From 6032f36b10fd30a4df51b7b5a296b552a8f0d901 Mon Sep 17 00:00:00 2001
From: Hashim Sharif <hsharif3@miranda.cs.illinois.edu>
Date: Thu, 1 Apr 2021 16:59:25 -0500
Subject: [PATCH] Making OpenCL code conditional on HPVM_USE_OPENCL
---
hpvm/projects/hpvm-rt/hpvm-rt.cpp | 245 +++++++++++++++++++++++++++++-
1 file changed, 241 insertions(+), 4 deletions(-)
diff --git a/hpvm/projects/hpvm-rt/hpvm-rt.cpp b/hpvm/projects/hpvm-rt/hpvm-rt.cpp
index dff4f0e4a2..f0716378fe 100644
--- a/hpvm/projects/hpvm-rt/hpvm-rt.cpp
+++ b/hpvm/projects/hpvm-rt/hpvm-rt.cpp
@@ -1,4 +1,6 @@
-#include <CL/cl.h>
+
+//#define HPVM_USE_OPENCL 1
+
#include <algorithm>
#include <cassert>
#include <cstdio>
@@ -8,9 +10,14 @@
#include <map>
#include <pthread.h>
#include <string>
-
#include <unistd.h>
+#ifdef HPVM_USE_OPENCL
+
+#include <CL/cl.h>
+
+#endif
+
#if _POSIX_VERSION >= 200112L
#include <sys/time.h>
#endif
@@ -40,6 +47,9 @@ typedef struct {
std::vector<CircularBuffer<uint64_t> *> *isLastInputBuffers;
} DFNodeContext_CPU;
+
+#ifdef HPVM_USE_OPENCL
+
typedef struct {
cl_context clOCLContext;
cl_command_queue clCommandQue;
@@ -51,6 +61,9 @@ cl_context globalOCLContext;
cl_device_id *clDevices;
cl_command_queue globalCommandQue;
+#endif
+
+
MemTracker MTracker;
vector<DFGDepth> DStack;
// Mutex to prevent concurrent access by multiple thereads in pipeline
@@ -59,6 +72,8 @@ pthread_mutex_t ocl_mtx;
#define NUM_TESTS 1
hpvm_TimerSet kernel_timer;
+#ifdef HPVM_USE_OPENCL
+
static const char *getErrorString(cl_int error) {
switch (error) {
// run-time and JIT compiler errors
@@ -209,6 +224,15 @@ static inline void checkErr(cl_int err, cl_int success, const char *name) {
}
}
+#endif
+
+
+void openCLAbort(){
+ cout <<" ERROR: OpenCL NOT found!. Please Recompile with OpenCL - Make sure to have OpenCL on System \n ";
+ abort();
+}
+
+
/************************* Depth Stack Routines ***************************/
void llvm_hpvm_cpu_dstack_push(unsigned n, uint64_t limitX, uint64_t iX,
@@ -260,6 +284,9 @@ uint64_t llvm_hpvm_cpu_getDimInstance(unsigned level, unsigned dim) {
/********************** Memory Tracking Routines **************************/
void llvm_hpvm_track_mem(void *ptr, size_t size) {
+
+#ifdef HPVM_USE_OPENCL
+
DEBUG(cout << "Start tracking memory: " << ptr << flush << "\n");
MemTrackerEntry *MTE = MTracker.lookup(ptr);
if (MTE != NULL) {
@@ -269,9 +296,19 @@ void llvm_hpvm_track_mem(void *ptr, size_t size) {
DEBUG(cout << "Inserting ID " << ptr << " in the MemTracker Table\n");
MTracker.insert(ptr, size, MemTrackerEntry::HOST, ptr);
DEBUG(MTracker.print());
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
void llvm_hpvm_untrack_mem(void *ptr) {
+
+#ifdef HPVM_USE_OPENCL
+
DEBUG(cout << "Stop tracking memory: " << ptr << flush << "\n");
MemTrackerEntry *MTE = MTracker.lookup(ptr);
if (MTE == NULL) {
@@ -284,11 +321,22 @@ void llvm_hpvm_untrack_mem(void *ptr) {
clReleaseMemObject((cl_mem)MTE->getAddress());
MTracker.remove(ptr);
DEBUG(MTracker.print());
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
+
+#ifdef HPVM_USE_OPENCL
+
static void *llvm_hpvm_ocl_request_mem(void *ptr, size_t size,
DFNodeContext_OCL *Context, bool isInput,
bool isOutput) {
+
pthread_mutex_lock(&ocl_mtx);
DEBUG(cout << "[OCL] Request memory: " << ptr
<< " for context: " << Context->clOCLContext << flush << "\n");
@@ -343,13 +391,20 @@ static void *llvm_hpvm_ocl_request_mem(void *ptr, size_t size,
DEBUG(MTracker.print());
pthread_mutex_unlock(&ocl_mtx);
return d_input;
+
}
+#endif
+
+
void *llvm_hpvm_cpu_argument_ptr(void *ptr, size_t size) {
return llvm_hpvm_request_mem(ptr, size);
}
void *llvm_hpvm_request_mem(void *ptr, size_t size) {
+
+#ifdef HPVM_USE_OPENCL
+
pthread_mutex_lock(&ocl_mtx);
DEBUG(cout << "[CPU] Request memory: " << ptr << flush << "\n");
MemTrackerEntry *MTE = MTracker.lookup(ptr);
@@ -386,6 +441,13 @@ void *llvm_hpvm_request_mem(void *ptr, size_t size) {
DEBUG(MTracker.print());
pthread_mutex_unlock(&ocl_mtx);
return ptr;
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
/*************************** Timer Routines **********************************/
@@ -419,6 +481,9 @@ get_last_async(struct hpvm_TimerSet *timers) {
}
static void insert_marker(struct hpvm_TimerSet *tset, enum hpvm_TimerID timer) {
+
+#ifdef HPVM_USE_OPENCL
+
cl_int ciErrNum = CL_SUCCESS;
struct hpvm_async_time_marker_list **new_event = &(tset->async_markers);
@@ -441,10 +506,20 @@ static void insert_marker(struct hpvm_TimerSet *tset, enum hpvm_TimerID timer) {
if (ciErrNum != CL_SUCCESS) {
fprintf(stderr, "Error Enqueueing Marker!\n");
}
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
static void insert_submarker(struct hpvm_TimerSet *tset, char *label,
enum hpvm_TimerID timer) {
+
+#ifdef HPVM_USE_OPENCL
+
cl_int ciErrNum = CL_SUCCESS;
struct hpvm_async_time_marker_list **new_event = &(tset->async_markers);
@@ -467,10 +542,20 @@ static void insert_submarker(struct hpvm_TimerSet *tset, char *label,
if (ciErrNum != CL_SUCCESS) {
fprintf(stderr, "Error Enqueueing Marker!\n");
}
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
/* Assumes that all recorded events have completed */
static hpvm_Timestamp record_async_times(struct hpvm_TimerSet *tset) {
+
+#ifdef HPVM_USE_OPENCL
+
struct hpvm_async_time_marker_list *next_interval = NULL;
struct hpvm_async_time_marker_list *last_marker = get_last_async(tset);
hpvm_Timestamp total_async_time = 0;
@@ -517,6 +602,13 @@ static hpvm_Timestamp record_async_times(struct hpvm_TimerSet *tset) {
next_interval->timerID = INVALID_TIMERID;
return total_async_time;
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
static void accumulate_time(hpvm_Timestamp *accum, hpvm_Timestamp start,
@@ -733,6 +825,9 @@ void hpvm_AddSubTimer(struct hpvm_TimerSet *timers, char *label,
}
void hpvm_SwitchToTimer(struct hpvm_TimerSet *timers, enum hpvm_TimerID timer) {
+
+#ifdef HPVM_USE_OPENCL
+
// cerr << "Switch to timer: " << timer << flush << "\n";
/* Stop the currently running timer */
if (timers->current != hpvm_TimerID_NONE) {
@@ -849,10 +944,21 @@ void hpvm_SwitchToTimer(struct hpvm_TimerSet *timers, enum hpvm_TimerID timer) {
}
}
timers->current = timer;
+
+#else
+
+ openCLAbort();
+
+#endif
+
+
}
void hpvm_SwitchToSubTimer(struct hpvm_TimerSet *timers, char *label,
enum hpvm_TimerID category) {
+
+#ifdef HPVM_USE_OPENCL
+
struct hpvm_SubTimerList *subtimerlist =
timers->sub_timer_list[timers->current];
struct hpvm_SubTimer *curr =
@@ -1001,6 +1107,13 @@ void hpvm_SwitchToSubTimer(struct hpvm_TimerSet *timers, char *label,
}
timers->current = category;
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
void hpvm_PrintTimerSet(struct hpvm_TimerSet *timers) {
@@ -1069,6 +1182,9 @@ void hpvm_PrintTimerSet(struct hpvm_TimerSet *timers) {
}
void hpvm_DestroyTimerSet(struct hpvm_TimerSet *timers) {
+
+#ifdef HPVM_USE_OPENCL
+
/* clean up all of the async event markers */
struct hpvm_async_time_marker_list *event = timers->async_markers;
while (event != NULL) {
@@ -1106,6 +1222,13 @@ void hpvm_DestroyTimerSet(struct hpvm_TimerSet *timers) {
free(timers->sub_timer_list[i]);
}
}
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
/**************************** Pipeline API ************************************/
@@ -1304,10 +1427,13 @@ void llvm_hpvm_cpu_wait(void *graphID) {
DEBUG(cout << "\t... pthread Done!\n");
}
+
+#ifdef HPVM_USE_OPENCL
+
// Returns the platform name.
std::string getPlatformName(cl_platform_id pid) {
+
cl_int status;
-
size_t sz;
status = clGetPlatformInfo(pid, CL_PLATFORM_NAME, 0, NULL, &sz);
checkErr(status, CL_SUCCESS, "Query for platform name size failed");
@@ -1318,12 +1444,18 @@ std::string getPlatformName(cl_platform_id pid) {
const auto &tmp = std::string(name, name + sz);
delete[] name;
- return tmp;
+ return tmp;
}
+#endif
+
+
+#ifdef HPVM_USE_OPENCL
+
// Searches all platforms for the first platform whose name
// contains the search string (case-insensitive).
cl_platform_id findPlatform(const char *platform_name_search) {
+
cl_int status;
std::string search = platform_name_search;
@@ -1360,7 +1492,13 @@ cl_platform_id findPlatform(const char *platform_name_search) {
assert(false && "No matching platform found!");
}
+#endif
+
+
void *llvm_hpvm_ocl_initContext(enum hpvm::Target T) {
+
+#ifdef HPVM_USE_OPENCL
+
pthread_mutex_lock(&ocl_mtx);
DEBUG(std::string Target = T == hpvm::GPU_TARGET ? "GPU" : "SPIR");
DEBUG(cout << "Initializing Context for " << Target << " device\n");
@@ -1450,9 +1588,19 @@ void *llvm_hpvm_ocl_initContext(enum hpvm::Target T) {
pthread_mutex_unlock(&ocl_mtx);
return globalOCLContext;
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
void llvm_hpvm_ocl_clearContext(void *graphID) {
+
+#ifdef HPVM_USE_OPENCL
+
pthread_mutex_lock(&ocl_mtx);
DEBUG(cout << "Clear Context\n");
DFNodeContext_OCL *Context = (DFNodeContext_OCL *)graphID;
@@ -1464,9 +1612,19 @@ void llvm_hpvm_ocl_clearContext(void *graphID) {
cout << "Printing HPVM Timer: KernelTimer\n";
hpvm_PrintTimerSet(&kernel_timer);
pthread_mutex_unlock(&ocl_mtx);
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
void llvm_hpvm_ocl_argument_shared(void *graphID, int arg_index, size_t size) {
+
+#ifdef HPVM_USE_OPENCL
+
pthread_mutex_lock(&ocl_mtx);
DEBUG(cout << "Set Shared Memory Input:");
DEBUG(cout << "\tArgument Index = " << arg_index << ", Size = " << size
@@ -1477,10 +1635,20 @@ void llvm_hpvm_ocl_argument_shared(void *graphID, int arg_index, size_t size) {
cl_int errcode = clSetKernelArg(Context->clKernel, arg_index, size, NULL);
checkErr(errcode, CL_SUCCESS, "Failure to set shared memory argument");
pthread_mutex_unlock(&ocl_mtx);
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
void llvm_hpvm_ocl_argument_scalar(void *graphID, void *input, int arg_index,
size_t size) {
+
+#ifdef HPVM_USE_OPENCL
+
pthread_mutex_lock(&ocl_mtx);
DEBUG(cout << "Set Scalar Input:");
DEBUG(cout << "\tArgument Index = " << arg_index << ", Size = " << size
@@ -1491,10 +1659,20 @@ void llvm_hpvm_ocl_argument_scalar(void *graphID, void *input, int arg_index,
cl_int errcode = clSetKernelArg(Context->clKernel, arg_index, size, input);
checkErr(errcode, CL_SUCCESS, "Failure to set constant input argument");
pthread_mutex_unlock(&ocl_mtx);
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
void *llvm_hpvm_ocl_argument_ptr(void *graphID, void *input, int arg_index,
size_t size, bool isInput, bool isOutput) {
+
+#ifdef HPVM_USE_OPENCL
+
pthread_mutex_lock(&ocl_mtx);
DEBUG(cout << "Set Pointer Input:");
DEBUG(cout << "\tArgument Index = " << arg_index << ", Ptr = " << input
@@ -1518,9 +1696,19 @@ void *llvm_hpvm_ocl_argument_ptr(void *graphID, void *input, int arg_index,
DEBUG(cout << "\tDevicePtr = " << d_input << flush << "\n");
pthread_mutex_unlock(&ocl_mtx);
return d_input;
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
void *llvm_hpvm_ocl_output_ptr(void *graphID, int arg_index, size_t size) {
+
+#ifdef HPVM_USE_OPENCL
+
pthread_mutex_lock(&ocl_mtx);
DEBUG(cout << "Set device memory for Output Struct:");
DEBUG(cout << "\tArgument Index = " << arg_index << ", Size = " << size
@@ -1536,12 +1724,22 @@ void *llvm_hpvm_ocl_output_ptr(void *graphID, int arg_index, size_t size) {
DEBUG(cout << "\tDevicePtr = " << d_output << flush << "\n");
pthread_mutex_unlock(&ocl_mtx);
return d_output;
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
void llvm_hpvm_ocl_free(void *ptr) {}
void *llvm_hpvm_ocl_getOutput(void *graphID, void *h_output, void *d_output,
size_t size) {
+
+#ifdef HPVM_USE_OPENCL
+
pthread_mutex_lock(&ocl_mtx);
DEBUG(cout << "Get Output:\n");
DEBUG(cout << "\tHostPtr = " << h_output << ", DevicePtr = " << d_output
@@ -1555,11 +1753,21 @@ void *llvm_hpvm_ocl_getOutput(void *graphID, void *h_output, void *d_output,
checkErr(errcode, CL_SUCCESS, "[getOutput] Failure to read output");
pthread_mutex_unlock(&ocl_mtx);
return h_output;
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
void *llvm_hpvm_ocl_executeNode(void *graphID, unsigned workDim,
const size_t *localWorkSize,
const size_t *globalWorkSize) {
+
+#ifdef HPVM_USE_OPENCL
+
pthread_mutex_lock(&ocl_mtx);
size_t GlobalWG[3];
@@ -1610,6 +1818,13 @@ void *llvm_hpvm_ocl_executeNode(void *graphID, unsigned workDim,
pthread_mutex_unlock(&ocl_mtx);
return NULL;
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
//////////////////////////////////////////////////////////////////////////////
@@ -1656,6 +1871,9 @@ static char *LoadProgSource(const char *Filename, size_t *szFinalLength) {
}
void *llvm_hpvm_ocl_launch(const char *FileName, const char *KernelName) {
+
+#ifdef HPVM_USE_OPENCL
+
pthread_mutex_lock(&ocl_mtx);
DEBUG(cout << "Launch OCL Kernel\n");
// Initialize OpenCL
@@ -1717,14 +1935,31 @@ void *llvm_hpvm_ocl_launch(const char *FileName, const char *KernelName) {
pthread_mutex_unlock(&ocl_mtx);
return Context;
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
void llvm_hpvm_ocl_wait(void *graphID) {
+
+#ifdef HPVM_USE_OPENCL
+
pthread_mutex_lock(&ocl_mtx);
DEBUG(cout << "Wait\n");
DFNodeContext_OCL *Context = (DFNodeContext_OCL *)graphID;
clFinish(Context->clCommandQue);
pthread_mutex_unlock(&ocl_mtx);
+
+#else
+
+ openCLAbort();
+
+#endif
+
}
void llvm_hpvm_switchToTimer(void **timerSet, enum hpvm_TimerID timer) {
@@ -1749,3 +1984,5 @@ void *llvm_hpvm_initializeTimerSet() {
pthread_mutex_unlock(&ocl_mtx);
return TS;
}
+
+
--
GitLab