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
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
|
/*
* Copyright (c) 2017 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 <cassert>
#include "gemm_common.hpp"
#include "profiler.hpp"
#include "transform.hpp"
#include "mergeresults.hpp"
// Some macros used to decide how much working space to allocate.
// Round allocations up to the next cache line.
#define ALLOC_ROUND 64
#define ROUND_UP(x) ((((x) + ALLOC_ROUND-1) / ALLOC_ROUND) * ALLOC_ROUND)
// Implementation of the GemmCommon abstract class.
//
// This implementation interleaves the source matrices in blocks - good for
// larger matrices.
template<typename strategy, typename To, typename Tr>
class GemmInterleaved : public GemmCommon<To, Tr> {
typedef typename strategy::operand_type Toi;
typedef typename strategy::result_type Tri;
const unsigned int M;
const unsigned int N;
const unsigned int K;
const bool trA;
const bool trB;
const strategy strat;
unsigned int k_block = 0;
unsigned int x_block = 0;
unsigned int Mround = 0;
size_t get_a_working_size() const {
return ROUND_UP(sizeof(Toi) * k_block * Mround);
}
size_t get_b_working_size() const {
return ROUND_UP(sizeof(Toi) * x_block * k_block);
}
size_t get_c_working_size() const {
return ROUND_UP(sizeof(Tri) * x_block * strat.out_height);
}
public:
size_t get_working_size() const override {
return get_a_working_size() + get_b_working_size() + get_c_working_size();
}
GemmInterleaved(const CPUInfo *ci, const unsigned int M, const unsigned int N, const unsigned int K, const bool trA, const bool trB) : M(M), N(N), K(K), trA(trA), trB(trB), strat(ci) {
const unsigned int L1_size = ci->L1_size;
const unsigned int L2_size = ci->L2_size;
// Work out blocking parameters
// k_block: Each iteration will consume (out_width + out_height)
// operands - so how many iterations will fill the L1?
k_block = L1_size / (sizeof(Toi) * (strat.out_width + strat.out_height));
// Needs to be a multiple of the K unroll level.
k_block /= strat.k_unroll;
k_block *= strat.k_unroll;
// Now tune to presented problem size; this is how many blocks we need.
int num_k_blocks = (K + (k_block - 1)) / k_block;
// So divide the space equally into that many blocks.
k_block = (K + num_k_blocks - 1) / num_k_blocks;
// And round UP to the K unroll level required.
k_block = (k_block + strat.k_unroll - 1) / strat.k_unroll;
k_block *= strat.k_unroll;
// x_block: Work out how many rows (of length k_block) will fit in the L2
x_block = L2_size / (sizeof(Toi) * k_block);
// Needs to be a multiple of the kernel output width.
x_block /= strat.out_width;
x_block *= strat.out_width;
// And tune to the presented problem size.
int num_x_blocks = (N + (x_block - 1)) / x_block;
x_block = (N + num_x_blocks - 1) / num_x_blocks;
x_block = (x_block + strat.out_width - 1) / strat.out_width;
x_block *= strat.out_width;
// Work out the rounded size of M - needed for some buffers.
Mround = (M + (strat.out_height - 1)) / strat.out_height;
Mround *= strat.out_height;
}
// Actually execute the GEMM.
void execute(const To *A, const int lda, const To *B, const int ldb, Tr *C, const int ldc, const Tr alpha, const Tr beta, void *working_space) const override {
assert(working_space);
profiler prof;
int8_t *working_space_bytes = reinterpret_cast<int8_t *>(working_space);
intptr_t working_space_int = reinterpret_cast<intptr_t>(working_space_bytes);
size_t diff = 0;
if (working_space_int & 0xF) {
diff = 0x10 - (working_space_int & 0xF);
}
Toi * const a_panel = reinterpret_cast<Toi *>(working_space_bytes + diff);
Toi * const b_panel = reinterpret_cast<Toi *>(working_space_bytes + get_a_working_size() + diff);
Tri * const c_panel = reinterpret_cast<Tri *>(working_space_bytes + get_a_working_size() + get_b_working_size() + diff);
for (unsigned int k0=0; k0<K; k0 += k_block) {
unsigned int kmax = k0 + k_block;
if (kmax > K) kmax = K;
// Figure out how many "K" the kernel will actually process.
int kern_k = ((kmax - k0) + (strat.k_unroll - 1)) / strat.k_unroll;
kern_k *= strat.k_unroll;
prof(PROFILE_PREPA, (M * (kmax-k0) * sizeof(Toi)), [&](void) {
if (trA ^ strategy::A_transpose) {
Transform<strategy::A_interleave, strategy::A_block, true>(a_panel, A, lda, 0, M, k0, kmax);
} else {
Transform<strategy::A_interleave, strategy::A_block, false>(a_panel, A, lda, 0, M, k0, kmax);
}
});
for (unsigned int x0=0; x0<N; x0 += x_block) {
unsigned int xmax = x0 + x_block;
if (xmax > N) xmax = N;
int bblocks = (xmax - x0 + strat.out_width - 1) / strat.out_width;
prof(PROFILE_PREPB, (xmax-x0) * (kmax-k0) * sizeof(Toi), [&](void) {
if (trB ^ strategy::B_transpose) {
Transform<strategy::B_interleave, strategy::B_block, true>(b_panel, B, ldb, x0, xmax, k0, kmax);
} else {
Transform<strategy::B_interleave, strategy::B_block, false>(b_panel, B, ldb, x0, xmax, k0, kmax);
}
});
for (unsigned int y=0; y<M; y+=strat.out_height) {
unsigned int ymax = y + strat.out_height;
if (ymax > M) ymax = M;
prof(PROFILE_KERNEL, (strat.out_height * bblocks * strat.out_width * kern_k), [&](void) { strat.kernel(a_panel + (y * kern_k), b_panel, c_panel, 1, bblocks, kern_k); });
prof(PROFILE_MERGE, (strat.out_height * bblocks * strat.out_width * sizeof(Tr)), [&](void) { MergeResults<strategy::out_width, strategy::out_height>(C, c_panel, ldc, y, ymax, x0, xmax, alpha, (k0==0 ? beta : static_cast<Tr>(1))); });
}
}
}
}
};
|
