Neko 1.99.7
A portable framework for high-order spectral element flow simulations
Loading...
Searching...
No Matches
opr_opgrad.cu
Go to the documentation of this file.
1/*
2 Copyright (c) 2021-2026, The Neko Authors
3 All rights reserved.
4
5 Redistribution and use in source and binary forms, with or without
6 modification, are permitted provided that the following conditions
7 are met:
8
9 * Redistributions of source code must retain the above copyright
10 notice, this list of conditions and the following disclaimer.
11
12 * Redistributions in binary form must reproduce the above
13 copyright notice, this list of conditions and the following
14 disclaimer in the documentation and/or other materials provided
15 with the distribution.
16
17 * Neither the name of the authors nor the names of its
18 contributors may be used to endorse or promote products derived
19 from this software without specific prior written permission.
20
21 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
24 FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
25 COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
26 INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
27 BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
29 CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN
31 ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
32 POSSIBILITY OF SUCH DAMAGE.
33*/
34
35#include <string.h>
36#include <stdlib.h>
37#include <stdio.h>
38#include "opgrad_kernel.h"
39#include "elem_block_tune.h"
40#include <device/device_config.h>
41#include <device/cuda/check.h>
42
43extern "C" {
44 #include <common/neko_log.h>
45}
46
47template < const int >
48int tune_opgrad(void *ux, void *uy, void *uz, void *u,
49 void *dx, void *dy, void *dz,
50 void *drdx, void *dsdx, void *dtdx,
51 void *drdy, void *dsdy, void *dtdy,
52 void *drdz, void *dsdz, void *dtdz,
53 void *w3, int *nel, int *lx, int *eb_sel, int *ch_sel);
54
55extern "C" {
56
60 void cuda_opgrad(void *ux, void *uy, void *uz, void *u,
61 void *dx, void *dy, void *dz,
62 void *drdx, void *dsdx, void *dtdx,
63 void *drdy, void *dsdy, void *dtdy,
64 void *drdz, void *dsdz, void *dtdz,
65 void *w3, int *nel, int *lx) {
66
67 static int autotune[17] = { 0 };
68 /* elements per block candidate chosen by the tuner */
69 static int autotune_eb[17] = { 0 };
70 /* chunk candidate chosen for the 1d variant */
71 static int autotune_ch[17] = { 0 };
72
73 const dim3 nthrds_1d(1024, 1, 1);
74 const dim3 nthrds_kstep((*lx), (*lx), 1);
75 const dim3 nblcks((*nel), 1, 1);
76 const cudaStream_t stream = (cudaStream_t) glb_cmd_queue;
77
78#define CASE_1D(LX, C) \
79 opgrad_kernel_1d<real, LX, NEKO_CHUNKS(LX, C)> \
80 <<<nblcks, NEKO_CHUNKS_NTHRDS(LX, C), 0, stream>>> \
81 ((real *) ux, (real *) uy, (real *) uz, (real *) u, \
82 (real *) dx, (real *) dy, (real *) dz, \
83 (real *) drdx, (real *) dsdx, (real *) dtdx, \
84 (real *) drdy, (real *) dsdy, (real *) dtdy, \
85 (real *) drdz, (real *) dsdz, (real *) dtdz, \
86 (real *) w3); \
87 CUDA_CHECK(cudaGetLastError());
88
89/* Runtime dispatch onto the tuned chunk candidate */
90#define CASE_1D_SEL(LX, SEL) \
91 switch (SEL) { \
92 case 0: CASE_1D(LX, 0); break; \
93 case 1: CASE_1D(LX, 1); break; \
94 case 2: CASE_1D(LX, 2); break; \
95 default: CASE_1D(LX, 3); break; \
96 }
97
98
99#define CASE_KSTEP(LX, C) \
100 opgrad_kernel_kstep<real, LX, NEKO_EB(LX, C)> \
101 <<<NEKO_EB_NBLCKS(*nel, LX, C), NEKO_EB_NTHRDS(LX, C), 0, stream>>> \
102 ((real *) ux, (real *) uy, (real *) uz, (real *) u, \
103 (real *) dx, (real *) dy, (real *) dz, \
104 (real *) drdx, (real *) dsdx, (real *) dtdx, \
105 (real *) drdy, (real *) dsdy, (real *) dtdy, \
106 (real *) drdz, (real *) dsdz, (real *) dtdz, \
107 (real *) w3, *nel); \
108 CUDA_CHECK(cudaGetLastError());
109
110/* Runtime dispatch onto the tuned candidate */
111#define CASE_KSTEP_SEL(LX, SEL) \
112 switch (SEL) { \
113 case 0: CASE_KSTEP(LX, 0); break; \
114 case 1: CASE_KSTEP(LX, 1); break; \
115 default: CASE_KSTEP(LX, 2); break; \
116 }
117
118#define CASE(LX) \
119 case LX: \
120 if(autotune[LX] == 0 ) { \
121 autotune[LX]=tune_opgrad<LX>(ux, uy, uz, u, \
122 dx, dy, dz, \
123 drdx, dsdx, dtdx, \
124 drdy, dsdy, dtdy, \
125 drdz, dsdz, dtdz, \
126 w3, nel, lx, &autotune_eb[LX], \
127 &autotune_ch[LX]); \
128 } else if (autotune[LX] == 1 ) { \
129 CASE_1D_SEL(LX, autotune_ch[LX]); \
130 } else if (autotune[LX] == 2 ) { \
131 CASE_KSTEP_SEL(LX, autotune_eb[LX]); \
132 } \
133 break
134
135 switch(*lx) {
136 CASE(2);
137 CASE(3);
138 CASE(4);
139 CASE(5);
140 CASE(6);
141 CASE(7);
142 CASE(8);
143 CASE(9);
144 CASE(10);
145 CASE(11);
146 CASE(12);
147 CASE(13);
148 CASE(14);
149 CASE(15);
150 CASE(16);
151 default:
152 {
153 fprintf(stderr, __FILE__ ": size not supported: %d\n", *lx);
154 exit(1);
155 }
156 }
157 }
158}
159
160template < const int LX >
161int tune_opgrad(void *ux, void *uy, void *uz, void *u,
162 void *dx, void *dy, void *dz,
163 void *drdx, void *dsdx, void *dtdx,
164 void *drdy, void *dsdy, void *dtdy,
165 void *drdz, void *dsdz, void *dtdz,
166 void *w3, int *nel, int *lx, int *eb_sel, int *ch_sel) {
167 cudaEvent_t start,stop;
168 float time1[NEKO_CHUNKS_CANDIDATES];
169 int best1 = 0;
170 float time2[NEKO_EB_CANDIDATES];
171 const int rounds = neko_tune_rounds();
172 const int iters = neko_tune_iters();
173 const int sweep = neko_eb_sweep();
174 int best = 0;
175 int retval;
176
177 for (int c = 0; c < NEKO_EB_CANDIDATES; c++) {
178 time2[c] = NEKO_TUNE_INIT;
179 }
180 for (int c = 0; c < NEKO_CHUNKS_CANDIDATES; c++) {
181 time1[c] = NEKO_TUNE_INIT;
182 }
183
184 const dim3 nthrds_1d(1024, 1, 1);
185 const dim3 nthrds_kstep((*lx), (*lx), 1);
186 const dim3 nblcks((*nel), 1, 1);
187 const cudaStream_t stream = (cudaStream_t) glb_cmd_queue;
188
189 char *env_value = NULL;
190 char neko_log_buf[80];
191
192 env_value=getenv("NEKO_AUTOTUNE");
193
194 sprintf(neko_log_buf, "Autotune opgrad (lx: %d)", *lx);
195 log_section(neko_log_buf);
196
197 *eb_sel = 0;
198 *ch_sel = 0;
199
200 if(env_value) {
201 if( !strcmp(env_value,"1D") ) {
202 *ch_sel = neko_chunks_env();
203 CASE_1D_SEL(LX, *ch_sel);
204 sprintf(neko_log_buf,"Set by env : 1 (1D, %d chunk)",
205 NEKO_CHUNKS_SEL(LX, *ch_sel));
206 log_message(neko_log_buf);
207 log_end_section();
208 return 1;
209 } else if( !strcmp(env_value,"KSTEP") ) {
210 {
211 const int c = neko_eb_env();
212 *eb_sel = c;
213 CASE_KSTEP_SEL(LX, c);
214 sprintf(neko_log_buf,"Set by env : 2 (KSTEP, %d elem/block)",
215 NEKO_EB_SEL(LX, c));
216 }
217 log_message(neko_log_buf);
218 log_end_section();
219 return 2;
220 } else {
221 sprintf(neko_log_buf, "Invalid value set for NEKO_AUTOTUNE");
222 log_error(neko_log_buf);
223 }
224 }
225
226 cudaEventCreate(&start);
227 cudaEventCreate(&stop);
228
229 /* Warm every variant before timing anything: each specialisation has to be
230 resident and the clocks at steady state, or whichever is timed first is
231 measured on a colder part */
232 for (int i = 0; i < NEKO_TUNE_WARMUP; i++) {
233 CASE_1D(LX, 0);
234 CASE_1D(LX, 1);
235 CASE_1D(LX, 2);
236 CASE_1D(LX, 3);
237 CASE_KSTEP(LX, 0);
238 if (sweep) {
239 CASE_KSTEP(LX, 1);
240 CASE_KSTEP(LX, 2);
241 }
242 }
243
244 /* Interleaved rounds, best time per variant */
245 for (int r = 0; r < rounds; r++) {
246 NEKO_TUNE_TIME(time1, CASE_1D, LX, 0, iters);
247 NEKO_TUNE_TIME(time1, CASE_1D, LX, 1, iters);
248 NEKO_TUNE_TIME(time1, CASE_1D, LX, 2, iters);
249 NEKO_TUNE_TIME(time1, CASE_1D, LX, 3, iters);
250 NEKO_TUNE_TIME(time2, CASE_KSTEP, LX, 0, iters);
251 if (sweep) {
252 NEKO_TUNE_TIME(time2, CASE_KSTEP, LX, 1, iters);
253 NEKO_TUNE_TIME(time2, CASE_KSTEP, LX, 2, iters);
254 }
255 }
256
257 NEKO_TUNE_LOG(LX, time1, time2);
258 NEKO_TUNE_BEST(time1, best1, NEKO_CHUNKS_CANDIDATES);
259 NEKO_TUNE_BEST(time2, best, NEKO_EB_CANDIDATES);
260 *eb_sel = best;
261 *ch_sel = best1;
262
263 if (time1[best1] < time2[best]) {
264 retval = 1;
265 } else {
266 retval = 2;
267 }
268
269 /* Leave the chosen kernel's output in place: the tuner stands in for a real
270 evaluation and the variants do not sum in the same order */
271 if (retval == 1) {
272 CASE_1D_SEL(LX, best1);
273 } else {
274 CASE_KSTEP_SEL(LX, best);
275 }
276
277 if (retval == 1) {
278 sprintf(neko_log_buf, "Chose : 1 (1D, %d chunk)",
279 NEKO_CHUNKS_SEL(LX, best1));
280 } else {
281 sprintf(neko_log_buf, "Chose : 2 (KSTEP, %d elem/block)",
282 NEKO_EB_SEL(LX, best));
283 }
284 log_message(neko_log_buf);
285 log_end_section();
286 return retval;
287}
ale_add_kinematics_kernel
__global__ void ale_add_kinematics_kernel(const int n, T *__restrict__ wx, T *__restrict__ wy, T *__restrict__ wz, const T *__restrict__ x_ref, const T *__restrict__ y_ref, const T *__restrict__ z_ref, const T *__restrict__ phi, const T *__restrict__ x, const T *__restrict__ y, const T *__restrict__ z, const kinematics_params_t kin_params)
Definition ale_kinematics_kernel.h:52
drdy
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ drdy
Definition ax_helm_full_kernel.h:54
drdz
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ drdz
Definition ax_helm_full_kernel.h:55
dsdz
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dsdz
Definition ax_helm_full_kernel.h:58
dsdy
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dsdy
Definition ax_helm_full_kernel.h:57
i
const int i
Definition ax_helm_full_kernel.h:95
dtdy
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dtdy
Definition ax_helm_full_kernel.h:60
u
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ u
Definition ax_helm_full_kernel.h:46
dx
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dx
Definition ax_helm_full_kernel.h:49
drdx
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ drdx
Definition ax_helm_full_kernel.h:53
dtdz
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dtdz
Definition ax_helm_full_kernel.h:61
dsdx
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dsdx
Definition ax_helm_full_kernel.h:56
dtdx
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dtdx
Definition ax_helm_full_kernel.h:59
dz
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dz
Definition ax_helm_full_kernel.h:51
dy
__global__ void T *__restrict__ T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ dy
Definition ax_helm_full_kernel.h:50
w3
__global__ void const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ const T *__restrict__ w3
Definition cdtp_kernel.h:115
NEKO_CHUNKS_CANDIDATES
#define NEKO_CHUNKS_CANDIDATES
Definition elem_block.h:125
NEKO_EB_CANDIDATES
#define NEKO_EB_CANDIDATES
Definition elem_block.h:63
NEKO_EB_SEL
#define NEKO_EB_SEL(LX, SEL)
Definition elem_block.h:108
NEKO_CHUNKS_SEL
#define NEKO_CHUNKS_SEL(LX, SEL)
Definition elem_block.h:147
NEKO_TUNE_TIME
#define NEKO_TUNE_TIME(T, LAUNCH, LX, C, ITERS)
Definition elem_block_tune.h:119
neko_eb_env
static int neko_eb_env()
Definition elem_block_tune.h:79
neko_tune_rounds
static int neko_tune_rounds()
Definition elem_block_tune.h:102
NEKO_TUNE_LOG
#define NEKO_TUNE_LOG(LX, T1, T2)
Definition elem_block_tune.h:138
NEKO_TUNE_INIT
#define NEKO_TUNE_INIT
Definition elem_block_tune.h:65
NEKO_TUNE_BEST
#define NEKO_TUNE_BEST(T, BEST, N)
Definition elem_block_tune.h:130
neko_tune_iters
static int neko_tune_iters()
Definition elem_block_tune.h:110
neko_chunks_env
static int neko_chunks_env()
Definition elem_block_tune.h:91
neko_eb_sweep
static int neko_eb_sweep()
Definition elem_block_tune.h:68
NEKO_TUNE_WARMUP
#define NEKO_TUNE_WARMUP
Definition elem_block_tune.h:66
uy
__global__ void T *__restrict__ uy
Definition opgrad_kernel.h:128
uz
__global__ void T *__restrict__ T *__restrict__ uz
Definition opgrad_kernel.h:129
device_config.h
elem_block_tune.h
opgrad_kernel.h
neko_log.h
log_error
void log_error(char *msg)
log_message
void log_message(char *msg)
log_end_section
void log_end_section()
log_section
void log_section(char *msg)
CASE_KSTEP_SEL
#define CASE_KSTEP_SEL(LX, SEL)
CASE
#define CASE(LX)
CASE_1D_SEL
#define CASE_1D_SEL(LX, SEL)
cuda_opgrad
void cuda_opgrad(void *ux, void *uy, void *uz, void *u, void *dx, void *dy, void *dz, void *drdx, void *dsdx, void *dtdx, void *drdy, void *dsdy, void *dtdy, void *drdz, void *dsdz, void *dtdz, void *w3, int *nel, int *lx)
Definition opr_opgrad.cu:60
CASE_KSTEP
#define CASE_KSTEP(LX, C)
CASE_1D
#define CASE_1D(LX, C)
tune_opgrad
int tune_opgrad(void *ux, void *uy, void *uz, void *u, void *dx, void *dy, void *dz, void *drdx, void *dsdx, void *dtdx, void *drdy, void *dsdy, void *dtdy, void *drdz, void *dsdz, void *dtdz, void *w3, int *nel, int *lx, int *eb_sel, int *ch_sel)
Definition opr_opgrad.cu:161