xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_gemm/std_transforms_fixed.hpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2018 Arm Limited.
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 
26 #include "convolver.hpp"
27 #include "mergeresults.hpp"
28 #include "transform.hpp"
29 #include "interleave_indirect.hpp"
30 
31 namespace arm_gemm {
32 
33 /*
34  * Define "standard" transforms for the blocked GEMMs with fixed vector
35  * length.
36  *
37  * This assumes that A is interleaved 'height' ways, B is interleaved
38  * 'width' ways and transposed, and that the merge needs to work in 'height'
39  * x 'width' blocks.
40  *
41  * The optional 'block' parameter is for kernels using dot-product type
42  * instructions like UDOT and SDOT.
43  */
44 template<typename TOperand, typename TResult, unsigned int height, unsigned int width, unsigned int block=1, bool integrate_sums=false>
45 class StdTransformsFixed
46 {
47 public:
48     template<typename TIn>
PrepareA(TOperand * out,const TIn * in,const int stride,const int y0,const int ymax,const int k0,const int kmax,int32_t row_sum_multiplier) const49     void PrepareA(TOperand *out, const TIn *in, const int stride, const int y0,
50                   const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) const {
51         Interleave<height, block, VLType::None>(out, in, stride, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
52     }
53 
54     template<typename TIn>
PrepareA_indirect(TOperand * out,const TIn * const * const * ptr,size_t stringlen,size_t rounded_stringlen,const int y0,const int ymax,const int k0,const int kmax,int32_t row_sum_multiplier)55     void PrepareA_indirect(TOperand *out, const TIn * const * const *ptr, size_t stringlen, size_t rounded_stringlen, const int y0,
56                            const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) {
57         IndirectInterleave<height, block, VLType::None>(out, ptr, stringlen, rounded_stringlen, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
58     }
59 
60     template<typename TIn>
PrepareA_convolution(TOperand * out,const TIn * ptr,size_t stride,const convolver<TIn> & conv,size_t rounded_stringlen,const int y0,const int ymax,const int k0,const int kmax,int32_t row_sum_multiplier)61     void PrepareA_convolution(TOperand *out, const TIn *ptr, size_t stride, const convolver<TIn> &conv, size_t rounded_stringlen,
62                               const int y0, const int ymax, const int k0, const int kmax, int32_t row_sum_multiplier) {
63         ConvolutionInterleave<height, block, VLType::None>(out, ptr, stride, conv, rounded_stringlen, y0, ymax, k0, kmax, integrate_sums, row_sum_multiplier);
64     }
65 
66     template<typename TIn>
PrepareB(TOperand * out,const TIn * in,const int stride,const int x0,const int xmax,const int k0,const int kmax) const67     void PrepareB(TOperand *out, const TIn *in, const int stride, const int x0,
68                   const int xmax, const int k0, const int kmax) const {
69         Transform<width, block,  true>(out, in, stride, x0, xmax, k0, kmax);
70     }
71 
72     template<typename TOut>
Merge(TOut * out,const TResult * in,int stride,int y0,int ymax,int x0,int xmax,const TOut * bias,const Activation act,bool append) const73     void Merge(TOut *out, const TResult *in, int stride, int y0, int ymax, int x0, int xmax, const TOut *bias, const Activation act, bool append) const {
74         MergeResults<width, height>(out, in, stride, y0, ymax, x0, xmax, bias, act, append);
75     }
76 };
77 
78 } // namespace arm_gemm
79