Remove flagcx_hetero.cc and hostGpuMemAlloc.h (#7)

flagcx/core/group.cc

@@ -14,6 +14,7 @@
 #include "assert.h"
 #include "net.h"
 #include "adaptor.h"
+#include "launch_kernel.h"
 
 __thread int flagcxGroupDepth = 0;
 __thread bool flagcxGroupJobAbortFlag = false;

flagcx/core/launch_kernel.cc

@@ -0,0 +1,7 @@
+#include "launch_kernel.h"
+
+void cpuAsyncLaunch(void *_args){
+    struct hostLaunchArgs *args = (struct hostLaunchArgs *) _args;
+    while(!args->stopLaunch);
+    __atomic_store_n(&args->retLaunch, 1, __ATOMIC_RELAXED);
+}

flagcx/core/launch_kernel.h

@@ -0,0 +1,27 @@
+#ifndef FLAGCX_LAUNCH_KERNEL_H_
+#define FLAGCX_LAUNCH_KERNEL_H_
+
+#include "topo.h"
+#include "debug.h"
+#include <stdlib.h>
+#include <getopt.h>
+#include <memory.h>
+#include <stdio.h>
+#include <math.h>
+#include <iostream>
+#include <iomanip>
+#include <unistd.h>
+#include <dlfcn.h>
+#include "adaptor.h"
+#include "utils.h"
+#include "param.h"
+
+struct hostLaunchArgs{
+    volatile bool stopLaunch;
+    volatile bool retLaunch;
+};
+
+void cpuAsyncLaunch(void *_args);
+
+#endif
+

flagcx/core/proxy.h

@@ -14,7 +14,7 @@
 #include <pthread.h>
 #include "info.h"
 #include "net.h"
-#include "hostGpuMemAlloc.h"
+#include "launch_kernel.h"
 
 enum flagcxProxyOpState { flagcxProxyOpNone, flagcxProxyOpReady, flagcxProxyOpProgress };
 

flagcx/core/topo.cc

@@ -12,7 +12,6 @@
 #include <sys/stat.h>
 #include <fcntl.h>
 #include "bootstrap.h"
-#include "hostGpuMemAlloc.h"
 
 #define BUSID_SIZE (sizeof("0000:00:00.0"))
 #define BUSID_REDUCED_SIZE (sizeof("0000:00"))

flagcx/core/topo.h

@@ -249,4 +249,14 @@ static int mirrorBits(int val, int pow2) {
   for (int b=1, mb=(pow2>>1); b<pow2; b<<=1, mb>>=1) if (val & b) mirror |= mb;
   return mirror;
 }
+
+
+#ifdef CREATE_DEVICE_TOPO_API
+#define DEVICE_TOPO_API_EXTERN
+#else
+#define DEVICE_TOPO_API_EXTERN extern
+#endif
+
+DEVICE_TOPO_API_EXTERN flagcxResult_t (*flagcxTopoGetLocalNet)(int gpu, char *name);
+
 #endif

flagcx/core/transport.cc

@@ -5,7 +5,6 @@
 #include "net.h"
 #include "topo.h"
 #include "adaptor.h"
-#include "hostGpuMemAlloc.h"
 #define ENABLE_TIMER 0
 #include "timer.h"
 

flagcx/service/load_devapi.cc

@@ -0,0 +1,45 @@
+#define CREATE_DEVICE_TOPO_API 
+#include "topo.h"
+#include "debug.h"
+#include <stdlib.h>
+#include <getopt.h>
+#include <memory.h>
+#include <stdio.h>
+#include <math.h>
+#include <iostream>
+#include <iomanip>
+#include <unistd.h>
+#include <dlfcn.h>
+#include "adaptor.h"
+#include "utils.h"
+#include "param.h"
+
+using namespace std;
+
+struct flagcxDeviceAdaptor devRunTimeApi;
+
+void *dlsymCheck(void *handle, const char *funcName){
+    void *funcPtr = dlsym(handle, funcName);
+    if(funcPtr == NULL) INFO(FLAGCX_INIT, "fail to load symbol %s", funcName);
+    return funcPtr;
+}
+
+#define LOADSYMBOL(handle,api) do{ \
+    api = (typeof(api)) dlsymCheck(handle, #api); \
+}while(0);
+
+flagcxResult_t loadDeviceSymbol(){
+    void *libHandle = dlopen("./libmylib.so", RTLD_LAZY);
+    if(libHandle == nullptr){
+        const char* useNet = flagcxGetEnv("FLAGCX_USENET");
+        if(useNet == NULL){
+            INFO(FLAGCX_INIT, "fail to open libmylib.so");
+            return flagcxRemoteError;
+        }
+        return flagcxSuccess;
+    }
+
+    LOADSYMBOL(libHandle, flagcxTopoGetLocalNet);
+    return flagcxSuccess;
+}
+