1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
|
/*
* Copyright (c) 2017-2020 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
#include <stdio.h>
#include "arm_gemm.hpp"
#include "ndrange.hpp"
#ifdef CYCLE_PROFILING
#include "profiler.hpp"
#endif
namespace arm_gemm {
// Implementation of the GemmCommon abstract class.
//
// This is implementation is for native GEMM with no transposition.
//
// By default the source data is used in-place, but if type conversion is
// needed we need to allocate working space (CURRENTLY NOT IMPLEMENTED).
template<typename strategy, typename To, typename Tr>
class GemmNative : public GemmCommon<To, Tr> {
typedef typename strategy::operand_type Toi;
typedef typename strategy::result_type Tri;
const unsigned int _Msize;
const unsigned int _Nsize;
const unsigned int _Ksize;
const unsigned int _nbatches;
const unsigned int _nmultis;
const Activation _act;
const CPUInfo * const _ci;
const unsigned int _k_block;
const unsigned int _n_block;
const NDRange<4> _window_range;
static unsigned int compute_k_block(const GemmArgs &args) {
return args._Ksize;
}
static unsigned int compute_n_block(const GemmArgs &args) {
if ((args._cfg != nullptr) && args._cfg->outer_block_size > 0) {
return args._cfg->outer_block_size;
} else {
return args._Nsize;
}
}
public:
GemmNative(GemmNative &) = delete;
GemmNative & operator= (GemmNative &) = delete;
GemmNative(const GemmArgs &args)
: _Msize(args._Msize), _Nsize(args._Nsize), _Ksize(args._Ksize),
_nbatches(args._nbatches), _nmultis(args._nmulti),
_act(args._act), _ci(args._ci),
_k_block(compute_k_block(args)), _n_block(compute_n_block(args)),
_window_range(iceildiv(_Msize, strategy::out_height()), _nbatches, iceildiv(_Nsize, _n_block), _nmultis) { }
// Window is amount per multi multiplied by total number of multis.
ndrange_t get_window_size() const override {
return { _window_range.total_size(), 1u, 1u, 1u, 1u, 1u };
}
// Native GEMMs can always be dynamically scheduled (whether requested or not)
bool supports_dynamic_scheduling() const override {
return true;
}
// Actually execute the GEMM.
void execute_1d(unsigned int start, unsigned int end, int) {
#ifdef CYCLE_PROFILING
profiler prof;
#endif
strategy strat(_ci);
static_assert(std::is_same<To, Toi>::value, "gemm_native: Operand types must be the same.");
static_assert(std::is_same<Tr, Tri>::value, "gemm_native: Result types must be the same.");
auto p = _window_range.iterator(start, end);
if (p.done()) {
return;
}
do {
unsigned int y0 = p.dim(0) * strategy::out_height();
unsigned int ymax = std::min(p.dim0_max() * strategy::out_height(), _Msize);
unsigned int batch = p.dim(1);
unsigned int n0 = p.dim(2) * _n_block;
unsigned int nmax = std::min(n0 + _n_block, _Nsize);
unsigned int multi = p.dim(3);
#ifdef CYCLE_PROFILING
auto p = prof.ScopedProfiler(PROFILE_KERNEL, (ymax-y0) * (nmax - n0) * _Ksize);
#endif
strat.kernel(this->_Aptr + (multi * this->_A_multi_stride) + (batch * this->_A_batch_stride) + (y0 * this->_lda), this->_lda,
this->_Bptr + (multi * this->_B_multi_stride) + n0, this->_ldb,
this->_Cptr + (multi * this->_C_multi_stride) + (batch * this->_C_batch_stride) + (y0 * this->_ldc) + n0, this->_ldc,
(ymax-y0), (nmax-n0), _Ksize,
(strategy::supports_bias() && this->_bias) ? this->_bias + (multi * this->_bias_multi_stride) + n0 : nullptr,
_act, false);
// Add bias externally if needed
if (!strategy::supports_bias() && this->_bias) {
bias_adder(this->_Cptr + (multi * this->_C_multi_stride) + (batch * this->_C_batch_stride) + (y0 * this->_ldc) + n0, this->_ldc,
this->_bias + (multi * this->_bias_multi_stride) + n0,
(ymax - y0), (nmax - n0));
}
} while (p.next_dim1());
}
//Execute
void execute(const ndcoord_t& work_range, const ndcoord_t& thread_locator, int threadid) override {
UNUSED(thread_locator);
const auto start = work_range.get_position(0);
const auto stop = work_range.get_position_end(0);
execute_1d(start, stop, threadid);
}
};
} // namespace arm_gemm
|
