/
time_GPU.cu
353 lines (330 loc) · 19.4 KB
/
time_GPU.cu
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
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
/***
nvcc -std=c++11 -o GPU_timing.exe time_GPU.cu -gencode arch=compute_75,code=sm_75 -O3
***/
#define NUM_STREAMS 4
#include "utils/experiment_helpers.h"
#include "helpers_CPU/dynamicsGradient.h" // GPU requires CPU for partial kernels
#include "helpers_GPU/dynamicsGradient.cuh"
#if TEST_FOR_EQUIVALENCE
dim3 dimms(1,1);
#else
// mainly single loops a few double loops where dimy = NUM_POS
// dimx and single loops ocassionally very high but generally O(NUM_POS)
// with small constants. As such we set dimx = 32 which is the warp size
// to avoid divergence of warps during double loops (ideally) and this is
// not too large that it will overtax the GPU scheduler
dim3 dimms(32,NUM_POS);
#endif
template <typename T, int TEST_ITERS, bool MPC_MODE = false, bool VEL_DAMPING = true>
__global__
void kern_single_full(T *d_dqdd, T *d_mem_vol, T *d_mem_const, T *s_fext = nullptr){
int starty, dy, startx, dx; doubleLoopVals_GPU(&starty,&dy,&startx,&dx); int start, delta; singleLoopVals_GPU(&start,&delta);
__shared__ T s_mem_const[CONST_VALS]; T *s_T = &s_mem_const[0]; T *s_I = &s_mem_const[36*NUM_POS];
__shared__ T s_mem_vol[COMPLETELY_FUSED_VALS]; T *s_q = &s_mem_vol[0]; T *s_qd = &s_mem_vol[NUM_POS];
T *s_qdd = &s_mem_vol[3*NUM_POS]; T *s_Minv = &s_mem_vol[4*NUM_POS]; //T *s_u = &s_mem_vol[2*NUM_POS];
__shared__ T s_sinq[NUM_POS]; __shared__ T s_cosq[NUM_POS]; __shared__ T s_temp[6*NUM_POS];
__shared__ T s_v[6*NUM_POS]; __shared__ T s_a[6*NUM_POS]; __shared__ T s_f[6*NUM_POS];
__shared__ T s_cdu[2*NUM_POS*NUM_POS]; __shared__ T s_dqdd[2*NUM_POS*NUM_POS];
// Load in vars to shared mem
#pragma unroll
for (int ind = start; ind < CONST_VALS; ind += delta){s_mem_const[ind] = d_mem_const[ind];}
#pragma unroll
for (int ind = start; ind < COMPLETELY_FUSED_VALS; ind += delta){s_mem_vol[ind] = d_mem_vol[ind];}
#pragma unroll
for (int ind = start; ind < NUM_POS; ind += delta){s_sinq[ind] = sin(s_q[ind]); s_cosq[ind] = cos(s_q[ind]);}
__syncthreads();
// loop for all test iters
for (int i = 0; i < TEST_ITERS; i++){
// build Tmats
updateTransforms_GPU<T>(s_T,s_sinq,s_cosq); __syncthreads();
// dqdd/dtau = Minv and dqdd/dq(d) = -Minv*dc/dq(d) note: dM term in dq drops out if you compute c with fd qdd per carpentier
FD_helpers_GPU_vaf<T,MPC_MODE>(s_v,s_a,s_f,s_qd,s_qdd,s_I,s_T,s_temp,s_fext); __syncthreads();
inverseDynamicsGradient_GPU<T,VEL_DAMPING>(s_cdu,s_qd,s_v,s_a,s_f,s_I,s_T,0); __syncthreads();
// finally compute the final dqdd by multiplying by Minv
finish_dqdd_GPU<T>(s_dqdd,s_cdu,s_Minv); __syncthreads();
}
#pragma unroll
for (int ind = start; ind < 2*NUM_POS*NUM_POS; ind += delta){d_dqdd[ind] = s_dqdd[ind];}
}
template <typename T, int TEST_ITERS, bool MPC_MODE = false, bool VEL_DAMPING = true>
__global__
void kern_single_vaf_dcdu(T *d_cdu, T *d_mem_vol, T *d_mem_const, T *s_fext = nullptr){
int starty, dy, startx, dx; doubleLoopVals_GPU(&starty,&dy,&startx,&dx); int start, delta; singleLoopVals_GPU(&start,&delta);
__shared__ T s_mem_const[CONST_VALS]; T *s_T = &s_mem_const[0]; T *s_I = &s_mem_const[36*NUM_POS];
__shared__ T s_mem_vol[FUSED_VALS]; T *s_q = &s_mem_vol[0]; T *s_qd = &s_mem_vol[NUM_POS]; T *s_qdd = &s_mem_vol[3*NUM_POS];
// T *s_u = &s_mem_vol[2*NUM_POS];
__shared__ T s_sinq[NUM_POS]; __shared__ T s_cosq[NUM_POS]; __shared__ T s_temp[6*NUM_POS];
__shared__ T s_v[6*NUM_POS]; __shared__ T s_a[6*NUM_POS]; __shared__ T s_f[6*NUM_POS];
__shared__ T s_cdu[2*NUM_POS*NUM_POS];
// Load in vars to shared mem
#pragma unroll
for (int ind = start; ind < CONST_VALS; ind += delta){s_mem_const[ind] = d_mem_const[ind];}
#pragma unroll
for (int ind = start; ind < FUSED_VALS; ind += delta){s_mem_vol[ind] = d_mem_vol[ind];}
#pragma unroll
for (int ind = start; ind < NUM_POS; ind += delta){s_sinq[ind] = sin(s_q[ind]); s_cosq[ind] = cos(s_q[ind]);}
__syncthreads();
// loop for all test iters
for (int i = 0; i < TEST_ITERS; i++){
// build Tmats
updateTransforms_GPU<T>(s_T,s_sinq,s_cosq); __syncthreads();
// dqdd/dtau = Minv and dqdd/dq(d) = -Minv*dc/dq(d) note: dM term in dq drops out if you compute c with fd qdd per carpentier
FD_helpers_GPU_vaf<T,MPC_MODE>(s_v,s_a,s_f,s_qd,s_qdd,s_I,s_T,s_temp,s_fext); __syncthreads();
inverseDynamicsGradient_GPU<T,VEL_DAMPING>(s_cdu,s_qd,s_v,s_a,s_f,s_I,s_T,0); __syncthreads();
}
#pragma unroll
for (int ind = start; ind < 2*NUM_POS*NUM_POS; ind += delta){d_cdu[ind] = s_cdu[ind];}
}
template <typename T, int TEST_ITERS, bool MPC_MODE = false, bool VEL_DAMPING = true>
__global__
void kern_single_vaf(T *d_vaf, T *d_mem_vol, T *d_mem_const, T *s_fext = nullptr){
int starty, dy, startx, dx; doubleLoopVals_GPU(&starty,&dy,&startx,&dx); int start, delta; singleLoopVals_GPU(&start,&delta);
__shared__ T s_mem_const[CONST_VALS]; T *s_T = &s_mem_const[0]; T *s_I = &s_mem_const[36*NUM_POS];
__shared__ T s_mem_vol[FUSED_VALS]; T *s_q = &s_mem_vol[0]; T *s_qd = &s_mem_vol[NUM_POS]; T *s_qdd = &s_mem_vol[3*NUM_POS];
// T *s_u = &s_mem_vol[2*NUM_POS];
__shared__ T s_sinq[NUM_POS]; __shared__ T s_cosq[NUM_POS]; __shared__ T s_temp[6*NUM_POS];
__shared__ T s_v[6*NUM_POS]; __shared__ T s_a[6*NUM_POS]; __shared__ T s_f[6*NUM_POS];
// Load in vars to shared mem
#pragma unroll
for (int ind = start; ind < CONST_VALS; ind += delta){s_mem_const[ind] = d_mem_const[ind];}
#pragma unroll
for (int ind = start; ind < FUSED_VALS; ind += delta){s_mem_vol[ind] = d_mem_vol[ind];}
#pragma unroll
for (int ind = start; ind < NUM_POS; ind += delta){s_sinq[ind] = sin(s_q[ind]); s_cosq[ind] = cos(s_q[ind]);}
__syncthreads();
// loop for all test iters
for (int i = 0; i < TEST_ITERS; i++){
// build Tmats
updateTransforms_GPU<T>(s_T,s_sinq,s_cosq); __syncthreads();
// dqdd/dtau = Minv and dqdd/dq(d) = -Minv*dc/dq(d) note: dM term in dq drops out if you compute c with fd qdd per carpentier
FD_helpers_GPU_vaf<T,MPC_MODE>(s_v,s_a,s_f,s_qd,s_qdd,s_I,s_T,s_temp,s_fext); __syncthreads();
}
#pragma unroll
for (int ind = start; ind < 6*NUM_POS; ind += delta){d_vaf[ind] = s_v[ind]; d_vaf[ind+6*NUM_POS] = s_a[ind]; d_vaf[ind+12*NUM_POS] = s_f[ind];}
}
template<typename T, int TEST_ITERS, int NUM_THREADS_TEST, int NUM_TIME_STEPS_TEST, bool MPC_MODE = false, bool VEL_DAMPING = true>
void test(){
// allocate and load on CPU
T *h_dqdd = (T *)malloc(2*NUM_POS*NUM_POS*NUM_TIME_STEPS_TEST*sizeof(T));
T *h_cdu = (T *)malloc(2*NUM_POS*NUM_POS*NUM_TIME_STEPS_TEST*sizeof(T));
T *h_vaf = (T *)malloc(6*NUM_POS*3*NUM_TIME_STEPS_TEST*sizeof(T));
T *h_mem_const = (T *)malloc(CONST_VALS*NUM_THREADS_TEST*sizeof(T)); // T,I
for(int tid = 0; tid < NUM_THREADS_TEST; tid++){
initTransforms(&h_mem_const[CONST_VALS*tid]); initInertiaTensors(&h_mem_const[CONST_VALS*tid + 36*NUM_POS]);
}
traj<T,NUM_TIME_STEPS_TEST> *testTraj = new traj<T,NUM_TIME_STEPS_TEST>;
for (int k = 0; k < NUM_TIME_STEPS_TEST; k++){
for(int i = 0; i < NUM_POS; i++){
#if TEST_FOR_EQUIVALENCE
testTraj->knots[k].q[i] = 0.1;
testTraj->knots[k].qd[i] = 0.1;
testTraj->knots[k].u[i] = 0.1;
#else
testTraj->knots[k].q[i] = getRand<T>();
testTraj->knots[k].qd[i] = getRand<T>();
testTraj->knots[k].u[i] = getRand<T>();
#endif
}
forwardDynamics<T,MPC_MODE,VEL_DAMPING,1>(&(testTraj->knots[k]));
}
// and load into gpu structured memory (on the CPU)
T *h_mem_vol_split = (T *)malloc(SPLIT_VALS*NUM_TIME_STEPS_TEST*sizeof(T)); // q,qd,u,qdd,v,a,f
T *h_Minv = (T *)malloc(NUM_POS*NUM_POS*NUM_TIME_STEPS_TEST*sizeof(T)); // for finish in split kernel
T *h_mem_vol_fused = (T *)malloc(FUSED_VALS*NUM_TIME_STEPS_TEST*sizeof(T)); // q,qd,u,qdd
T *h_mem_vol_completely_fused = (T *)malloc(COMPLETELY_FUSED_VALS*NUM_TIME_STEPS_TEST*sizeof(T)); // q,qd,u,qdd,Minv
for (int k = 0; k < NUM_TIME_STEPS_TEST; k++){
T *h_mem_vol_splitk = &h_mem_vol_split[SPLIT_VALS*k];
T *h_mem_vol_fusedk = &h_mem_vol_fused[FUSED_VALS*k];
T *h_mem_vol_completely_fusedk = &h_mem_vol_completely_fused[COMPLETELY_FUSED_VALS*k];
for(int i = 0; i < NUM_POS; i++){
h_mem_vol_splitk[i] = testTraj->knots[k].q[i];
h_mem_vol_splitk[i+NUM_POS] = testTraj->knots[k].qd[i];
h_mem_vol_splitk[i+2*NUM_POS] = testTraj->knots[k].u[i];
h_mem_vol_splitk[i+3*NUM_POS] = testTraj->knots[k].qdd[i];
}
for(int i = 0; i < 4*NUM_POS; i++){
h_mem_vol_fusedk[i] = h_mem_vol_splitk[i]; h_mem_vol_completely_fusedk[i] = h_mem_vol_splitk[i];
}
for(int i = 0; i < NUM_POS*NUM_POS; i++){
h_mem_vol_completely_fusedk[4*NUM_POS + i] = testTraj->knots[k].Minv[i];
h_Minv[NUM_POS*NUM_POS*k + i] = testTraj->knots[k].Minv[i];
}
}
// allocate and copy to GPU
T *d_dqdd; gpuErrchk(cudaMalloc((void**)&d_dqdd, 2*NUM_POS*NUM_POS*NUM_TIME_STEPS_TEST*sizeof(T)));
T *d_cdu; gpuErrchk(cudaMalloc((void**)&d_cdu, 2*NUM_POS*NUM_POS*NUM_TIME_STEPS_TEST*sizeof(T)));
T *d_vaf; gpuErrchk(cudaMalloc((void**)&d_vaf, 6*NUM_POS*3*NUM_TIME_STEPS_TEST*sizeof(T)));
T *d_mem_const; gpuErrchk(cudaMalloc((void**)&d_mem_const, CONST_VALS*sizeof(T))); // T,I
T *d_mem_vol_split; gpuErrchk(cudaMalloc((void**)&d_mem_vol_split, SPLIT_VALS*NUM_TIME_STEPS_TEST*sizeof(T))); // q,qd,u,qdd,v,a,f
T *d_mem_vol_fused; gpuErrchk(cudaMalloc((void**)&d_mem_vol_fused, FUSED_VALS*NUM_TIME_STEPS_TEST*sizeof(T))); // q,qd,u,qdd
T *d_mem_vol_completely_fused; gpuErrchk(cudaMalloc((void**)&d_mem_vol_completely_fused, COMPLETELY_FUSED_VALS*NUM_TIME_STEPS_TEST*sizeof(T))); // q,qd,u,qdd,Minv
cudaStream_t *streams = (cudaStream_t *)malloc(NUM_STREAMS*sizeof(cudaStream_t));
int priority, minPriority, maxPriority;
gpuErrchk(cudaDeviceGetStreamPriorityRange(&minPriority, &maxPriority));
for(int i=0; i<NUM_STREAMS; i++){priority = std::min(minPriority+i,maxPriority);
gpuErrchk(cudaStreamCreateWithPriority(&(streams[i]),cudaStreamNonBlocking,priority));
}
gpuErrchk(cudaMemcpy(d_mem_const,h_mem_const,CONST_VALS*sizeof(T),cudaMemcpyHostToDevice));
gpuErrchk(cudaMemcpy(d_mem_vol_split,h_mem_vol_split,SPLIT_VALS*NUM_TIME_STEPS_TEST*sizeof(T),cudaMemcpyHostToDevice));
gpuErrchk(cudaMemcpy(d_mem_vol_fused,h_mem_vol_fused,FUSED_VALS*NUM_TIME_STEPS_TEST*sizeof(T),cudaMemcpyHostToDevice));
gpuErrchk(cudaMemcpy(d_mem_vol_completely_fused,h_mem_vol_completely_fused,COMPLETELY_FUSED_VALS*NUM_TIME_STEPS_TEST*sizeof(T),cudaMemcpyHostToDevice));
gpuErrchk(cudaDeviceSynchronize());
// create resusable threads for any multithreaded CPU calls
ReusableThreads<NUM_THREADS_TEST> threads;
// time kernel with uncompressed and compressed memory copy
struct timespec start, end;
if(NUM_TIME_STEPS_TEST == 1){
#if TEST_FOR_EQUIVALENCE
#define SINGLE_TEST_ITERS (TEST_ITERS*1)
#else
#define SINGLE_TEST_ITERS (TEST_ITERS*10)
#endif
clock_gettime(CLOCK_MONOTONIC,&start);
kern_single_full<T,SINGLE_TEST_ITERS,MPC_MODE,VEL_DAMPING><<<1,dimms>>>(d_dqdd,d_mem_vol_completely_fused,d_mem_const);
gpuErrchk(cudaDeviceSynchronize());
clock_gettime(CLOCK_MONOTONIC,&end);
printf("Single Call vaf+dc/du+dqdd/du %fus\n",time_delta_us_timespec(start,end)/static_cast<double>(SINGLE_TEST_ITERS));
gpuErrchk(cudaMemcpyAsync(h_dqdd,d_dqdd,NUM_POS*2*NUM_POS*sizeof(T),cudaMemcpyDeviceToHost, streams[0])); gpuErrchk(cudaDeviceSynchronize());
#if TEST_FOR_EQUIVALENCE
printf("q,qd,qdd,u\n");
printMat<T,1,NUM_POS>(h_mem_vol_completely_fused,1);
printMat<T,1,NUM_POS>(&h_mem_vol_completely_fused[NUM_POS],1);
printMat<T,1,NUM_POS>(&h_mem_vol_completely_fused[3*NUM_POS],1);
printMat<T,1,NUM_POS>(&h_mem_vol_completely_fused[2*NUM_POS],1);
printf("Minv\n");
printMat<T,NUM_POS,NUM_POS>(&h_mem_vol_completely_fused[4*NUM_POS],NUM_POS);
printf("dqdd/dq\n");
printMat<T,NUM_POS,NUM_POS>(h_dqdd,NUM_POS);
printf("dqdd/dqd\n");
printMat<T,NUM_POS,NUM_POS>(&h_dqdd[NUM_POS*NUM_POS],NUM_POS);
#endif
clock_gettime(CLOCK_MONOTONIC,&start);
kern_single_vaf_dcdu<T,SINGLE_TEST_ITERS,MPC_MODE,VEL_DAMPING><<<1,dimms>>>(d_dqdd,d_mem_vol_fused,d_mem_const);
gpuErrchk(cudaDeviceSynchronize());
clock_gettime(CLOCK_MONOTONIC,&end);
printf("Single Call vaf+dc/du %fus\n",time_delta_us_timespec(start,end)/static_cast<double>(SINGLE_TEST_ITERS));
clock_gettime(CLOCK_MONOTONIC,&start);
kern_single_vaf<T,SINGLE_TEST_ITERS,MPC_MODE,VEL_DAMPING><<<1,dimms>>>(d_vaf,d_mem_vol_fused,d_mem_const);
gpuErrchk(cudaDeviceSynchronize());
clock_gettime(CLOCK_MONOTONIC,&end);
printf("Single Call vaf %fus\n",time_delta_us_timespec(start,end)/static_cast<double>(SINGLE_TEST_ITERS));
printf("----------------------------------------\n");
}
// time each end to end option (and memory back outs etc.)
else{
// ----------------------------------------//
// Split: CPU vaf > GPU dc/du > CPU finish //
// ----------------------------------------//
std::vector<double> times = {};
for(int iter = 0; iter < TEST_ITERS; iter++){
clock_gettime(CLOCK_MONOTONIC,&start);
dynamicsGradientReusableThreaded_start<T,NUM_THREADS_TEST,NUM_TIME_STEPS_TEST,MPC_MODE,VEL_DAMPING>(h_mem_vol_split,h_mem_const,&threads);
clock_gettime(CLOCK_MONOTONIC,&end);
times.push_back(time_delta_us_timespec(start,end));
}
printf("[N:%d]: Split - CPU start: ",NUM_TIME_STEPS_TEST); printStats<PRINT_DISTRIBUTIONS_GLOBAL>(×);
std::vector<double> times2 = {};
for(int iter = 0; iter < TEST_ITERS; iter++){
clock_gettime(CLOCK_MONOTONIC,&start);
gpuErrchk(cudaMemcpy(d_mem_vol_split,h_mem_vol_split,SPLIT_VALS*NUM_TIME_STEPS_TEST*sizeof(T),cudaMemcpyHostToDevice));
dynamicsGradientKernel_split<T,NUM_TIME_STEPS_TEST,VEL_DAMPING><<<NUM_TIME_STEPS_TEST,dimms>>>(d_cdu,d_mem_vol_split,d_mem_const);
gpuErrchk(cudaDeviceSynchronize());
gpuErrchk(cudaMemcpy(h_cdu,d_cdu,2*NUM_POS*NUM_POS*NUM_TIME_STEPS_TEST*sizeof(T),cudaMemcpyDeviceToHost));
clock_gettime(CLOCK_MONOTONIC,&end);
times2.push_back(time_delta_us_timespec(start,end));
}
printf("[N:%d]: Split - GPU Compute + I/O: ",NUM_TIME_STEPS_TEST); printStats<PRINT_DISTRIBUTIONS_GLOBAL>(×2);
std::vector<double> times3 = {};
for(int iter = 0; iter < TEST_ITERS; iter++){
clock_gettime(CLOCK_MONOTONIC,&start);
dynamicsGradientKernel_split<T,NUM_TIME_STEPS_TEST,VEL_DAMPING><<<NUM_TIME_STEPS_TEST,dimms>>>(d_cdu,d_mem_vol_split,d_mem_const);
gpuErrchk(cudaDeviceSynchronize());
clock_gettime(CLOCK_MONOTONIC,&end);
times3.push_back(time_delta_us_timespec(start,end));
}
printf("[N:%d]: Split - GPU Compute: ",NUM_TIME_STEPS_TEST); printStats<PRINT_DISTRIBUTIONS_GLOBAL>(×3);
std::vector<double> times4 = {};
for(int iter = 0; iter < TEST_ITERS; iter++){
clock_gettime(CLOCK_MONOTONIC,&start);
dynamicsGradientReusableThreaded_finish<T,NUM_THREADS_TEST,NUM_TIME_STEPS_TEST,MPC_MODE,VEL_DAMPING>(h_dqdd,h_cdu,h_Minv,&threads);
clock_gettime(CLOCK_MONOTONIC,&end);
times4.push_back(time_delta_us_timespec(start,end));
}
printf("[N:%d] Split - CPU finish: ",NUM_TIME_STEPS_TEST); printStats<PRINT_DISTRIBUTIONS_GLOBAL>(×4);
printf("\n");
// ---------------------------------//
// Fused: GPU vaf+dcdu > CPU finish //
// ---------------------------------//
std::vector<double> times5 = {};
for(int iter = 0; iter < TEST_ITERS; iter++){
clock_gettime(CLOCK_MONOTONIC,&start);
gpuErrchk(cudaMemcpy(d_mem_vol_fused,h_mem_vol_fused,FUSED_VALS*NUM_TIME_STEPS_TEST*sizeof(T),cudaMemcpyHostToDevice));
dynamicsGradientKernel_fused<T,NUM_TIME_STEPS_TEST,VEL_DAMPING><<<NUM_TIME_STEPS_TEST,dimms>>>(d_dqdd,d_mem_vol_fused,d_mem_const);
gpuErrchk(cudaDeviceSynchronize());
gpuErrchk(cudaMemcpy(h_dqdd,d_dqdd,2*NUM_POS*NUM_POS*NUM_TIME_STEPS_TEST*sizeof(T),cudaMemcpyDeviceToHost));
clock_gettime(CLOCK_MONOTONIC,&end);
times5.push_back(time_delta_us_timespec(start,end));
}
printf("[N:%d] Fused - GPU compute + I/O: ",NUM_TIME_STEPS_TEST); printStats<PRINT_DISTRIBUTIONS_GLOBAL>(×5);
std::vector<double> times6 = {};
for(int iter = 0; iter < TEST_ITERS; iter++){
clock_gettime(CLOCK_MONOTONIC,&start);
dynamicsGradientKernel_fused<T,NUM_TIME_STEPS_TEST,VEL_DAMPING><<<NUM_TIME_STEPS_TEST,dimms>>>(d_dqdd,d_mem_vol_fused,d_mem_const);
gpuErrchk(cudaDeviceSynchronize());
clock_gettime(CLOCK_MONOTONIC,&end);
times6.push_back(time_delta_us_timespec(start,end));
}
printf("[N:%d] Fused - GPU compute: ",NUM_TIME_STEPS_TEST); printStats<PRINT_DISTRIBUTIONS_GLOBAL>(×6);
std::vector<double> times7 = {};
for(int iter = 0; iter < TEST_ITERS; iter++){
clock_gettime(CLOCK_MONOTONIC,&start);
dynamicsGradientReusableThreaded_finish<T,NUM_THREADS_TEST,NUM_TIME_STEPS_TEST,MPC_MODE,VEL_DAMPING>(h_dqdd,h_cdu,h_Minv,&threads);
clock_gettime(CLOCK_MONOTONIC,&end);
times7.push_back(time_delta_us_timespec(start,end));
}
printf("[N:%d] Fused - CPU finish: ",NUM_TIME_STEPS_TEST); printStats<PRINT_DISTRIBUTIONS_GLOBAL>(×7);
printf("\n");
// -----------------------------//
// Completely Fused: All on GPU //
// -----------------------------//
std::vector<double> times8 = {};
for(int iter = 0; iter < TEST_ITERS; iter++){
clock_gettime(CLOCK_MONOTONIC,&start);
gpuErrchk(cudaMemcpy(d_mem_vol_completely_fused,h_mem_vol_completely_fused,COMPLETELY_FUSED_VALS*NUM_TIME_STEPS_TEST*sizeof(T),cudaMemcpyHostToDevice));
dynamicsGradientKernel<T,NUM_TIME_STEPS_TEST,MPC_MODE,VEL_DAMPING><<<NUM_TIME_STEPS_TEST,dimms>>>(d_dqdd,d_mem_vol_completely_fused,d_mem_const);
gpuErrchk(cudaDeviceSynchronize());
gpuErrchk(cudaMemcpy(h_dqdd,d_dqdd,2*NUM_POS*NUM_POS*NUM_TIME_STEPS_TEST*sizeof(T),cudaMemcpyDeviceToHost));
clock_gettime(CLOCK_MONOTONIC,&end);
times8.push_back(time_delta_us_timespec(start,end));
}
printf("[N:%d] Completely Fused - GPU Compute + I/O: ",NUM_TIME_STEPS_TEST); printStats<PRINT_DISTRIBUTIONS_GLOBAL>(×8);
std::vector<double> times9= {};
for(int iter = 0; iter < TEST_ITERS; iter++){
clock_gettime(CLOCK_MONOTONIC,&start);
dynamicsGradientKernel<T,NUM_TIME_STEPS_TEST,MPC_MODE,VEL_DAMPING><<<NUM_TIME_STEPS_TEST,dimms>>>(d_dqdd,d_mem_vol_completely_fused,d_mem_const);
gpuErrchk(cudaDeviceSynchronize());
clock_gettime(CLOCK_MONOTONIC,&end);
times9.push_back(time_delta_us_timespec(start,end));
}
printf("[N:%d] Completely Fused - GPU Compute: ",NUM_TIME_STEPS_TEST); printStats<PRINT_DISTRIBUTIONS_GLOBAL>(×9);
printf("\n");
printf("----------------------------------------\n");
}
// free all
free(h_vaf); free(h_cdu); free(h_dqdd); free(h_mem_const); delete testTraj;
free(h_mem_vol_split); free(h_Minv); free(h_mem_vol_fused); free(h_mem_vol_completely_fused);
gpuErrchk(cudaFree(d_dqdd)); gpuErrchk(cudaFree(d_cdu)); gpuErrchk(cudaFree(d_mem_const));
gpuErrchk(cudaFree(d_mem_vol_split)); gpuErrchk(cudaFree(d_mem_vol_fused)); gpuErrchk(cudaFree(d_mem_vol_completely_fused));
for(int i=0; i<NUM_STREAMS; i++){gpuErrchk(cudaStreamDestroy(streams[i]));} free(streams);
}
int main(void){
test<float,TEST_ITERS_GLOBAL,CPU_THREADS_GLOBAL,1,MPC_MODE_GLOBAL,VEL_DAMPING_GLOBAL>();
#if !TEST_FOR_EQUIVALENCE
test<float,TEST_ITERS_GLOBAL,CPU_THREADS_GLOBAL,10,MPC_MODE_GLOBAL,VEL_DAMPING_GLOBAL>();
test<float,TEST_ITERS_GLOBAL,CPU_THREADS_GLOBAL,16,MPC_MODE_GLOBAL,VEL_DAMPING_GLOBAL>();
test<float,TEST_ITERS_GLOBAL,CPU_THREADS_GLOBAL,32,MPC_MODE_GLOBAL,VEL_DAMPING_GLOBAL>();
test<float,TEST_ITERS_GLOBAL,CPU_THREADS_GLOBAL,64,MPC_MODE_GLOBAL,VEL_DAMPING_GLOBAL>();
test<float,TEST_ITERS_GLOBAL,CPU_THREADS_GLOBAL,128,MPC_MODE_GLOBAL,VEL_DAMPING_GLOBAL>();
test<float,TEST_ITERS_GLOBAL,CPU_THREADS_GLOBAL,256,MPC_MODE_GLOBAL,VEL_DAMPING_GLOBAL>();
#endif
}