EIC code displayed by LXR |
|
|
|
|||
File indexing completed on 2026-05-10 08:48:20
0001 //===- MatmulOptimizer.h -------------------------------------------------===// 0002 // 0003 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 0004 // See https://llvm.org/LICENSE.txt for license information. 0005 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 0006 // 0007 //===----------------------------------------------------------------------===// 0008 0009 #ifndef POLLY_MATMULOPTIMIZER_H 0010 #define POLLY_MATMULOPTIMIZER_H 0011 0012 #include "isl/isl-noexceptions.h" 0013 0014 namespace llvm { 0015 class TargetTransformInfo; 0016 } 0017 0018 namespace polly { 0019 class Dependences; 0020 0021 /// Apply the BLIS matmul optimization pattern if possible. 0022 /// 0023 /// Make the loops containing the matrix multiplication be the innermost 0024 /// loops and apply the BLIS matmul optimization pattern. BLIS implements 0025 /// gemm as three nested loops around a macro-kernel, plus two packing 0026 /// routines. The macro-kernel is implemented in terms of two additional 0027 /// loops around a micro-kernel. The micro-kernel is a loop around a rank-1 0028 /// (i.e., outer product) update. 0029 /// 0030 /// For a detailed description please see [1]. 0031 /// 0032 /// The order of the loops defines the data reused in the BLIS implementation 0033 /// of gemm ([1]). In particular, elements of the matrix B, the second 0034 /// operand of matrix multiplication, are reused between iterations of the 0035 /// innermost loop. To keep the reused data in cache, only elements of matrix 0036 /// A, the first operand of matrix multiplication, should be evicted during 0037 /// an iteration of the innermost loop. To provide such a cache replacement 0038 /// policy, elements of the matrix A can, in particular, be loaded first and, 0039 /// consequently, be least-recently-used. 0040 /// 0041 /// In our case matrices are stored in row-major order instead of 0042 /// column-major order used in the BLIS implementation ([1]). It affects only 0043 /// on the form of the BLIS micro kernel and the computation of its 0044 /// parameters. In particular, reused elements of the matrix B are 0045 /// successively multiplied by specific elements of the matrix A. 0046 /// 0047 /// Refs.: 0048 /// [1] - Analytical Modeling is Enough for High Performance BLIS 0049 /// Tze Meng Low, Francisco D Igual, Tyler M Smith, Enrique S Quintana-Orti 0050 /// Technical Report, 2014 0051 /// http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf 0052 /// 0053 /// @see ScheduleTreeOptimizer::createMicroKernel 0054 /// @see ScheduleTreeOptimizer::createMacroKernel 0055 /// @see getMicroKernelParams 0056 /// @see getMacroKernelParams 0057 /// 0058 /// TODO: Implement the packing transformation. 0059 /// 0060 /// @param Node The node that contains a band to be optimized. The node 0061 /// is required to successfully pass 0062 /// ScheduleTreeOptimizer::isMatrMultPattern. 0063 /// @param TTI Target Transform Info. 0064 /// @param D The dependencies. 0065 /// 0066 /// @returns The transformed schedule or nullptr if the optimization 0067 /// cannot be applied. 0068 isl::schedule_node 0069 tryOptimizeMatMulPattern(isl::schedule_node Node, 0070 const llvm::TargetTransformInfo *TTI, 0071 const Dependences *D); 0072 0073 } // namespace polly 0074 #endif // POLLY_MATMULOPTIMIZER_H
| [ Source navigation ] | [ Diff markup ] | [ Identifier search ] | [ general search ] |
|
This page was automatically generated by the 2.3.7 LXR engine. The LXR team |
|