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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
|
/* Ergo, version 3.8, a program for linear scaling electronic structure
* calculations.
* Copyright (C) 2019 Elias Rudberg, Emanuel H. Rubensson, Pawel Salek,
* and Anastasia Kruchinina.
*
* This program is free software: you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program. If not, see <http://www.gnu.org/licenses/>.
*
* Primary academic reference:
* Ergo: An open-source program for linear-scaling electronic structure
* calculations,
* Elias Rudberg, Emanuel H. Rubensson, Pawel Salek, and Anastasia
* Kruchinina,
* SoftwareX 7, 107 (2018),
* <http://dx.doi.org/10.1016/j.softx.2018.03.005>
*
* For further information about Ergo, see <http://www.ergoscf.org>.
*/
/** @file mm_kernel_inner_sse2_A.h
*
* @brief Templates for efficient gemm kernels. For architectures
* with SSE2 or higher.
*
* @author Emanuel H. Rubensson
* @date 2009
*
*/
#ifndef MM_KERNEL_INNER_SSE2_A_H
#define MM_KERNEL_INNER_SSE2_A_H
#include "common.h"
#include "vector_intrin.h"
/** Matrix multiplication template for architectures with SSE2 or higher
* and compilers that support C++ intrinsics for access to SSE
* instructions.
*
* Choice of template parameters:
* - T_M and T_N should be chosen so that the T_M x T_N matrix C
* fits in registers. For example T_M == T_N == 4
* - T_K should be chosen so that the generated code fits in L1 instruction cache.
* For example T_K == 128.
* - T_real and T_reg must go together.
* Example:
* - <T_real, T_reg> == <double, __m128d>
* - <T_real, T_reg> == <float, __m128>
*
* The public typedefs and static members specify how the matrices must be
* stored.
*
*/
template<typename T_real, typename T_reg, int T_M, int T_N, int T_K>
class MM_kernel_inner_sse2_A {
public:
typedef T_real real; /**< Real number type (usually float or double) */
static int const M = T_M; /**< Number of rows of A and C. */
static int const N = T_N; /**< Number of columns of B and C. */
static int const K = T_K; /**< Number of columns of A and rows of B. */
protected:
static int const floats_per_register = ( sizeof(T_reg) / sizeof(real) );
/**< Number of real numbers that fit in one register. */
private:
/** Template for packing of matrix elements. */
template<int T_ROWS_kernel, int T_COLS_kernel, typename T_ordering_kernel, int T_repetitions>
class Pack;
public:
typedef Pack< M, K, Ordering_col_wise, 1 > Pack_type_A; /**< Type that can (should) be used to pack A. */
typedef Pack< K, N, Ordering_row_wise, floats_per_register > Pack_type_B; /**< Type that can (should) be used to pack B. */
typedef Pack< M, N, Ordering_col_wise, 1 > Pack_type_C; /**< Type that can (should) be used to pack C. */
// Consider passing derived class from std::vector as arguments
// that have compile time constant length that can be checked at
// compile time using static asserts
/** Executes the matrix-matrix multiply C += A B with the three matrices A, B, and C
* stored according to the static members and typedefs of this class.
*/
static void exec( real const * const * const A,
real const * const * const B,
real * const C,
int const i = 1,
int const offset_A = 0,
int const offset_B = 0,
int const offset_C = 0 );
template<int T_offset_A, int T_offset_B, int T_offset_C>
static void exec( real const * const * const A,
real const * const * const B,
real * const C,
int const i = 1 );
protected:
template<int T_loop_index, int T_end>
struct Loop {
static inline void ALWAYS_INLINE set_to_zero( Vector_intrin<real, T_reg> * X_reg ) {
X_reg[T_loop_index].set_to_zero();
Loop<T_loop_index+1, T_end>::set_to_zero( X_reg );
}
static inline void ALWAYS_INLINE inner( int const row_A_reg, // == row_C_reg
int const row_B,
Vector_intrin<real, T_reg> const & A_reg,
Vector_intrin<real, T_reg> * C_reg,
real const * B_packed ) {
Vector_intrin<real, T_reg> B_reg;
B_reg.load_p( &B_packed[row_B * T_N * floats_per_register +
T_loop_index * floats_per_register] );
B_reg *= A_reg;
C_reg[row_A_reg + T_loop_index * T_M / floats_per_register] += B_reg;
Loop<T_loop_index+1, T_end>::inner( row_A_reg, row_B,
A_reg, C_reg,
B_packed );
}
static inline void ALWAYS_INLINE middle( int const col_A, // == row_B
Vector_intrin<real, T_reg> * C_reg,
real const * A,
real const * B_packed ) {
Vector_intrin<real, T_reg> A_reg;
A_reg.load_p( &A[col_A * T_M + T_loop_index * floats_per_register] );
// Loop over cols of B
Loop<0, T_N>::inner( T_loop_index, // == row_A_reg == row_C_reg
col_A, // == row_B
A_reg,
C_reg,
B_packed );
Loop<T_loop_index+1, T_end>::middle( col_A, C_reg, A, B_packed );
}
static inline void ALWAYS_INLINE outer( int const start_i,
Vector_intrin<real, T_reg> * C_reg,
real const * A,
real const * B_packed ) {
// Loop over (register) rows of A and C
Loop<0, T_M/floats_per_register>::middle( start_i + T_loop_index,
C_reg,
A,
B_packed );
Loop<T_loop_index+1, T_end>::outer( start_i, C_reg, A, B_packed );
}
static inline void ALWAYS_INLINE add( Vector_intrin<real, T_reg> * X_reg,
real const * X ) {
X_reg[T_loop_index] += &X[T_loop_index * floats_per_register];
Loop<T_loop_index+1, T_end>::add( X_reg, X );
}
static inline void ALWAYS_INLINE store( Vector_intrin<real, T_reg> const * X_reg,
real * X ) {
X_reg[T_loop_index].store_p( &X[T_loop_index * floats_per_register] );
Loop<T_loop_index+1, T_end>::store( X_reg, X );
}
static inline void ALWAYS_INLINE multiple_loop( Vector_intrin<real, T_reg> * C_reg,
real const * const * const A,
real const * const * const B ) {
// Loop over columns of A and rows of B^T
Loop<0, T_K>::outer( 0, C_reg, A[T_loop_index], B[T_loop_index] );
Loop<T_loop_index+1, T_end>::multiple_loop( C_reg, A, B );
}
};
template<int T_end>
struct Loop<T_end, T_end> {
static inline void ALWAYS_INLINE set_to_zero( Vector_intrin<real, T_reg> * X_reg ) {}
static inline void ALWAYS_INLINE inner( int const row_A_reg, // == row_C_reg
int const row_B,
Vector_intrin<real, T_reg> const & A_reg,
Vector_intrin<real, T_reg> * C_reg,
real const * B_packed ) {}
static inline void ALWAYS_INLINE middle( int const col_A, // == row_B
Vector_intrin<real, T_reg> * C_reg,
real const * A,
real const * B_packed ) {}
static inline void ALWAYS_INLINE outer( int const start_i,
Vector_intrin<real, T_reg> * C_reg,
real const * A,
real const * B_packed ) {}
static inline void ALWAYS_INLINE add( Vector_intrin<real, T_reg> * X_reg,
real const * X ) {}
static inline void ALWAYS_INLINE store( Vector_intrin<real, T_reg> const * X_reg,
real * X ) {}
static inline void ALWAYS_INLINE multiple_loop( Vector_intrin<real, T_reg> * C_reg,
real const * const * const A,
real const * const * const B ) {}
};
};
// Doesn't matter if inlined or not... same performance for 4, 4, 5
// IT DOES MATTER WHEN OUTER CONTROL STRUCTURE IS ADDED ON TOP!!!
// THEN, INLINING IS BAD
template<typename real, typename T_reg, int T_M, int T_N, int T_K>
void MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::exec(real const * const * const A,
real const * const * const B,
real * const C,
int const i,
int const offset_A,
int const offset_B,
int const offset_C) {
STATIC_ASSERT_DEBUG(!(T_M%floats_per_register), TEMPLATE_ARGUMENT_T_M_MUST_BE_MULTIPLE_OF_floats_per_register);
Vector_intrin<real, T_reg> C_reg[T_M * T_N / floats_per_register];
MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_M * T_N / floats_per_register>::set_to_zero( C_reg );
#if 1 // I loose a bit performance because of the offsets
for (int ind = 0; ind < i; ++ind)
MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_K>::outer( 0, C_reg, A[ind] + offset_A, B[ind] + offset_B );
//// Loop over columns of A and rows of B^T
// MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_K>::outer( 0, C_reg, A, B );
MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_M * T_N / floats_per_register>::add( C_reg, C + offset_C);
MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_M * T_N / floats_per_register>::store( C_reg, C + offset_C);
#else
for (int ind = 0; ind < i; ++ind)
MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_K>::outer( 0, C_reg, A[ind], B[ind] );
//// Loop over columns of A and rows of B^T
// MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_K>::outer( 0, C_reg, A, B );
MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_M * T_N / floats_per_register>::add( C_reg, C);
MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_M * T_N / floats_per_register>::store( C_reg, C);
#endif
} // end exec
template<typename real, typename T_reg, int T_M, int T_N, int T_K>
template<int T_offset_A, int T_offset_B, int T_offset_C>
void MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::exec( real const * const * const A,
real const * const * const B,
real * const C,
int const i ) {
STATIC_ASSERT_DEBUG(!(T_M%floats_per_register), TEMPLATE_ARGUMENT_T_M_MUST_BE_MULTIPLE_OF_floats_per_register);
Vector_intrin<real, T_reg> C_reg[T_M * T_N / floats_per_register];
MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_M * T_N / floats_per_register>::set_to_zero( C_reg );
for (int ind = 0; ind < i; ++ind)
MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_K>::outer( 0, C_reg, A[ind] + T_offset_A, B[ind] + T_offset_B );
//// Loop over columns of A and rows of B^T
// MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_K>::outer( 0, C_reg, A, B );
MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_M * T_N / floats_per_register>::add( C_reg, C + T_offset_C);
MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::template Loop<0, T_M * T_N / floats_per_register>::store( C_reg, C + T_offset_C);
} // end exec template
/** Class template for packing of matrix elements prior to matrix-matrix multiply. */
template<typename real, typename T_reg, int T_M, int T_N, int T_K>
template<int T_rows, int T_cols, typename T_ordering_kernel, int T_repetitions>
class MM_kernel_inner_sse2_A<real, T_reg, T_M, T_N, T_K>::Pack {
public:
static int const size_packed = T_rows * T_cols * T_repetitions;
static int const rows = T_rows;
static int const cols = T_cols;
template<typename T_ordering_matrix>
struct Assign_to_packed {
typedef real * PtrTypePacked; /**< Type of packed pointer - note the absence of const qualifiers. */
typedef real const * const PtrType; /**< Type of matrix pointer - note the presence of const qualifiers. */
inline static void exec( PtrType X, PtrTypePacked X_packed,
int const row_k,
int const col_k,
int const rows_total_matrix,
int const cols_total_matrix ) {
for ( int ir = 0; ir < T_repetitions; ++ir)
X_packed[ T_ordering_kernel::get( row_k, col_k, T_rows, T_cols ) * T_repetitions + ir ]
= X[ T_ordering_matrix::get(row_k, col_k, rows_total_matrix, cols_total_matrix) ];
}
};
template<typename T_ordering_matrix>
struct Extract_from_packed {
typedef real const * const PtrTypePacked; /**< Type of packed pointer - note the presence of const qualifiers. */
typedef real * PtrType; /**< Type of matrix pointer - note the absence of const qualifiers. */
inline static void exec( PtrType X, PtrTypePacked X_packed,
int const row_k,
int const col_k,
int const rows_total_matrix,
int const cols_total_matrix ) {
for ( int ir = 0; ir < T_repetitions; ++ir)
X[ T_ordering_matrix::get(row_k, col_k, rows_total_matrix, cols_total_matrix) ] =
X_packed[ T_ordering_kernel::get( row_k, col_k, T_rows, T_cols ) * T_repetitions + ir ];
}
};
template<template<typename T_ordering> class T_assign, typename T_ordering_matrix>
static void exec(typename T_assign<T_ordering_matrix>::PtrType X,
typename T_assign<T_ordering_matrix>::PtrTypePacked X_packed,
int const rows_total_matrix, int const cols_total_matrix) {
// Loop over columns of kernel
for ( int col_k = 0; col_k < T_cols; ++col_k ) {
// Loop over rows of kernel
for ( int row_k = 0; row_k < T_rows; ++row_k ) {
T_assign<T_ordering_matrix>::exec( X, X_packed,
row_k, col_k,
rows_total_matrix, cols_total_matrix );
}
}
} // end exec()
/** Convenience function for assignments to packed matrix.
* The template argument specifies how the original (unpacked) matrix
* is stored (e.g. column- or rowwise)
*/
template<typename T_ordering_matrix>
inline static void pack(real const * const X, real * X_packed,
int const rows_total_matrix, int const cols_total_matrix) {
exec< Assign_to_packed, T_ordering_matrix >(X, X_packed, rows_total_matrix, cols_total_matrix);
}
/** Convenience function for extracting matrix from packed matrix.
* The template argument specifies how the unpacked matrix
* is stored (e.g. column- or rowwise)
*/
template<typename T_ordering_matrix>
inline static void unpack(real * X, real const * const X_packed,
int const rows_total_matrix, int const cols_total_matrix) {
exec< Extract_from_packed, T_ordering_matrix >(X, X_packed, rows_total_matrix, cols_total_matrix);
}
};
#endif
|
