Line data Source code
1 : /*----------------------------------------------------------------------------*/
2 : /* CP2K: A general program to perform molecular dynamics simulations */
3 : /* Copyright 2000-2024 CP2K developers group <https://cp2k.org> */
4 : /* */
5 : /* SPDX-License-Identifier: BSD-3-Clause */
6 : /*----------------------------------------------------------------------------*/
7 :
8 : #include <assert.h>
9 : #include <stddef.h>
10 : #include <string.h>
11 :
12 : #if defined(__LIBXSMM)
13 : #include <libxsmm.h>
14 : #if !defined(DBM_LIBXSMM_PREFETCH)
15 : //#define DBM_LIBXSMM_PREFETCH LIBXSMM_GEMM_PREFETCH_AL2_AHEAD
16 : #define DBM_LIBXSMM_PREFETCH LIBXSMM_GEMM_PREFETCH_NONE
17 : #endif
18 : #if LIBXSMM_VERSION4(1, 17, 0, 3710) > LIBXSMM_VERSION_NUMBER
19 : #define libxsmm_dispatch_gemm libxsmm_dispatch_gemm_v2
20 : #endif
21 : #endif
22 :
23 : #include "dbm_hyperparams.h"
24 : #include "dbm_multiply_cpu.h"
25 :
26 : /*******************************************************************************
27 : * \brief Prototype for BLAS dgemm.
28 : * \author Ole Schuett
29 : ******************************************************************************/
30 : void dgemm_(const char *transa, const char *transb, const int *m, const int *n,
31 : const int *k, const double *alpha, const double *a, const int *lda,
32 : const double *b, const int *ldb, const double *beta, double *c,
33 : const int *ldc);
34 :
35 : /*******************************************************************************
36 : * \brief Private convenient wrapper to hide Fortran nature of dgemm_.
37 : * \author Ole Schuett
38 : ******************************************************************************/
39 2426131 : static inline void dbm_dgemm(const char transa, const char transb, const int m,
40 : const int n, const int k, const double alpha,
41 : const double *a, const int lda, const double *b,
42 : const int ldb, const double beta, double *c,
43 : const int ldc) {
44 :
45 2426131 : dgemm_(&transa, &transb, &m, &n, &k, &alpha, a, &lda, b, &ldb, &beta, c,
46 : &ldc);
47 : }
48 :
49 : /*******************************************************************************
50 : * \brief Private hash function based on Szudzik's elegant pairing.
51 : * Using unsigned int to return a positive number even after overflow.
52 : * https://en.wikipedia.org/wiki/Pairing_function#Other_pairing_functions
53 : * https://stackoverflow.com/a/13871379
54 : * http://szudzik.com/ElegantPairing.pdf
55 : * \author Ole Schuett
56 : ******************************************************************************/
57 41532266 : static inline unsigned int hash(const dbm_task_t task) {
58 41532266 : const unsigned int m = task.m, n = task.n, k = task.k;
59 41532266 : const unsigned int mn = (m >= n) ? m * m + m + n : m + n * n;
60 41532266 : const unsigned int mnk = (mn >= k) ? mn * mn + mn + k : mn + k * k;
61 41532266 : return mnk;
62 : }
63 :
64 : /*******************************************************************************
65 : * \brief Internal routine for executing the tasks in given batch on the CPU.
66 : * \author Ole Schuett
67 : ******************************************************************************/
68 132860 : void dbm_multiply_cpu_process_batch(const int ntasks, dbm_task_t batch[ntasks],
69 : const double alpha,
70 : const dbm_pack_t *pack_a,
71 : const dbm_pack_t *pack_b,
72 132860 : dbm_shard_t *shard_c) {
73 :
74 132860 : if (0 >= ntasks) { // nothing to do
75 22229 : return;
76 : }
77 110631 : dbm_shard_allocate_promised_blocks(shard_c);
78 :
79 : #if defined(__LIBXSMM)
80 :
81 : // Sort tasks approximately by m,n,k via bucket sort.
82 110631 : int buckets[BATCH_NUM_BUCKETS];
83 110631 : memset(buckets, 0, BATCH_NUM_BUCKETS * sizeof(int));
84 20876764 : for (int itask = 0; itask < ntasks; ++itask) {
85 20766133 : const int i = hash(batch[itask]) % BATCH_NUM_BUCKETS;
86 20766133 : ++buckets[i];
87 : }
88 110631000 : for (int i = 1; i < BATCH_NUM_BUCKETS; ++i) {
89 110520369 : buckets[i] += buckets[i - 1];
90 : }
91 110631 : assert(buckets[BATCH_NUM_BUCKETS - 1] == ntasks);
92 110631 : int batch_order[ntasks];
93 20876764 : for (int itask = 0; itask < ntasks; ++itask) {
94 20766133 : const int i = hash(batch[itask]) % BATCH_NUM_BUCKETS;
95 20766133 : --buckets[i];
96 20766133 : batch_order[buckets[i]] = itask;
97 : }
98 :
99 : // Prepare arguments for libxsmm's kernel-dispatch.
100 110631 : const int flags = LIBXSMM_GEMM_FLAG_TRANS_B; // transa = "N", transb = "T"
101 110631 : const int prefetch = DBM_LIBXSMM_PREFETCH;
102 110631 : int kernel_m = 0, kernel_n = 0, kernel_k = 0;
103 110631 : dbm_task_t task_next = batch[batch_order[0]];
104 :
105 : #if (LIBXSMM_GEMM_PREFETCH_NONE != DBM_LIBXSMM_PREFETCH)
106 : double *data_a_next = NULL, *data_b_next = NULL, *data_c_next = NULL;
107 : #endif
108 : #if LIBXSMM_VERSION2(1, 17) < LIBXSMM_VERSION_NUMBER
109 : libxsmm_gemmfunction kernel_func = NULL;
110 : #else
111 110631 : libxsmm_dmmfunction kernel_func = NULL;
112 110631 : const double beta = 1.0;
113 : #endif
114 :
115 : // Loop over tasks.
116 20876764 : for (int itask = 0; itask < ntasks; ++itask) {
117 20766133 : const dbm_task_t task = task_next;
118 20766133 : task_next = batch[batch_order[(itask + 1) < ntasks ? (itask + 1) : itask]];
119 :
120 20766133 : if (task.m != kernel_m || task.n != kernel_n || task.k != kernel_k) {
121 : #if LIBXSMM_VERSION2(1, 17) < LIBXSMM_VERSION_NUMBER
122 : const libxsmm_gemm_shape shape = libxsmm_create_gemm_shape(
123 : task.m, task.n, task.k, task.m /*lda*/, task.n /*ldb/transb*/,
124 : task.m /*ldc*/, LIBXSMM_DATATYPE_F64 /*aprec*/,
125 : LIBXSMM_DATATYPE_F64 /*bprec*/, LIBXSMM_DATATYPE_F64 /*cprec*/,
126 : LIBXSMM_DATATYPE_F64 /*calcp*/);
127 : kernel_func = (LIBXSMM_FEQ(1.0, alpha)
128 : ? libxsmm_dispatch_gemm(shape, (libxsmm_bitfield)flags,
129 : (libxsmm_bitfield)prefetch)
130 : : NULL);
131 : #else
132 1389912 : kernel_func = libxsmm_dmmdispatch(task.m, task.n, task.k, NULL /*lda*/,
133 : NULL /*ldb*/, NULL /*ldc*/, &alpha,
134 : &beta, &flags, &prefetch);
135 : #endif
136 1389912 : kernel_m = task.m;
137 1389912 : kernel_n = task.n;
138 1389912 : kernel_k = task.k;
139 : }
140 :
141 : // gemm_param wants non-const data even for A and B
142 20766133 : double *const data_a = pack_a->data + task.offset_a;
143 20766133 : double *const data_b = pack_b->data + task.offset_b;
144 20766133 : double *const data_c = shard_c->data + task.offset_c;
145 :
146 20766133 : if (kernel_func != NULL) {
147 : #if LIBXSMM_VERSION2(1, 17) < LIBXSMM_VERSION_NUMBER
148 : libxsmm_gemm_param gemm_param;
149 : gemm_param.a.primary = data_a;
150 : gemm_param.b.primary = data_b;
151 : gemm_param.c.primary = data_c;
152 : #if (LIBXSMM_GEMM_PREFETCH_NONE != DBM_LIBXSMM_PREFETCH)
153 : gemm_param.a.quaternary = pack_a->data + task_next.offset_a;
154 : gemm_param.b.quaternary = pack_b->data + task_next.offset_b;
155 : gemm_param.c.quaternary = shard_c->data + task_next.offset_c;
156 : #endif
157 : kernel_func(&gemm_param);
158 : #elif (LIBXSMM_GEMM_PREFETCH_NONE != DBM_LIBXSMM_PREFETCH)
159 : kernel_func(data_a, data_b, data_c, pack_a->data + task_next.offset_a,
160 : pack_b->data + task_next.offset_b,
161 : shard_c->data + task_next.offset_c);
162 : #else
163 18340002 : kernel_func(data_a, data_b, data_c);
164 : #endif
165 : } else {
166 2426131 : dbm_dgemm('N', 'T', task.m, task.n, task.k, alpha, data_a, task.m, data_b,
167 : task.n, 1.0, data_c, task.m);
168 : }
169 : }
170 : #else
171 : // Fallback to BLAS when libxsmm is not available.
172 : for (int itask = 0; itask < ntasks; ++itask) {
173 : const dbm_task_t task = batch[itask];
174 : const double *data_a = &pack_a->data[task.offset_a];
175 : const double *data_b = &pack_b->data[task.offset_b];
176 : double *data_c = &shard_c->data[task.offset_c];
177 : dbm_dgemm('N', 'T', task.m, task.n, task.k, alpha, data_a, task.m, data_b,
178 : task.n, 1.0, data_c, task.m);
179 : }
180 : #endif
181 : }
182 :
183 : // EOF
|
