void Scheduler::issueInstruction(InstRef &IR) {
// Release buffered resources.
const InstrDesc &Desc = IR.getInstruction()->getDesc();
releaseBuffers(Desc.Buffers);
notifyReleasedBuffers(Desc.Buffers);
// Issue IS to the underlying pipelines and notify listeners.
SmallVector<std::pair<ResourceRef, double>, 4> Pipes;
issueInstructionImpl(IR, Pipes);
notifyInstructionIssued(IR, Pipes);
if (IR.getInstruction()->isExecuted())
notifyInstructionExecuted(IR);
}
NuPlayerVPPProcessor::~NuPlayerVPPProcessor() {
quitThread();
if (mWorker != NULL) {
delete mWorker;
mWorker = NULL;
}
if (mThreadRunning == false) {
releaseBuffers();
}
mNuPlayerVPPProcessor = NULL;
LOGI("===== VPPInputCount = %d =====", mInputCount);
}
void UdrServerDataStream::actOnReceive(IpcConnection *conn)
{
const char *moduleName = "UdrServerDataStream::actOnReceive()";
IpcMessageObjType t;
NABoolean somethingArrived = FALSE;
//
// A note about control flow in this method. It was originally
// written under the assumption that only one message arrives at a
// time. It would call getNextReceiveMsg() only once, then process
// incoming objects, then return. It turns out this creates a memory
// leak. Internally within the stream, the incoming message buffer
// never moved from the "receive" list to the "in use" list and
// never got cleaned up. Now this method calls getNextReceiveMsg()
// until it returns FALSE. The result is that after the incoming
// message has been processed, the next getNextReceiveMsg() call
// will move the request buffer to the "in use" list where it later
// gets cleaned up by a call to cleanupBuffers() or
// releaseBuffers().
//
while (getNextReceiveMsg(t))
{
somethingArrived = TRUE;
while (getNextObjType(t))
{
switch (t)
{
case UDR_MSG_DATA_HEADER:
{
// Do nothing for now except extract the object from the stream
UdrDataHeader *h = new (receiveMsgObj()) UdrDataHeader(this);
} // case UDR_MSG_DATA_HEADER
break;
case UDR_MSG_CONTINUE_REQUEST:
{
// Extract the object from the stream and call SPInfo's work()
UdrContinueMsg *m = new (receiveMsgObj()) UdrContinueMsg(this);
if(spinfo_->getParamStyle() == COM_STYLE_TM)
spinfo_->workTM();
else
spinfo_->work();
} // case UDR_MSG_CONTINUE_REQUEST
break;
case UDR_MSG_DATA_REQUEST:
{
UdrDataBuffer *request = new (receiveMsgObj()) UdrDataBuffer(this);
if (udrGlob_->verbose_ &&
udrGlob_->traceLevel_ >= TRACE_DETAILS &&
udrGlob_->showMain_)
{
ServerDebug( "");
ServerDebug("[UdrServ (%s)] Invoke Request has %ld tupps.",
moduleName,
request->getSqlBuffer()->getTotalTuppDescs());
ServerDebug("[UdrServ (%s)] Invoke Request SQL Buffer",
moduleName);
displaySqlBuffer(request->getSqlBuffer(),
(Lng32) request->getSqlBufferLength());
}
udrGlob_->numReqInvokeSP_++;
// Let the SPInfo process the request
spinfo_->setCurrentRequest(request);
spinfo_->work();
} // case UDR_MSG_DATA_REQUEST
break;
case UDR_MSG_TMUDF_DATA_HEADER:
{
// Extract the object. UDR Handle and RS Handle are
// interesting things in this message.
UdrTmudfDataHeader *h = new (receiveMsgObj()) UdrTmudfDataHeader(this);
UdrHandle udrHandle = h->getHandle();
NABoolean anyMore = getNextObjType(t);
UDR_ASSERT(anyMore, "DATA_REQUEST must follow TMUDF DATA_HEADER");
UDR_ASSERT(t == UDR_MSG_DATA_REQUEST,
"DATA_REQUEST must follow TMUDF DATA_HEADER");
UdrDataBuffer *request = new (receiveMsgObj())
UdrDataBuffer(this,
FALSE); //Avoid unpack.
// Let the SPInfo process the request
spinfo_->setCurrentRequest(request);
spinfo_->workTM();
}
break;
case UDR_MSG_RS_DATA_HEADER:
{
// Extract the object. UDR Handle and RS Handle are
// interesting things in this message.
UdrRSDataHeader *h = new (receiveMsgObj()) UdrRSDataHeader(this);
UdrHandle udrHandle = h->getHandle();
RSHandle rsHandle = h->getRSHandle();
NABoolean anyMore = getNextObjType(t);
UDR_ASSERT(anyMore, "DATA_REQUEST must follow RS_DATA_HEADER");
UDR_ASSERT(t == UDR_MSG_DATA_REQUEST,
"DATA_REQUEST must follow RS_DATA_HEADER");
UdrDataBuffer *request = new (receiveMsgObj()) UdrDataBuffer(this);
udrGlob_->numReqRSFetch_++;
processAnRSFetchMessage(udrGlob_, *this, udrHandle,
rsHandle, request);
// We need the request buffer in a state where the stream
// knows it can be freed. We call SqlBuffer::bufferFull() to
// accomplish this even though the method name is a bit
// misleading.
SqlBuffer *sqlBuf = request->getSqlBuffer();
UDR_ASSERT(sqlBuf,
"UDR request buffer is corrupt or contains no data");
sqlBuf->bufferFull();
}
break;
case UDR_MSG_RS_CONTINUE:
{
UdrRSContinueMsg *rsContinue =
new (receiveMsgObj()) UdrRSContinueMsg(this);
udrGlob_->numReqRSContinue_++;
processAnRSContinueMessage(udrGlob_, *this, *rsContinue);
}
break;
default:
{
UDR_ASSERT(FALSE,
"Unknown message type arrived on UDR data stream");
} // default
break;
} // switch (t)
} // while (getNextObjType(t))
} // while (getNextReceiveMsg(t))
// Make sure all reply buffers have been given to the connection. If
// the only objects that arrived during this callback were continue
// requests, then this action allows reply buffers associated with
// those continue requests to propagate out of the stream and onto
// the connection.
//
// If numOfInputBuffers() is > 0 then we do not need to do anything
// at this point. This callback will be invoked again when the
// incoming message is complete and ready to be processed.
Lng32 numInputBuffers = numOfInputBuffers();
if (somethingArrived && numInputBuffers == 0 &&
spinfo_->getCurrentRequest() == NULL)
{
responseDone();
if (udrGlob_->verbose_ &&
udrGlob_->traceLevel_ >= TRACE_IPMS &&
udrGlob_->showMain_)
{
ServerDebug("[UdrServ (%s)] All messages marked as replied", moduleName);
}
// Cleanup any unpacked message buffers containing only objects
// that are no longer in use, as determined by the virtual method
// IpcMessageObj::msgObjIsFree()
releaseBuffers(); // Do final garbage collection
}
} // UdrServerDataStream::actOnReceive()
~PyBufferWrapper(void)
{
releaseBuffers();
}
PyBufferWrapper(PyObject *const data, const uint32_t count, const bool writable)
:
m_count(count),
m_length(UINT32_MAX)
{
memset(m_buffers, 0, sizeof(double*) * MAX_BUFFERS);
memset(m_views, 0, sizeof(Py_buffer) * MAX_BUFFERS);
if (!((data) && PyTuple_Check(data)))
{
PY_EXCEPTION("Input Python object does not appear to be a tuple!");
throw std::invalid_argument("Invalid buffer object!");
}
if (!((count > 0) && (count <= MAX_BUFFERS) && (static_cast<uint32_t>(PyTuple_Size(data)) >= count)))
{
PY_EXCEPTION("Input tuple does not have the required minimum size!");
throw std::invalid_argument("Invalid buffer count!");
}
PyObject *pyArrays[MAX_BUFFERS];
for (uint32_t c = 0; c < count; ++c)
{
if (!(pyArrays[c] = PyTuple_GetItem(data, static_cast<Py_ssize_t>(c))))
{
PY_EXCEPTION("Failed to obtain the n-th element of the given input tuple!");
throw std::invalid_argument("PyTuple_GetItem() error!");
}
if (PyObject_IsInstance(pyArrays[c], g_arrayClass) <= 0)
{
PY_EXCEPTION("The n-th element of the given tuple is not an array object!");
throw std::invalid_argument("PyObject_IsInstance() error!");
}
}
for (uint32_t c = 0; c < count; ++c)
{
if (!PyObject_CheckBuffer(pyArrays[c]))
{
PY_EXCEPTION("The provided array object does not support the 'buffer' protocol!");
releaseBuffers();
throw std::invalid_argument("PyObject_CheckBuffer() error!");
}
if (PyObject_GetBuffer(pyArrays[c], &m_views[c], (writable ? PyBUF_WRITABLE : 0)) < 0)
{
PY_EXCEPTION("Failed to get the internal buffer from array object!");
releaseBuffers();
throw std::invalid_argument("PyObject_GetBuffer() error!");
}
if (m_views[c].itemsize != sizeof(double))
{
PY_EXCEPTION("The provided array has an invalid item size! Not 'double' array?");
releaseBuffers();
throw std::invalid_argument("Py_buffer.itemsize is invalid!");
}
if ((m_length = std::min(m_length, static_cast<uint32_t>(m_views[c].len / m_views[c].itemsize))) < 1)
{
PY_EXCEPTION("The size of the provided buffers is way too small!");
releaseBuffers();
throw std::invalid_argument("Py_buffer.len is invalid!");
}
if (!(m_buffers[c] = reinterpret_cast<double*>(m_views[c].buf)))
{
PY_EXCEPTION("Pointer to the arrays's buffer appears to be NULL!");
releaseBuffers();
throw std::invalid_argument("Py_buffer.buf is invalid!");
}
}
}
/** Run the OpenCL kernel */
short runOpenCL(struct benchmark *bench)
{
short j;
#ifdef DEBUG
printf("TEST OPENCL\n");
#endif
/**********************
Initializations
**********************/
error=clGetPlatformIDs(1,&platform,NULL);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Error to get platform ID : %d\n",error);
goto error;
}
error=clGetDeviceIDs(platform,CL_DEVICE_TYPE_GPU,1,&device,NULL);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Error to get device ID : %d\n",error);
goto error;
}
if (error != CL_SUCCESS)
{
fprintf(stderr,"Can't get device info : %d\n",error);
goto error;
}
context = clCreateContext(0,1,&device,NULL,NULL,&error);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Error to create context : %d\n",error);
goto error;
}
size_t maxWorkItemDim;
size_t maxWorkGroupSize;
size_t workItemSize[10];
error = clGetDeviceInfo(device,CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS,sizeof(size_t),&maxWorkItemDim,NULL);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Can't get max work item dimensions : %d\n",error);
goto errorContext;
}
error = clGetDeviceInfo(device,CL_DEVICE_MAX_WORK_ITEM_SIZES,maxWorkItemDim*sizeof(size_t),workItemSize,NULL);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Can't get mwork item sizes : %d\n",error);
goto errorContext;
}
error = clGetDeviceInfo(device,CL_DEVICE_MAX_WORK_GROUP_SIZE,sizeof(size_t),&maxWorkGroupSize,NULL);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Can't get max work item dimensions : %d\n",error);
goto errorContext;
}
queue = clCreateCommandQueue(context, device, CL_QUEUE_PROFILING_ENABLE ,&error);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Error to create command queue : %d\n",error);
goto errorContext;
}
/**********************
Memory allocations
**********************/
long i;
#ifdef DEBUG
printf("Create buffers\n");
#endif
double createBufTime;
createBufTime = createBuffers(context,bench);
/**********************
OpenCL kernel
**********************/
cl_program program;
char *fileContent;
FILE *f;
struct stat fState;
char path[256];
strcpy(path,"Kernels/");
strcat(path,bench->kernel);
stat(path,&fState);
f=fopen(path,"r");
fileContent=malloc(fState.st_size*sizeof(char));
fread(fileContent,sizeof(char),fState.st_size,f);
fclose(f);
program=clCreateProgramWithSource(context,1,(const char**)&fileContent,&fState.st_size,&error);
free(fileContent);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Can't create program : %d\n",error);
goto errorBuffer;
}
/*error=clBuildProgram(program,1,&device,"-cl-fast-relaxed-math",NULL,NULL);*/
error=clBuildProgram(program,1,&device,"",NULL,NULL);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Can't build program : %d\n",error);
goto errorProgram;
}
cl_kernel kernel=clCreateKernel(program,"mainKernel",&error);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Can't create kernel : %d\n",error);
goto errorProgram;
}
/**********************
Launching the kernel
**********************/
#ifdef DEBUG
printf("Set args\n");
#endif
setArgs(kernel,bench);
cl_ulong lStart;
cl_ulong lEnd;
double fTimeInSeconds;
double fFLOPS;
#ifdef DEBUG
printf("Compute\n");
printf("%d\n",bench->worksizeDim);
for(i=0;i<bench->worksizeDim;i++)
{
printf(" GLOBAL -> %d\n",bench->global_ws[i]);
printf(" LOCAL -> %d\n",bench->local_ws[i]);
}
#endif
double writeBufTime;
writeBufTime=writeInputs(queue,bench);
double computeTime=getCurrentTime();
error=clEnqueueNDRangeKernel(queue, kernel,bench->worksizeDim, NULL,bench->global_ws, bench->local_ws,0,NULL,&event);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Can't enqueue kernel : %d\n",error);
goto errorKernel;
}
clFinish(queue);
computeTime= getCurrentTime() - computeTime;
#ifdef DEBUG
printf("Read results\n");
#endif
double readBufTime;
readBufTime=readResults(queue,bench);
error = clGetEventProfilingInfo(event,CL_PROFILING_COMMAND_START,sizeof(cl_ulong),&lStart,NULL);
error |= clGetEventProfilingInfo(event,CL_PROFILING_COMMAND_END ,sizeof(cl_ulong),&lEnd,NULL);
if (error != CL_SUCCESS)
{
fprintf(stderr,"Can't get profiling info : %d\n",error);
goto errorEvent;
}
fTimeInSeconds = ((double)(lEnd-lStart)) / 1000000000.0;
/* Send timing */
write(newsockfd,"t\n",2*sizeof(char));
sprintf(result,"%f\n",createBufTime);
write(newsockfd,result,strlen(result));
sprintf(result,"%f\n",writeBufTime);
write(newsockfd,result,strlen(result));
sprintf(result,"%f\n",fTimeInSeconds);
write(newsockfd,result,strlen(result));
sprintf(result,"%f\n",readBufTime);
write(newsockfd,result,strlen(result));
/*sprintf(result,"%f\n",computeTime);
write(newsockfd,result,strlen(result));
*/
/* Send results */
sendResults(bench);
errorEvent:
clReleaseEvent(event);
errorKernel:
clReleaseKernel(kernel);
/**********************
Cleanup
**********************/
errorProgram:
clReleaseProgram(program);
errorBuffer:
clReleaseCommandQueue(queue);
releaseBuffers(bench);
errorContext:
clReleaseContext(context);
/**
HORRIBLE error processing. It may create memory leaks. It will have to be improved.
*/
error:
return (error != CL_SUCCESS);
}
kit::Sphere::~Sphere()
{
releaseBuffers();
}
