Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 1 | /* |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2018 ARM Limited. |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #pragma once |
| 25 | |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 26 | #include <cstddef> |
| 27 | |
| 28 | namespace arm_gemm { |
| 29 | |
| 30 | // Abstract class for the GEMM/GEMV functions. |
| 31 | // |
| 32 | // GEMM implementations may be "native" (never require any input |
| 33 | // permutation), "pretransposed" (require permutation up-front) or require |
| 34 | // working space (permute as they go along). This interface should support |
| 35 | // all of them. |
| 36 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 37 | template<typename To, typename Tr> |
| 38 | class GemmCommon { |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 39 | protected: |
| 40 | const To *_Aptr=nullptr; |
| 41 | int _lda=0; |
| 42 | const To *_Bptr=nullptr; |
| 43 | int _ldb=0; |
| 44 | Tr *_Cptr=nullptr; |
| 45 | int _ldc=0; |
| 46 | |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 47 | public: |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 48 | /* Pass in the pointers to the arrays to be operated on and their |
| 49 | * strides. This has a default implementation that just captures them |
| 50 | * all in protected members. If B is pretransposed (see below) then the |
| 51 | * settings for B here are ignored. */ |
| 52 | virtual void set_arrays(const To *A, const int lda, const To *B, const int ldb, Tr *C, const int ldc) { |
| 53 | _Aptr = A; |
| 54 | _lda = lda; |
| 55 | _Bptr = B; |
| 56 | _ldb = ldb; |
| 57 | _Cptr = C; |
| 58 | _ldc = ldc; |
| 59 | } |
| 60 | |
| 61 | /* For threading, we divide the work into some number of units and work |
| 62 | * out internally what unit corresponds to what work. This returns the |
| 63 | * total number of units. */ |
| 64 | virtual unsigned int get_window_size() const = 0; |
| 65 | |
| 66 | /* The maximum thread count is specified when the GEMM is created. Some |
| 67 | * implementations need to know how many threads will actually run in |
| 68 | * order to work properly. |
| 69 | * |
| 70 | * In some cases, after creating the GEMM the number of threads needs to |
| 71 | * be reduced (e.g. not enough work to split across threads). This |
| 72 | * method allows the number of actual threads to be run to be set (must |
| 73 | * be equal or lower). |
| 74 | * |
| 75 | * This has an empty default implementation, as GEMMs which don't care |
| 76 | * about thread count can safely ignore this. |
| 77 | */ |
| 78 | virtual void set_nthreads(int nthreads) { }; |
| 79 | |
| 80 | /* Actually do the work. Provide a threadid to index any per-thread |
| 81 | * buffers, and a start/end range to indicate which work to do. */ |
| 82 | virtual void execute(unsigned int start, unsigned int end, int threadid) = 0; |
| 83 | |
| 84 | /*** Working space interface (optional) ***/ |
| 85 | /* Total number of bytes of temporary working space needed. If zero, it's not necessary to call set_working_space(). */ |
| 86 | virtual size_t get_working_size() const { return 0; } |
| 87 | /* Provide working space buffer - the void * passed in must remain allocated for the duration of any execute calls. */ |
| 88 | virtual void set_working_space(void *) { }; |
| 89 | |
| 90 | /*** "Pretransposed" interface (optional) ***/ |
| 91 | /* Is this object set up for pretranspose? If so, pretranspose_array() needs to be called before execute(); */ |
| 92 | virtual bool B_is_pretransposed() const { return false; } |
| 93 | /* Does pretranspose still need to be done? */ |
| 94 | virtual bool B_pretranspose_required() const { return false; } |
| 95 | /* Total number of bytes of space needed for pretransposed arrays. */ |
| 96 | virtual size_t get_B_pretransposed_array_size() const { return 0; } |
| 97 | /* Perform pretranspose - the void * passed in must remain allocated for the duration of any execute calls. */ |
| 98 | virtual void pretranspose_B_array(void *buffer, const To *, const int) { }; |
| 99 | |
| 100 | // Destructor |
Moritz Pflanzer | beabe3b | 2017-08-31 14:56:32 +0100 | [diff] [blame] | 101 | virtual ~GemmCommon() { } |
| 102 | }; |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 103 | |
| 104 | } // namespace arm_gemm |