Mercurial > hg > Game > Cerium

changeset 1553:21ddbc06dcde draft

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
fix GpuScheduler
author Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
date Sun, 17 Feb 2013 19:42:36 +0900
parents 40a554d45fc6
children bd97edaed523
files TaskManager/Gpu/GpuScheduler.cc example/fft/task_init.cc example/many_task/main.cc
diffstat 3 files changed, 26 insertions(+), 22 deletions(-) [+]
line wrap: on
line diff
--- a/TaskManager/Gpu/GpuScheduler.cc	Sun Feb 17 14:51:45 2013 +0900
+++ b/TaskManager/Gpu/GpuScheduler.cc	Sun Feb 17 19:42:36 2013 +0900
@@ -45,7 +45,11 @@
         exit(EXIT_FAILURE);
     }
     context = clCreateContext(NULL, 1, &device_id, NULL, NULL, &ret);
-    command_queue = clCreateCommandQueue(context, device_id, 0, &ret);
+    command_queue = clCreateCommandQueue(context, device_id, 0,&ret);
+    if (ret<0) {
+        const char *msg=convert_error_status(ret);
+        error(msg);
+    }
 }
 
 GpuScheduler::~GpuScheduler()
@@ -68,9 +72,7 @@
     int cur = 0;
     memaddr reply[2];
     cl_kernel *kernel = new cl_kernel[2];
-    cl_event event[2];
-    event[0] = NULL;
-    event[1] = NULL;
+    cl_event *event = new cl_event[2];
     cl_mem *memin[2];
     cl_mem *memout[2];
     HTask::htask_flag flag;
@@ -142,9 +144,14 @@
 
                 // set arg count
                 cl_mem memparam = clCreateBuffer(context, CL_MEM_READ_ONLY,
-                                                 sizeof(memaddr)*nextTask->param_count, NULL, NULL);
-                ret = clEnqueueWriteBuffer(command_queue, memparam, CL_TRUE, 0,
-                                           sizeof(memaddr)*nextTask->param_count,nextTask->param(0), 0, NULL, NULL);
+                                                 sizeof(memaddr)*nextTask->param_count, NULL, &ret);
+                if (ret<0) {
+                    const char *msg=convert_error_status(ret);
+                    error(msg);
+                }
+
+                ret = clEnqueueWriteBuffer(command_queue, memparam, CL_TRUE, 0,sizeof(memaddr)*nextTask->param_count,
+                                           nextTask->param(0), 0, NULL, NULL);
                 if (ret<0) {
                     const char *msg=convert_error_status(ret);
                     error(msg);
@@ -230,30 +237,27 @@
                     const char *msg=convert_error_status(ret);
                     error(msg);
                 }
-                // ndrange flagが0ならdim,global_work_size[0],local_work_size[0] = 1で固定に
-                // clEnqueueNDRange
-                // (command_queue, kernel[cur], dim, NULL,global_work_size[0],local_work_size[0],NULL&ev);
 
                 for(int i=0;i<nextTask->outData_count;i++) { // read output data
                     ListElement *output_buf = flag.flip? nextTask->inData(i) :nextTask->outData(i);
                     if (output_buf->size==0) break;
                     ret = clEnqueueReadBuffer(command_queue, memout[cur][i], CL_TRUE, 0,
-                                              output_buf->size, output_buf->addr, 0, NULL, NULL);
+                                              output_buf->size, output_buf->addr, 0, NULL, &event[cur]);
                     if (ret<0) {
                         const char *msg=convert_error_status(ret);
                         error(msg);
                     }
                 }
-                
-                ret = clEnqueueMarkerWithWaitList(command_queue,0,NULL,&event[cur]);
+                    
                 if (ret<0) {
                     const char *msg=convert_error_status(ret);
                     error(msg);
                 }
                 reply[cur] = (memaddr)tasklist->waiter;
                 //clFlush(command_queue); // waiting for queued task
-                if (event[1-cur] != NULL) {
+                if (event[1-cur] != 0) {
                     ret=clWaitForEvents(1,&event[1-cur]);
+                    event[1-cur]=NULL;
                 }
                 if (ret<0) {
                     const char *msg=convert_error_status(ret);
@@ -275,7 +279,7 @@
             }
         }
         clFlush(command_queue); // waiting for queued task
-        //ret=clWaitForEvents(1,&event[1-cur]);
+        clFinish(command_queue); // waiting for queued task
         connector->mail_write(reply[1-cur]);
         
         connector->mail_write((memaddr)MY_SPE_STATUS_READY);
--- a/example/fft/task_init.cc	Sun Feb 17 14:51:45 2013 +0900
+++ b/example/fft/task_init.cc	Sun Feb 17 19:42:36 2013 +0900
@@ -15,12 +15,12 @@
 task_init(void)
 {
 #ifdef __CERIUM_GPU__
-    GpuSchedRegister(SPIN_FACT, "gpu/fft.cl", "spinFact");
-    GpuSchedRegister(BIT_REVERSE, "gpu/fft.cl", "bitReverse");
-    GpuSchedRegister(NORMALIZATION, "gpu/fft.cl", "norm");
-    GpuSchedRegister(BUTTERFLY, "gpu/fft.cl", "butterfly");
-    GpuSchedRegister(TRANSEPOSE, "gpu/fft.cl", "transpose");
-    GpuSchedRegister(HIGH_PASS_FILTER, "gpu/fft.cl", "highPassFilter");
+    GpuSchedRegister(SPIN_FACT, "gpu/spinFact.cl", "spinFact");
+    GpuSchedRegister(BIT_REVERSE, "gpu/bitReverse.cl", "bitReverse");
+    GpuSchedRegister(NORMALIZATION, "gpu/norm.cl", "norm");
+    GpuSchedRegister(BUTTERFLY, "gpu/butterfly.cl", "butterfly");
+    GpuSchedRegister(TRANSEPOSE, "gpu/transpose.cl", "transpose");
+    GpuSchedRegister(HIGH_PASS_FILTER, "gpu/passFilter.cl", "highPassFilter");
 #else
     SchedRegisterTask(QUICK_SORT, QuickSort);
     SchedRegisterTask(SPIN_FACT,spinFact);
--- a/example/many_task/main.cc	Sun Feb 17 14:51:45 2013 +0900
+++ b/example/many_task/main.cc	Sun Feb 17 19:42:36 2013 +0900
@@ -77,7 +77,7 @@
 static void
 show_data(void)
 {
-    /*
+    /* 
     puts("-----------------------------------------------");
     for(int i = 0; i < sorter.data_length; i++) {
         printf("data[%02d].index = %d\n", i, sorter.data[i].index);