202108060624

examples/hdl4secnn/cnn/cnncell/cnncell_convolution.c

@@ -98,6 +98,7 @@ MODULE_DECLARE(cnncell_convolution)
 	float* line_out;
 	short *coeff;
 	short *bias;
+	int reset;
 #if CNNCELLDEBUG
 	FILE* pDumpFile;
 #endif
@@ -113,7 +114,7 @@ enum convstate {
 };
 
 DEFINE_FUNC(cnncell_convolution_gen_output_coeff) {
-	if (vget(nwReset) == 0) {
+	if (pobj->reset == 0) {
 		vput(wCoeffRead, 0);
 		vput(bCoeffReadAddr, 0);
 	}
@@ -154,7 +155,7 @@ DEFINE_FUNC(cnncell_convolution_gen_output_coeff) {
 
 DEFINE_FUNC(cnncell_convolution_gen_output_read) {
 	vput(wDataRead,0);
-	if (vget(nwReset) != 0) {
+	if (pobj->reset != 0) {
 		int state = vget(state);
 		if (state == CONV_START) { 
 			if (vget(wDataReadValid)) {
@@ -229,14 +230,20 @@ END_DEFINE_FUNC
 
 DEFINE_FUNC(cnncell_convolution_clktick) {
 	//wDataRead = 0;
-	if (vget(nwReset) == 0) {
-		vput(state, READ_COEFF);
-		vput(index, 0);
-		vput(readline, 0);
-		vput(writeline, 0);
+	if (pobj->reset == 0) {
+		if (vget(nwReset) == 0) {
+			vput(state, READ_COEFF);
+			vput(index, 0);
+			vput(readline, 0);
+			vput(writeline, 0);
+		}
+		else {
+			pobj->reset = 1;
+		}
 	}
-	else {
+	if (pobj->reset) {
 		int state = vget(state);
+		pobj->reset = 1;
 		if (state == READ_COEFF) {
 			int coeffvalid;
 			coeffvalid = vget(wCoeffReadValid);
@@ -390,6 +397,7 @@ MODULE_INIT(cnncell_convolution)
 	pobj->pw = (int)MODULE_PARAM(10);
 	pobj->ph = (int)MODULE_PARAM(11);
 	pobj->actfunc= (int)MODULE_PARAM(12);
+	pobj->reset = 0;
 	pobj->coeffcount = pobj->output_c * pobj->kw * pobj->kh * pobj->input_c;
 	pobj->coeff = (short *)malloc(pobj->coeffcount * sizeof(short));
 	pobj->bias = (short *)malloc(pobj->output_c * sizeof(short));

hdl4sesim/src/hdl4sesim.c

@@ -47,6 +47,8 @@
 #include "hdl4secell.h"
 #include "threadlock.h"
 
+#include "windows.h"
+
 #define IMPLEMENT_GUID
 #include "hdl4sesim.h"
 #undef IMPLEMENT_GUID
@@ -216,6 +218,8 @@ static int hdl4sesim_hdl4se_simulator_SetReset(HOBJECT object, int reset)
 	return 0;
 }
 
+#define THREADCOUNT 4
+
 static int hdl4sesim_hdl4se_simulator_ClkTick(HOBJECT object)
 {
 	sHDL4SESim* pobj;
@@ -226,14 +230,11 @@ static int hdl4sesim_hdl4se_simulator_ClkTick(HOBJECT object)
 	for (i = 0; i < pobj->singlethreadmodules.itemcount; i++) {
 		hdl4se_module_ClkTick(&((IHDL4SEModuleVar*)(pobj->singlethreadmodules.array[i]))->data);
 	}
-#pragma omp parallel for num_threads(2)
-	//for (i = 0; i < pobj->multithreadmodules.itemcount; i++) {
+#pragma omp parallel for num_threads(THREADCOUNT)
 	for (i = 0; i < pobj->multithreadmodules.itemcount; i++) {
-	//	int j = i / 16 + (i % 16) * 16;
-	//	if (j < pobj->multithreadmodules.itemcount)
-			hdl4se_module_ClkTick(&((IHDL4SEModuleVar*)(pobj->multithreadmodules.array[i]))->data);
+		hdl4se_module_ClkTick(&((IHDL4SEModuleVar*)(pobj->multithreadmodules.array[i]))->data);
 	}
-	return 0;
+	return 0; 
 }
 
 static int hdl4sesim_hdl4se_simulator_Setup(HOBJECT object)
@@ -248,7 +249,7 @@ static int hdl4sesim_hdl4se_simulator_Setup(HOBJECT object)
 	for (i = 0; i < pobj->singlethreadmodules.itemcount; i++) {
 		hdl4se_module_Setup(&((IHDL4SEModuleVar*)(pobj->singlethreadmodules.array[i]))->data);
 	}
-#pragma omp parallel for num_threads(2)
+#pragma omp parallel for num_threads(THREADCOUNT)
 	for (i = 0; i < pobj->multithreadmodules.itemcount; i++) {
 		hdl4se_module_Setup(&((IHDL4SEModuleVar*)(pobj->multithreadmodules.array[i]))->data);
 	}