Mercurial > hg > Members > Moririn
annotate src/parallel_execution/CUDAWorker.cbc @ 305:ec0a5b4fba05
CUDAWorker
| author | Shinji KONO <kono@ie.u-ryukyu.ac.jp> |
|---|---|
| date | Tue, 14 Feb 2017 12:15:58 +0900 |
| parents | 9755206813cb |
| children | 7dd5a7d52a67 |
| 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); |
| 233 | 16 |
|
304
9755206813cb
helper_string.h for ANSI C
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
303
diff
changeset
|
17 Worker* createCUDAWorker(struct Context* context, int id, Queue* queue) { |
| 233 | 18 struct Worker* worker = ALLOC(context, Worker); |
| 305 | 19 struct CUDAWorker* cudaWorker = new CUDAWorker(); |
| 20 worker->worker = (union Data*)cudaWorker; | |
| 244 | 21 worker->tasks = queue; |
| 305 | 22 cudaWorker->id = id; |
| 302 | 23 worker->taskReceive = C_taskReceiveCUDAWorker; |
| 24 worker->shutdown = C_shutdownCUDAWorker; | |
| 25 pthread_create(&worker->worker->CUDAWorker.thread, NULL, (void*)&start_CUDAworker, worker); | |
|
304
9755206813cb
helper_string.h for ANSI C
Shinji KONO <kono@ie.u-ryukyu.ac.jp>
parents:
303
diff
changeset
|
26 return worker; |
| 183 | 27 } |
| 28 | |
| 303 | 29 static void start_CUDAworker(Worker* worker) { |
| 305 | 30 CUDAWorker* cudaWorker = (CUDAWorker*)worker->worker; |
| 31 cudaWorker->context = NEW(struct Context); | |
| 32 initContext(cudaWorker->context); | |
| 33 Gearef(cudaWorker->context, Worker)->worker = (union Data*)worker; | |
| 34 cudaWorker->num_stream = 1; // number of stream | |
| 303 | 35 |
| 36 // initialize and load kernel | |
| 305 | 37 cudaWorker->stream = NEWN(cudaWorker->num_stream, CUstream ); |
| 303 | 38 checkCudaErrors(cuInit(0)); |
| 305 | 39 checkCudaErrors(cuDeviceGet(&cudaWorker->device, 0)); |
| 40 checkCudaErrors(cuCtxCreate(&cudaWorker->cuCtx, CU_CTX_SCHED_SPIN, cudaWorker->device)); | |
| 41 if (cudaWorker->num_stream) { | |
| 42 for (int i=0;i<cudaWorker->num_stream;i++) | |
| 43 checkCudaErrors(cuStreamCreate(&cudaWorker->stream[i],0)); | |
| 303 | 44 } |
| 45 | |
| 305 | 46 goto meta(cudaWorker->context, C_taskReceiveCUDAWorker); |
| 233 | 47 } |
| 48 | |
| 305 | 49 __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
|
50 queue->queue = (union Data*)worker->tasks; |
| 305 | 51 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
|
52 goto meta(context, worker->tasks->take); |
| 227 | 53 } |
| 54 | |
| 302 | 55 __code taskReceiveCUDAWorker_stub(struct Context* context) { |
| 56 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
|
57 } |
|
77faa28128b4
Add taskSend for TaskManager
Tatsuki IHA <e125716@ie.u-ryukyu.ac.jp>
parents:
217
diff
changeset
|
58 |
| 305 | 59 __code getTaskCUDA(struct Worker* worker, struct Context* task) { |
| 240 | 60 if (!task) |
| 236 | 61 return; // end thread |
| 230 | 62 task->worker = worker; |
| 305 | 63 enum Code taskCg = task->next; |
| 64 task->next = C_odgCommitCUDA; // set CG after task exec | |
| 65 goto meta(task, taskCg); | |
|
169
ea7b11f3e717
Using Queue Interface
Tatsuki IHA <e125716@ie.u-ryukyu.ac.jp>
parents:
149
diff
changeset
|
66 } |
|
109
059b26a250cc
Change put_queue process
Tatsuki IHA <e125716@ie.u-ryukyu.ac.jp>
parents:
103
diff
changeset
|
67 |
| 305 | 68 __code getTaskCUDA_stub(struct Context* context) { |
|
247
ce262b2c1daf
Fix createTask for main
Tatsuki IHA <e125716@ie.u-ryukyu.ac.jp>
parents:
244
diff
changeset
|
69 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
|
70 struct Context* task = &Gearef(context, Queue)->data->Context; |
| 305 | 71 goto getTaskCUDA(context, worker, task); |
| 72 } | |
| 73 | |
| 74 __code odgCommitCUDA(struct LoopCounter* loopCounter, struct Queue* queue, struct Context* task) { | |
| 75 int i = loopCounter->i ; | |
| 76 if(task->odg + i < task->maxOdg) { | |
| 77 queue->queue = (union Data*)GET_WAIT_LIST(task->data[task->odg+i]); | |
| 78 queue->next = C_odgCommitCUDA1; | |
| 79 goto meta(context, queue->queue->Queue.take); | |
| 80 } | |
| 81 loopCounter->i = 0; | |
| 82 goto meta(context, C_taskReceiveCUDAWorker); | |
|
95
3e28ee215c0e
modify twice, use OSAtomiceCompareAndSwap
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
diff
changeset
|
83 } |
| 101 | 84 |
| 305 | 85 __code odgCommitCUDA_stub(struct Context* context) { |
| 86 struct Context* workerContext = context->worker->worker->CUDAWorker.context; | |
| 87 goto odgCommitCUDA(workerContext, | |
| 88 Gearef(workerContext, LoopCounter), | |
| 89 Gearef(workerContext, Queue), | |
| 90 context); | |
| 98 | 91 } |
| 305 | 92 |
| 93 __code odgCommitCUDA1(struct TaskManager* taskManager, struct Context* task) { | |
| 94 if(__sync_fetch_and_sub(&task->idgCount, 1)) { | |
| 95 if(task->idgCount == 0) { | |
| 96 taskManager->taskManager = (union Data*)task->taskManager; | |
| 97 taskManager->context = task; | |
| 98 taskManager->next = C_odgCommitCUDA; | |
| 99 goto meta(context, task->taskManager->spawn); | |
| 100 } | |
| 101 } else { | |
| 102 goto meta(context, C_odgCommitCUDA1); | |
| 103 } | |
| 104 } | |
| 227 | 105 |
| 305 | 106 __code odgCommitCUDA1_stub(struct Context* context) { |
| 107 struct Context* task = &Gearef(context, Queue)->data->Context; | |
| 108 goto odgCommitCUDA1(context, | |
| 109 Gearef(context, TaskManager), | |
| 110 task); | |
| 111 | |
| 112 } | |
| 113 | |
| 114 | |
| 115 __code shutdownCUDAWorker(struct Context* context, CUDAWorker* worker) { | |
| 116 for (int i=0;i<worker->num_stream;i++) | |
| 117 checkCudaErrors(cuStreamDestroy(worker->stream[i])); | |
| 118 checkCudaErrors(cuCtxDestroy(worker->cuCtx)); | |
| 230 | 119 } |
| 227 | 120 |
| 303 | 121 __code shutdownCUDAWorker_stub(struct Context* context) { |
| 305 | 122 CUDAWorker* worker = (CUDAWorker *)GearImpl(context, Worker, worker); |
| 303 | 123 goto shutdownCUDAWorker(context,worker); |
| 227 | 124 } |