Mercurial > hg > Members > Moririn
annotate src/parallel_execution/CUDAWorker.cbc @ 316:54d203daf06b
CUDAtwice.cbc is called.
| author | Shinji KONO <kono@ie.u-ryukyu.ac.jp> |
|---|---|
| date | Wed, 15 Feb 2017 16:25:23 +0900 |
| parents | 1839586f5b41 |
| children | 51aa65676e37 |
| rev | line source |
|---|---|
| 303 | 1 #include <stdio.h> |
| 2 #include <sys/time.h> | |
| 3 #include <string.h> | |
| 4 #include <stdlib.h> | |
|
304
9755206813cb
helper_string.h for ANSI C
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
303
diff
changeset
|
5 #include <libkern/OSAtomic.h> |
| 305 | 6 |
| 7 // includes, project | |
| 8 #include <driver_types.h> | |
| 9 #include <cuda_runtime.h> | |
| 303 | 10 #include <cuda.h> |
| 11 #include "helper_cuda.h" | |
| 12 | |
| 302 | 13 #include "../context.h" |
|
95
3e28ee215c0e
modify twice, use OSAtomiceCompareAndSwap
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
14 |
| 303 | 15 static void start_CUDAworker(Worker* worker); |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
16 static void cudaInit(struct CUDAWorker *cudaWorker,int phase) ; |
| 233 | 17 |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
18 volatile int cuda_initialized = 0; |
| 313 | 19 |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
20 Worker* createCUDAWorker(struct Context* context, int id, Queue* queue, TaskManagerImpl *im) { |
| 233 | 21 struct Worker* worker = ALLOC(context, Worker); |
| 305 | 22 struct CUDAWorker* cudaWorker = new CUDAWorker(); |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
23 |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
24 cudaInit(cudaWorker,0); |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
25 |
| 305 | 26 worker->worker = (union Data*)cudaWorker; |
| 244 | 27 worker->tasks = queue; |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
28 printf("createCUDAWorker %p\n",queue); |
| 305 | 29 cudaWorker->id = id; |
| 313 | 30 worker->shutdown = C_shutdownCUDAWorker; |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
31 // pthread_create(&worker->worker->CUDAWorker.thread, NULL, (void*)&start_CUDAworker, worker); |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
32 if (im) { |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
33 printf("im->worker %p\n",im->workers); |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
34 im->workers[0] = worker; |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
35 } |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
36 cuda_initialized = 1; |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
37 start_CUDAworker(worker); |
| 313 | 38 return worker; |
| 39 } | |
| 40 | |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
41 static void cudaInit(struct CUDAWorker *cudaWorker,int phase) { |
|
312
7dd5a7d52a67
USE_CUDAWorker flag only for CUDAtwice
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
305
diff
changeset
|
42 // initialize and load kernel |
|
7dd5a7d52a67
USE_CUDAWorker flag only for CUDAtwice
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
305
diff
changeset
|
43 cudaWorker->num_stream = 1; // number of stream |
| 314 | 44 // cudaWorker->stream = NEWN(cudaWorker->num_stream, CUstream ); |
| 45 printf("cudaInit 1\n"); | |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
46 if (phase==0) |
|
312
7dd5a7d52a67
USE_CUDAWorker flag only for CUDAtwice
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
305
diff
changeset
|
47 checkCudaErrors(cuInit(0)); |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
48 if (phase==0) |
|
312
7dd5a7d52a67
USE_CUDAWorker flag only for CUDAtwice
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
305
diff
changeset
|
49 checkCudaErrors(cuDeviceGet(&cudaWorker->device, 0)); |
| 314 | 50 printf("cudaInit 2\n"); |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
51 if (phase==0) |
|
312
7dd5a7d52a67
USE_CUDAWorker flag only for CUDAtwice
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
305
diff
changeset
|
52 checkCudaErrors(cuCtxCreate(&cudaWorker->cuCtx, CU_CTX_SCHED_SPIN, cudaWorker->device)); |
| 314 | 53 printf("cudaInit 3\n"); |
| 54 // if (cudaWorker->num_stream) { | |
| 55 // for (int i=0;i<cudaWorker->num_stream;i++) | |
| 56 // checkCudaErrors(cuStreamCreate(&cudaWorker->stream[i],0)); | |
| 57 // } | |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
58 CUdeviceptr devA; |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
59 checkCudaErrors(cuMemAlloc(&devA, 16)); |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
60 |
| 314 | 61 printf("cudaInit done\n"); |
| 183 | 62 } |
| 63 | |
| 303 | 64 static void start_CUDAworker(Worker* worker) { |
| 305 | 65 CUDAWorker* cudaWorker = (CUDAWorker*)worker->worker; |
| 66 cudaWorker->context = NEW(struct Context); | |
| 67 initContext(cudaWorker->context); | |
| 68 Gearef(cudaWorker->context, Worker)->worker = (union Data*)worker; | |
| 303 | 69 |
| 305 | 70 goto meta(cudaWorker->context, C_taskReceiveCUDAWorker); |
| 233 | 71 } |
| 72 | |
| 305 | 73 __code taskReceiveCUDAWorker(struct Worker* worker,struct Queue* queue) { |
|
247
ce262b2c1daf
Fix createTask for main
Tatsuki IHA <e125716@ie.u-ryukyu.ac.jp>
parents:
244
diff
changeset
|
74 queue->queue = (union Data*)worker->tasks; |
| 305 | 75 queue->next = C_getTaskCUDA; |
|
248
1ede5390cda2
Fix segmentation fault but not multi thread running
Tatsuki IHA <e125716@ie.u-ryukyu.ac.jp>
parents:
247
diff
changeset
|
76 goto meta(context, worker->tasks->take); |
| 227 | 77 } |
| 78 | |
| 302 | 79 __code taskReceiveCUDAWorker_stub(struct Context* context) { |
| 80 goto taskReceiveCUDAWorker(context, &Gearef(context, Worker)->worker->Worker, Gearef(context, Queue)); | |
|
222
77faa28128b4
Add taskSend for TaskManager
Tatsuki IHA <e125716@ie.u-ryukyu.ac.jp>
parents:
217
diff
changeset
|
81 } |
|
77faa28128b4
Add taskSend for TaskManager
Tatsuki IHA <e125716@ie.u-ryukyu.ac.jp>
parents:
217
diff
changeset
|
82 |
| 305 | 83 __code getTaskCUDA(struct Worker* worker, struct Context* task) { |
| 240 | 84 if (!task) |
| 236 | 85 return; // end thread |
|
316
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
86 // if (cuda_initialized==0 || 1) { |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
87 // CUDAWorker* cudaWorker = (CUDAWorker*)worker->worker; |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
88 // cudaInit(cudaWorker,1); |
|
54d203daf06b
CUDAtwice.cbc is called.
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
314
diff
changeset
|
89 // } |
| 314 | 90 worker->taskReceive = C_taskReceiveCUDAWorker; |
| 230 | 91 task->worker = worker; |
| 305 | 92 enum Code taskCg = task->next; |
| 93 task->next = C_odgCommitCUDA; // set CG after task exec | |
| 94 goto meta(task, taskCg); | |
|
169
ea7b11f3e717
Using Queue Interface
Tatsuki IHA <e125716@ie.u-ryukyu.ac.jp>
parents:
149
diff
changeset
|
95 } |
|
109
059b26a250cc
Change put_queue process
Tatsuki IHA <e125716@ie.u-ryukyu.ac.jp>
parents:
103
diff
changeset
|
96 |
| 305 | 97 __code getTaskCUDA_stub(struct Context* context) { |
|
247
ce262b2c1daf
Fix createTask for main
Tatsuki IHA <e125716@ie.u-ryukyu.ac.jp>
parents:
244
diff
changeset
|
98 Worker* worker = &Gearef(context,Worker)->worker->Worker; |
|
260
6b5444bbea8a
generated stub no comiple errors
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
257
diff
changeset
|
99 struct Context* task = &Gearef(context, Queue)->data->Context; |
| 305 | 100 goto getTaskCUDA(context, worker, task); |
| 101 } | |
| 102 | |
| 103 __code odgCommitCUDA(struct LoopCounter* loopCounter, struct Queue* queue, struct Context* task) { | |
| 104 int i = loopCounter->i ; | |
| 105 if(task->odg + i < task->maxOdg) { | |
| 106 queue->queue = (union Data*)GET_WAIT_LIST(task->data[task->odg+i]); | |
| 107 queue->next = C_odgCommitCUDA1; | |
| 108 goto meta(context, queue->queue->Queue.take); | |
| 109 } | |
| 110 loopCounter->i = 0; | |
| 111 goto meta(context, C_taskReceiveCUDAWorker); | |
|
95
3e28ee215c0e
modify twice, use OSAtomiceCompareAndSwap
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
112 } |
| 101 | 113 |
| 305 | 114 __code odgCommitCUDA_stub(struct Context* context) { |
| 115 struct Context* workerContext = context->worker->worker->CUDAWorker.context; | |
| 116 goto odgCommitCUDA(workerContext, | |
| 117 Gearef(workerContext, LoopCounter), | |
| 118 Gearef(workerContext, Queue), | |
| 119 context); | |
| 98 | 120 } |
| 305 | 121 |
| 122 __code odgCommitCUDA1(struct TaskManager* taskManager, struct Context* task) { | |
| 123 if(__sync_fetch_and_sub(&task->idgCount, 1)) { | |
| 124 if(task->idgCount == 0) { | |
| 125 taskManager->taskManager = (union Data*)task->taskManager; | |
| 126 taskManager->context = task; | |
| 127 taskManager->next = C_odgCommitCUDA; | |
| 128 goto meta(context, task->taskManager->spawn); | |
| 129 } | |
| 130 } else { | |
| 131 goto meta(context, C_odgCommitCUDA1); | |
| 132 } | |
| 133 } | |
| 227 | 134 |
| 305 | 135 __code odgCommitCUDA1_stub(struct Context* context) { |
| 136 struct Context* task = &Gearef(context, Queue)->data->Context; | |
| 137 goto odgCommitCUDA1(context, | |
| 138 Gearef(context, TaskManager), | |
| 139 task); | |
| 140 | |
| 141 } | |
| 142 | |
| 143 | |
| 144 __code shutdownCUDAWorker(struct Context* context, CUDAWorker* worker) { | |
| 314 | 145 // for (int i=0;i<worker->num_stream;i++) |
| 146 // checkCudaErrors(cuStreamDestroy(worker->stream[i])); | |
| 305 | 147 checkCudaErrors(cuCtxDestroy(worker->cuCtx)); |
| 230 | 148 } |
| 227 | 149 |
| 303 | 150 __code shutdownCUDAWorker_stub(struct Context* context) { |
| 305 | 151 CUDAWorker* worker = (CUDAWorker *)GearImpl(context, Worker, worker); |
| 303 | 152 goto shutdownCUDAWorker(context,worker); |
| 227 | 153 } |