202108111109

examples/hdl4secnn/googlenet/verilog/main.c

@@ -49,7 +49,7 @@ unsigned long long clocks = 0;
 
 static int running = 1;
 
-#define VCDOUTPUT 1
+#define VCDOUTPUT 0
 
 int cnnInit();
 

hdl4secell/include/hdl4secell.h

@@ -145,6 +145,7 @@ typedef struct sModuleFunction {
 	void* pobj;
 	MODULE_FUNC func;
 	MODULE_FUNC cur_func;
+	int callit;
 	PointerArray variables; /* 影响的variables */
 }ModuleFunction;
 
@@ -239,7 +240,16 @@ int hdl4seCheckVariableArray(PointerArray* array);
 ModuleFunction* functionCreateFunc(MODULE_FUNC func, void * pobj);
 ModuleFunction* functionCreateGetValue(ModuleVariable * var);
 int functionDestroy(ModuleFunction* func);
-int functionCallAndSetNone(ModuleFunction* func);
+
+//int functionCallAndSetNone(ModuleFunction* func);
+#define functionCallAndSetNone(obj) \
+do { \
+	ModuleFunction * func = (ModuleFunction *)obj; \
+	int callit = LOCKED_SWAP(func->callit, 0); \
+	if (callit) { \
+		func->func(func->pobj, func->var); \
+	} \
+} while (0)
 
 int hdl4se_split_string(const char* string, char** substr, int substrcount, int gap);
 unsigned int hdl4seGetParamUint32(const char* param);

hdl4secell/src/hdl4se_function.c

@@ -59,6 +59,7 @@ ModuleFunction* functionCreateFunc(MODULE_FUNC func, void* pobj)
 	function->func = function->cur_func = func;
 	function->pobj = pobj;
 	function->var = NULL;
+	function->callit = 1;
 	pointerarrayInit(&function->variables);
 	return function;
 }
@@ -77,12 +78,14 @@ static int gen_none(void* p, void* v)
 	return 0;
 }
 
+/*
 int functionCallAndSetNone(ModuleFunction* func)
 {
 	MODULE_FUNC genfunc;
 	genfunc = LOCKED_SWAP_POINTER(func->cur_func, gen_none);
 	return genfunc(func->pobj, func->var);
 }
+*/
 
 int functionDestroy(ModuleFunction* func)
 {

hdl4secell/src/hdl4se_module.c

@@ -110,16 +110,16 @@ do { \
 #define UpdateVariable(v) \
 do { \
 	if (v->type != VTYPE_REG) { \
-		threadlockLock(v->lock); \
-		if (v->updatefunc != 0) { \
-			if (v->updatefunc == 1) { \
+		int updatefunc; \
+		updatefunc = LOCKED_SWAP(v->updatefunc, 0); \
+		if (updatefunc != 0) { \
+			if (updatefunc == 1) { \
 				hdl4seUpdateVariableGenFunc(v); \
 			} \
 			else { \
 				hdl4seUpdateVariableGetValue(v); \
 			} \
 		} \
-		threadlockUnlock(v->lock); \
 	} \
 } while (0)
 #endif
@@ -681,7 +681,8 @@ int hdl4se_module_Setup(sGeneralModule* pobj)
 	
 	for (i = 0; i < pobj->funcs.itemcount; i++) {
 		ModuleFunction* function = (ModuleFunction *)pobj->funcs.array[i];
-		function->cur_func = function->func;
+		//function->cur_func = function->func;
+		function->callit = 1;
 	}
 	
 	return 0;

hdl4sesim/src/hdl4sesim.c

@@ -231,7 +231,7 @@ static int hdl4sesim_hdl4se_simulator_SetReset(HOBJECT object, int reset)
 	return 0;
 }
 
-#define THREADCOUNT 9
+#define THREADCOUNT 2
 
 static int hdl4sesim_hdl4se_simulator_ClkTick(HOBJECT object)
 {
@@ -297,6 +297,15 @@ softmax_168 complete
 0.0356  n04505470 typewriter keyboard
 clocks: 1884160, TSPD=1359.422799cps, LSPD=1170.285714cps
 
+clocks: 1802240, TSPD=4681.142857cps, LSPD=4096.000000cps
+softmax_168 complete
+0.4901  n03085013 computer keyboard, keypad
+0.1591  n04264628 space bar
+0.0543  n04074963 remote control, remote
+0.0438  n03832673 notebook, notebook computer
+0.0356  n04505470 typewriter keyboard
+clocks: 1818624, TSPD=4687.175258cps, LSPD=5461.333333cps
+
 threadcount=3
 clocks : 1867776, TSPD = 1552.598504cps, LSPD = 2048.000000cps
 softmax_168 complete
@@ -317,6 +326,14 @@ softmax_168 complete
 0.0356  n04505470 typewriter keyboard
 clocks: 1884160, TSPD=1312.089136cps, LSPD=1260.307692cps
 
+clocks: 1802240, TSPD=4767.830688cps, LSPD=4096.000000cps
+softmax_168 complete
+0.4901  n03085013 computer keyboard, keypad
+0.1591  n04264628 space bar
+0.0543  n04074963 remote control, remote
+0.0439  n03832673 notebook, notebook computer
+0.0356  n04505470 typewriter keyboard
+clocks: 1818624, TSPD=4760.795812cps, LSPD=4096.000000cps
 
 threadcount=2
 clocks: 1867776, TSPD=1562.992469cps, LSPD=1365.333333cps
@@ -338,6 +355,25 @@ softmax_168 complete
 0.0356  n04505470 typewriter keyboard
 clocks: 1818624, TSPD=1752.046243cps, LSPD=1820.444444cps
 
+usebigint == 0
+clocks: 1802240, TSPD=3746.860707cps, LSPD=4096.000000cps
+softmax_168 complete
+0.4901  n03085013 computer keyboard, keypad
+0.1591  n04264628 space bar
+0.0543  n04074963 remote control, remote
+0.0439  n03832673 notebook, notebook computer
+0.0356  n04505470 typewriter keyboard
+clocks: 1818624, TSPD=3749.740206cps, LSPD=4096.000000cps
+
+clocks: 1802240, TSPD=4114.703196cps, LSPD=4096.000000cps
+softmax_168 complete
+0.4901  n03085013 computer keyboard, keypad
+0.1591  n04264628 space bar
+0.0543  n04074963 remote control, remote
+0.0439  n03832673 notebook, notebook computer
+0.0356  n04505470 typewriter keyboard
+clocks: 1818624, TSPD=4114.533937cps, LSPD=4096.000000cps
+
 threadcount=1
 clocks: 1867776, TSPD=993.497872cps, LSPD=963.764706cps
 softmax_168 complete
@@ -348,6 +384,25 @@ softmax_168 complete
 0.0356  n04505470 typewriter keyboard
 clocks: 1884160, TSPD=993.755274cps, LSPD=1024.000000cps
 
+usebitint == 0
+clocks: 1802240, TSPD=2549.137199cps, LSPD=2730.666667cps
+softmax_168 complete
+0.4901  n03085013 computer keyboard, keypad
+0.1591  n04264628 space bar
+0.0543  n04074963 remote control, remote
+0.0439  n03832673 notebook, notebook computer
+0.0356  n04505470 typewriter keyboard
+clocks: 1818624, TSPD=2547.092437cps, LSPD=2340.571429cps
+
+clocks: 1802240, TSPD=2802.861586cps, LSPD=2730.666667cps
+softmax_168 complete
+0.4901  n03085013 computer keyboard, keypad
+0.1591  n04264628 space bar
+0.0543  n04074963 remote control, remote
+0.0439  n03832673 notebook, notebook computer
+0.0356  n04505470 typewriter keyboard
+clocks: 1818624, TSPD=2802.194145cps, LSPD=2730.666667cps
+
 i9 10900 32G
 threadcount=32
 clocks: 1867776, TSPD=8768.901408cps, LSPD=8192.000000cps

hdl4seutils/include/threadlock.h

@@ -61,7 +61,7 @@ void  mt_free(void*);
 #define LOCKED_INCREMENT(p) InterlockedIncrement(&p)
 #define LOCKED_DECREMENT(p) InterlockedDecrement(&p)
 #define LOCKED_SWAP_POINTER(p, v) InterlockedExchangePointer(&p, v)
-
+#define LOCKED_SWAP(p, v) InterlockedExchange(&p, v)
 #define STRDUP _strdup
 #endif
 
@@ -69,7 +69,7 @@ void  mt_free(void*);
 #define LOCKED_INCREMENT(p) __sync_fetch_and_add(&p, 1);
 #define LOCKED_DECREMENT(p) __sync_fetch_and_sub(&p, 1);
 #define LOCKED_SWAP_POINTER(p, v) __sync_val_compare_and_swap(&p, p, v)
-
+#define LOCKED_SWAP(p, v) __sync_val_compare_and_swap(&p, p, v)
 #define STRDUP strdup
 #endif
 
@@ -78,7 +78,14 @@ void  mt_free(void*);
 #define LOCKED_DECREMENT(p) p--
 static void* locked_swap_pointer(void** p, void* v)
 {
-	void* ret = v;
+	void* ret = *p;
+	*p = v;
+	return ret;
+}
+#define LOCKED_SWAP_POINTER(p, v) locked_swap_pointer(&p, v)
+static long locked_swap(long * p, long v)
+{
+	long ret = *p;
 	*p = v;
 	return ret;
 }