp x19, #0x20\n" + "ldr q18, [x24], #0x10\n" + "ldr q17, [x24], #0x10\n" + "ldr q16, [x24], #0x10\n" + "str q19, [x23, #0x0]\n" + "str q18, [x23, #0x10]\n" + "str q17, [x23, #0x20]\n" + "str q16, [x23, #0x30]\n" + "add x23, x23, %x[out_stride]\n" + "bge 12b\n" + "13:" // Tail row loop: Column loop skip + "cmp x19, #0x10\n" + "blt 15f\n" + "14:" // Tail row loop: width 16 loop: loop + "ldr q17, [x24], #0x10\n" + "sub x19, x19, #0x10\n" + "cmp x19, #0x10\n" + "ldr q16, [x24], #0x10\n" + "str q17, [x23, #0x0]\n" + "str q16, [x23, #0x10]\n" + "add x23, x23, #0x20\n" + "bge 14b\n" + "15:" // Tail row loop: width 16 loop: skip + "cmp x19, #0x4\n" + "blt 17f\n" + "16:" // Tail row loop: width 4 loop: loop + "ldr d16, [x24], #0x8\n" + "sub x19, x19, #0x4\n" + "cmp x19, #0x4\n" + "str d16, [x23, #0x0]\n" + "add x23, x23, #0x8\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 h16, [x24], #0x2\n" + "sub x19, x19, #0x1\n" + "cmp x19, #0x1\n" + "str h16, [x23, #0x0]\n" + "add x23, x23, #0x2\n" + "bge 18b\n" + "19:" // Tail row loop: width 1 loop: skip + "add %x[out], %x[out], #0x40\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), [width] "r" (width) + : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x19", "x20", "x21", "x22", "x23", "x24" + ); +} + +} // anonymous namespace + +template<> +void Transform<16, 1, true, VLType::None>( + float *out, const float *in, int stride, int x0, int xmax, int k0, int kmax) +{ + a64_transpose_interleave_64( + reinterpret_cast(out), + reinterpret_cast(in + k0 * stride + x0), + (xmax-x0) * sizeof(float) / 2, + stride * sizeof(float), + (kmax-k0) + ); +} + +template<> +void Transform<32, 1, true, VLType::None>( + __fp16 *out, const __fp16 *in, int stride, int x0, int xmax, int k0, int kmax) +{ + a64_transpose_interleave_64( + reinterpret_cast(out), + reinterpret_cast(in + k0 * stride + x0), + (xmax-x0) * sizeof(__fp16) / 2, + stride * sizeof(__fp16), + (kmax-k0) + ); +} + +template<> +void Transform<32, 1, true, VLType::None>( + uint16_t *out, const uint16_t *in, int stride, int x0, int xmax, int k0, int kmax) +{ + a64_transpose_interleave_64( + reinterpret_cast(out), + reinterpret_cast(in + k0 * stride + x0), + (xmax-x0) * sizeof(uint16_t) / 2, + stride * sizeof(uint16_t), + (kmax-k0) + ); +} + +#endif -- cgit v1.2.1