Mercurial > hg > Game > Cerium

changeset 1627:8b2d3ac19991 draft

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
fix
author Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
date Mon, 03 Jun 2013 17:02:51 +0900
parents 4fed76f4d101
children da928c43c6be
files TaskManager/Cell/CellTaskManagerImpl.h TaskManager/Fifo/FifoTaskManagerImpl.cc TaskManager/kernel/ppe/TaskManager.cc TaskManager/kernel/ppe/TaskManager.h example/fft/main.cc example/fft/ppe/butterfly.cc example/fft/ppe/highPassFilter.cc example/fft/ppe/transpose.cc
diffstat 8 files changed, 30 insertions(+), 48 deletions(-) [+]
line wrap: on
line diff
--- a/TaskManager/Cell/CellTaskManagerImpl.h	Thu May 30 18:35:08 2013 +0900
+++ b/TaskManager/Cell/CellTaskManagerImpl.h	Mon Jun 03 17:02:51 2013 +0900
@@ -12,7 +12,7 @@
 class CellTaskManagerImpl : public TaskManagerImpl {
 public:
     /* constructor */
- CellTaskManagerImpl(int num, Threads *cpus) : TaskManagerImpl(num,gpu) {speThreads = cpus;}
+ CellTaskManagerImpl(int num, int gpu, Threads *cpus) : TaskManagerImpl(num) {gpuNum = gpu; speThreads = cpus;}
     ~CellTaskManagerImpl();
 
     /* variables */
@@ -23,6 +23,7 @@
     FifoTaskManagerImpl *ppeManager;
     int spe_running;
     int spuIdle;
+    int gpuNum;
     /* functions */
     // system
     void init(int spuIdle,int useRefDma, int export_task_log);
--- a/TaskManager/Fifo/FifoTaskManagerImpl.cc	Thu May 30 18:35:08 2013 +0900
+++ b/TaskManager/Fifo/FifoTaskManagerImpl.cc	Mon Jun 03 17:02:51 2013 +0900
@@ -339,7 +339,7 @@
  */
 #ifdef __CERIUM_FIFO__
 TaskManagerImpl*
-create_impl(int num, int useRefDma)
+create_impl(int num, int gpu, int useRefDma)
 {
 #ifdef __CERIUM_PARALLEL__
 	//マルチコアverでコンパイルしたのにかかわらず、
@@ -354,7 +354,7 @@
         return new FifoTaskManagerImpl(num);
 	} else {
 		Threads *cpus = new CpuThreads(num,useRefDma);
-		return new CellTaskManagerImpl(num,cpus);
+		return new CellTaskManagerImpl(num,gpu,cpus);
 	}
 }
 #endif // __CERIUM_FIFO__
--- a/TaskManager/kernel/ppe/TaskManager.cc	Thu May 30 18:35:08 2013 +0900
+++ b/TaskManager/kernel/ppe/TaskManager.cc	Mon Jun 03 17:02:51 2013 +0900
@@ -4,7 +4,7 @@
 #include "Task.h"
 #include "SysFunc.h"
 
-TaskManager::TaskManager(int num, int gpu) : machineNum(num) : gpuNum(gpu)
+TaskManager::TaskManager(int num, int gpu) : machineNum(num) , gpuNum(gpu)
 {
 }
 
@@ -29,7 +29,7 @@
  *
  * てか、普通に TaskManagerImpl に関数持たせた方が早いか・・・？
  */
-extern TaskManagerImpl* create_impl(int,int);
+extern TaskManagerImpl* create_impl(int,int,int);
 
 void
 TaskManager::init(int spuIdle, int export_task_log, int useRefDma=0)
--- a/TaskManager/kernel/ppe/TaskManager.h	Thu May 30 18:35:08 2013 +0900
+++ b/TaskManager/kernel/ppe/TaskManager.h	Mon Jun 03 17:02:51 2013 +0900
@@ -11,7 +11,7 @@
 class TaskManager {
 public:
     /* constructor */
-    TaskManager(int num); // The number of threads
+    TaskManager(int num, int gpu); // The number of threads
     ~TaskManager();
 
     /* variables */
@@ -45,6 +45,7 @@
     void error(const char* str);
 private:
     int machineNum;
+    int gpuNum;
 }  ;
 
 #endif
--- a/example/fft/main.cc	Thu May 30 18:35:08 2013 +0900
+++ b/example/fft/main.cc	Mon Jun 03 17:02:51 2013 +0900
@@ -85,7 +85,6 @@
     size_t gws[2],lws[2];
     int length_dst = n[0]*n[0];
     int length_src = n[0]*n[0];
-    cl_uint dimension = 2;
 
     HTask* brev = manager->create_task(BIT_REVERSE);
     setWorkSize(gws,lws,n[0],n[0]);
@@ -106,23 +105,17 @@
     bfly->set_outData(0,dst,length_dst*sizeof(cl_float2));
     bfly->set_cpu(spe_cpu);
     bfly->wait_for(brev);
-    bfly->iterate(m_,gws[0],gws[1]);
+    bfly->iterate(gws[0],gws[1],m_);
     
     if (direction == inverse) { 
         HTask *norm = manager->create_task(NORMALIZATION);
         setWorkSize(gws,lws,n[0],n[0]);
-        norm->set_param(0,(memaddr)length_dst);
-        norm->set_param(1,(memaddr)dimension);
-        norm->set_param(2,(memaddr)gws[0]);
-        norm->set_param(3,(memaddr)gws[1]);
-        norm->set_param(4,(memaddr)lws[0]);
-        norm->set_param(5,(memaddr)lws[1]);
         norm->set_inData(0, n,sizeof(int));
         norm->set_outData(0, dst, length_dst*sizeof(cl_float2));
         norm->set_cpu(spe_cpu);
         norm->flip();
         norm->wait_for(bfly);
-        norm->spawn();
+        norm->iterate(gws[0],lws[0]);
     }
     return 0;
 }
@@ -150,7 +143,6 @@
 void
 run_start(TaskManager *manager,pgm_t ipgm)
 {
-    int dimension;
     int *n = new int[1];
     n[0] = ipgm.width;
     int *m = new int[1];
@@ -200,22 +192,15 @@
     // Transpose matrix 
     int length_r =n[0] * n[0];
     setWorkSize(gws,lws,n[0]/2,1);
-    dimension = 2;
     for (int i=0;i<2;i++) {
         trns[i]= manager->create_task(TRANSEPOSE);
-        trns[i]->set_param(0, (memaddr)length_r);
-        trns[i]->set_param(1,(memaddr)dimension);
-        trns[i]->set_param(2,(memaddr)gws[0]);
-        trns[i]->set_param(3,(memaddr)gws[1]);
-        trns[i]->set_param(4,(memaddr)lws[0]);
-        trns[i]->set_param(5,(memaddr)lws[1]);
         trns[i]->set_inData(0, rm, length_r*sizeof(cl_float2));
         trns[i]->set_inData(1, n,sizeof(int));
         trns[i]->set_outData(0, xm, length_r*sizeof(cl_float2));
         trns[i]->set_cpu(spe_cpu);
     }
     trns[0]->wait_for(sfac);
-    trns[0]->spawn();
+    trns[0]->iterate(gws[0],lws[0]);
     // Butterfly Operation 
     fftCore(manager, rm, xm, wm, m[0], forward);
 
@@ -224,25 +209,20 @@
     cl_int *radius = new cl_int[1];
     radius[0] = n[0]/8;
     setWorkSize(gws,lws,n[0]/2,1);
-    hpfl->set_param(0, (memaddr)length_r);
-    hpfl->set_param(1,(memaddr)dimension);
-    hpfl->set_param(2,(memaddr)gws[0]);
-    hpfl->set_param(3,(memaddr)gws[1]);
-    hpfl->set_param(4,(memaddr)lws[0]);
-    hpfl->set_param(5,(memaddr)lws[1]);
     hpfl->set_inData(0, n,sizeof(int));
     hpfl->set_inData(1, radius,sizeof(int));
     hpfl->set_outData(0, rm, length_r*sizeof(cl_float2));
     hpfl->set_cpu(spe_cpu);
     hpfl->wait_for(trns[0]);
-    hpfl->spawn();
+    hpfl->iterate(gws[0],lws[0]);
     // Inverse FFT
 
     // Butterfly Operation
     fftCore(manager,xm, rm, wm, m[0], inverse);
 
     // Transpose matrix
-    trns[1]->spawn();
+    setWorkSize(gws,lws,n[0],n[0]);
+    trns[1]->iterate(gws[0],lws[0]);
 
     // Butterfly Operation
 
--- a/example/fft/ppe/butterfly.cc	Thu May 30 18:35:08 2013 +0900
+++ b/example/fft/ppe/butterfly.cc	Mon Jun 03 17:02:51 2013 +0900
@@ -17,9 +17,9 @@
     
     cl_float2* x_out = (cl_float2*)s->get_outputAddr(0);
 
-    long iter = (long)s->get_param(0) + 1;
-    unsigned long gid = (unsigned long)s->get_param(1);
-    unsigned long nid = (unsigned long)s->get_param(2);
+    unsigned long gid = (unsigned long)s->get_param(0);
+    unsigned long nid = (unsigned long)s->get_param(1);
+    long iter = (long)s->get_param(2) + 1;
 
     int butterflySize = 1 << (iter-1);
     int butterflyGrpDist = 1 << iter;
--- a/example/fft/ppe/highPassFilter.cc	Thu May 30 18:35:08 2013 +0900
+++ b/example/fft/ppe/highPassFilter.cc	Mon Jun 03 17:02:51 2013 +0900
@@ -10,13 +10,13 @@
 static int
 highPassFilter(SchedTask* s,void* rbuf,void* wbuf)
 {
-    int* n = (int*)s->get_input(rbuf,0);
-    int* radius = (int*)s->get_input(rbuf,1);
+    int* n = (int*)s->get_inputAddr(0);
+    int* radius = (int*)s->get_inputAddr(1);
 
-    cl_float2* image = (cl_float2*)s->get_output(wbuf,0);
+    cl_float2* image = (cl_float2*)s->get_outputAddr(0);
 
-    unsigned int* xgid = (unsigned int*)s->global_get(FIRSTID);
-    unsigned int* ygid = (unsigned int*)s->global_get(SECONDID);
+    unsigned long xgid = (unsigned long)s->get_param(0);
+    unsigned long ygid = (unsigned long)s->get_param(1);
 
     cl_int2 n_2;
     n_2.x = n_2.y = n[0]>>1;
@@ -25,8 +25,8 @@
     mask.x = mask.y = n[0]-1;
 
     cl_int2 gid;
-    gid.x = (xgid[0] + n_2.x) & mask.x;
-    gid.y = (ygid[0] + n_2.y) & mask.y;
+    gid.x = (xgid + n_2.x) & mask.x;
+    gid.y = (ygid + n_2.y) & mask.y;
 
     cl_int2 diff;
     diff.x = n_2.x - gid.x;
@@ -46,8 +46,8 @@
         window.x = window.y = (int)-1L;
     }
 
-    image[ygid[0]*n[0]+xgid[0]].x = (float)((int)image[ygid[0]*n[0]+xgid[0]].x & window.x);
-    image[ygid[0]*n[0]+xgid[0]].y = (float)((int)image[ygid[0]*n[0]+xgid[0]].y & window.y);
+    image[ygid*n[0]+xgid].x = (float)((int)image[ygid*n[0]+xgid].x & window.x);
+    image[ygid*n[0]+xgid].y = (float)((int)image[ygid*n[0]+xgid].y & window.y);
 
     return 0;
 }
--- a/example/fft/ppe/transpose.cc	Thu May 30 18:35:08 2013 +0900
+++ b/example/fft/ppe/transpose.cc	Mon Jun 03 17:02:51 2013 +0900
@@ -16,11 +16,11 @@
 
     cl_float2* dst = (cl_float2*)s->get_outputAddr(0);
 
-    unsigned int* xgid = (unsigned int*)s->global_get(FIRSTID);
-    unsigned int* ygid = (unsigned int*)s->global_get(SECONDID);
+    unsigned long xgid = (unsigned long)s->get_param(0);
+    unsigned long ygid = (unsigned long)s->get_param(1);
 
-    unsigned int iid = ygid[0] * n[0] + xgid[0];
-    unsigned int oid = xgid[0] * n[0] + ygid[0];
+    unsigned int iid = ygid * n[0] + xgid;
+    unsigned int oid = xgid * n[0] + ygid;
 
     dst[oid] = src[iid];