202108051031

examples/hdl4secnn/cnn/cnncell/cnncell.h

@@ -61,6 +61,8 @@ DEFINE_GUID(CLSID_CNN_BUF_W4, 0x9d2c00a4, 0x519f, 0x4b07, 0x82, 0x1e, 0x94, 0x34
 
 int cnnInit();
 
+#define CNNCELLDEBUG 0
+
 #endif /* _ASMLANGUAGE */
 
 #ifdef __cplusplus

examples/hdl4secnn/cnn/cnncell/cnncell_convolution.c

@@ -85,8 +85,6 @@ IDLIST
 	VID(lineindex),
 END_IDLIST
 
-#define CONVDEBUG 1
-
 MODULE_DECLARE(cnncell_convolution)
 	int input_w, input_h, input_c;
 	int output_w, output_h, output_c;
@@ -100,7 +98,7 @@ MODULE_DECLARE(cnncell_convolution)
 	float* line_out;
 	short *coeff;
 	short *bias;
-#if CONVDEBUG
+#if CNNCELLDEBUG
 	FILE* pDumpFile;
 #endif
 END_MODULE_DECLARE(cnncell_convolution)
@@ -211,7 +209,7 @@ DEFINE_FUNC(cnncell_convolution_gen_line) {
 			/* bias */
 			//if (layer_info->bias_term)
 			data_acc += pobj->bias[cc] * pobj->bias_scale;
-#if CONVDEBUG
+#if CNNCELLDEBUG
 			fprintf(pobj->pDumpFile, "%d, %d, %d, %f\n", vget(writeline), ww, cc, data_acc);
 #endif
 			/* active function */
@@ -223,7 +221,7 @@ DEFINE_FUNC(cnncell_convolution_gen_line) {
 			pobj->line_out[ww * pobj->output_c + cc] = data_acc;
 		}
 	}
-#if CONVDEBUG
+#if CNNCELLDEBUG
 	fflush(pobj->pDumpFile);
 #endif
 }
@@ -398,7 +396,7 @@ MODULE_INIT(cnncell_convolution)
 	pobj->linesize = (pobj->input_w + 2 * pobj->pw) * pobj->input_c;
 	pobj->line_in = (float*)malloc(pobj->kh * pobj->linesize * sizeof(float));
 	pobj->line_out = (float*)malloc(pobj->output_w * pobj->output_c * sizeof(float));
-#if CONVDEBUG
+#if CNNCELLDEBUG
 	{
 		char filename[80];
 		sprintf(filename, "d:/gitwork/debugfile/cnncell_%s.txt", pobj->data.name);

examples/hdl4secnn/cnn/cnncell/cnncell_innerproduct.c

@@ -81,8 +81,6 @@ IDLIST
 	VID(index),
 END_IDLIST
 
-#define INPD_DEBUG 1
-
 MODULE_DECLARE(cnncell_innerproduct)
 	int input_w, input_h, input_c;
 	int output_w, output_h, output_c;
@@ -92,7 +90,7 @@ MODULE_DECLARE(cnncell_innerproduct)
 	short* bias;
 	float* inbuf;
 	float* outbuf;
-#ifdef INPD_DEBUG
+#if CNNCELLDEBUG
 	FILE* pDumpFile;
 #endif
 END_MODULE_DECLARE(cnncell_innerproduct)
@@ -167,11 +165,11 @@ DEFINE_FUNC(cnncell_innerproduct_calc) {
 		pobj->outbuf[oc] *= pobj->coeff_scale;
 		/* bias */
 		pobj->outbuf[oc] += bias[oc] * pobj->bias_scale;
-#ifdef INPD_DEBUG
+#if CNNCELLDEBUG
 		fprintf(pobj->pDumpFile, "%d, %f\n", oc, pobj->outbuf[oc]);
 #endif
 	}
-#ifdef INPD_DEBUG
+#if CNNCELLDEBUG
 	fflush(pobj->pDumpFile);
 #endif
 	return 0;
@@ -308,7 +306,7 @@ MODULE_INIT(cnncell_innerproduct)
 	pobj->bias = (short*)malloc(pobj->output_c * pobj->output_h * pobj->output_w * sizeof(short));
 	pobj->inbuf = (float*)malloc(pobj->input_c * pobj->input_h * pobj->input_w * sizeof(float));
 	pobj->outbuf = (float*)malloc(pobj->output_c * pobj->output_h * pobj->output_w * sizeof(float));
-#ifdef INPD_DEBUG
+#if CNNCELLDEBUG
 	{
 		char filename[80];
 		sprintf(filename, "d:/gitwork/debugfile/cnncell_%s.txt", pobj->data.name);

examples/hdl4secnn/cnn/cnncell/cnncell_lrn.c

@@ -65,8 +65,6 @@ IDLIST
 	VID(state),
 END_IDLIST
 
-#define LRNDEBUG 1
-
 MODULE_DECLARE(cnncell_lrn)
 	int w, h, c;
 	int region;
@@ -74,7 +72,7 @@ MODULE_DECLARE(cnncell_lrn)
 	float k, alpha, beta;
 	float* buf;
 	float* square;
-#if LRNDEBUG
+#if CNNCELLDEBUG
 	FILE* pDumpFile;
 #endif
 END_MODULE_DECLARE(cnncell_lrn)
@@ -147,7 +145,7 @@ DEFINE_FUNC(cnncell_lrn_gen_line) {
 			sum += square[pobj->local_size] - square[0];
 			square++;
 		}
-#if LRNDEBUG
+#if CNNCELLDEBUG
 		for (j = 0; j < pobj->c; j++) {
 			fprintf(pobj->pDumpFile, "%d, %f\n", j, pobj->buf[j]);
 		}
@@ -211,7 +209,7 @@ MODULE_INIT(cnncell_lrn)
 	pobj->alpha = (float)MODULE_PARAM_REAL(5);
 	pobj->beta = (float)MODULE_PARAM_REAL(6);
 	pobj->region = (int)MODULE_PARAM(7);
-#if LRNDEBUG
+#if CNNCELLDEBUG
 	{
 		char filename[80];
 		sprintf(filename, "d:/gitwork/debugfile/cnncell_%s.txt", pobj->data.name);

examples/hdl4secnn/cnn/cnncell/cnncell_pooling.c

@@ -69,8 +69,6 @@ IDLIST
 	VID(lineindex),
 END_IDLIST
 
-#define POOLINGDEBUG 1
-
 MODULE_DECLARE(cnncell_pooling)
 	int input_w, input_h, input_c;
 	int output_w, output_h, output_c;
@@ -80,7 +78,7 @@ MODULE_DECLARE(cnncell_pooling)
 	int linesize;
 	float* line_in;
 	float* line_out;
-#if POOLINGDEBUG
+#if CNNCELLDEBUG
 	FILE* pDumpFile;
 #endif
 
@@ -151,7 +149,7 @@ DEFINE_FUNC(cnncell_pooling_gen_line) {
 						}
 					}
 				}
-#if POOLINGDEBUG
+#if CNNCELLDEBUG
 				fprintf(pobj->pDumpFile, "%d, %d, %d, %f\n", vget(writeline), ww, cc, data_max);
 #endif
 #if 0
@@ -179,7 +177,7 @@ DEFINE_FUNC(cnncell_pooling_gen_line) {
 					}
 				}
 				data_total /= (pobj->kh * pobj->kw);
-#if POOLINGDEBUG
+#if CNNCELLDEBUG
 				fprintf(pobj->pDumpFile, "%d, %d, %d, %f\n", vget(writeline), ww, cc, data_total);
 #endif
 				/* active function */
@@ -194,7 +192,7 @@ DEFINE_FUNC(cnncell_pooling_gen_line) {
 			}
 		}
 	}
-#if POOLINGDEBUG
+#if CNNCELLDEBUG
 	fflush(pobj->pDumpFile);
 #endif
 }
@@ -322,7 +320,7 @@ MODULE_INIT(cnncell_pooling)
 	pobj->line_in = (float*)malloc(pobj->kh * pobj->linesize * sizeof(float));
 	memset(pobj->line_in, 0, pobj->kh* pobj->linesize * sizeof(float));
 	pobj->line_out = (float*)malloc(pobj->output_w * pobj->output_c * sizeof(float));
-#if POOLINGDEBUG
+#if CNNCELLDEBUG
 	{
 		char filename[80];
 		sprintf(filename, "d:/gitwork/debugfile/cnncell_%s.txt", pobj->data.name);

examples/hdl4secnn/cnn/cnncell/cnncell_softmax.c

@@ -65,13 +65,11 @@ IDLIST
 	VID(index),
 END_IDLIST
 
-#define SOFTMAX_DEBUG 1
-
 MODULE_DECLARE(cnncell_softmax)
 	int w, h, c;
 	float* inbuf;
 	float* outbuf;
-#ifdef SOFTMAX_DEBUG
+#if CNNCELLDEBUG
 	FILE* pDumpFile;
 #endif
 END_MODULE_DECLARE(cnncell_softmax)
@@ -104,7 +102,7 @@ DEFINE_FUNC(cnncell_softmax_calc) {
 	for (i = 0; i < pobj->c * pobj->w * pobj->h; i++) {
 		pobj->outbuf[i] /= total;
 	}
-#ifdef SOFTMAX_DEBUG
+#if CNNCELLDEBUG
 	for (i = 0; i < pobj->c * pobj->w * pobj->h; i++) {
 		fprintf(pobj->pDumpFile, "%d, %f\n", i, pobj->outbuf[i]);
 	}
@@ -192,7 +190,7 @@ MODULE_INIT(cnncell_softmax)
 	pobj->c = (int)MODULE_PARAM(2);
 	pobj->inbuf = (float*)malloc(pobj->c * pobj->h * pobj->w * sizeof(float));
 	pobj->outbuf = (float*)malloc(pobj->c * pobj->h * pobj->w * sizeof(float));
-#ifdef SOFTMAX_DEBUG
+#if CNNCELLDEBUG
 	{
 		char filename[80];
 		sprintf(filename, "d:/gitwork/debugfile/cnncell_%s.txt", pobj->data.name);

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();
 
@@ -66,6 +66,10 @@ int main(int argc, char* argv[])
 	int i;
 	int width;
 	int count, unitcount;
+
+	time_t starttime;
+	time_t lasttime;
+
 	IHDL4SEWaveOutput** vcdfile;
 	cnnInit();
 	sim = hdl4sesimCreateSimulator();
@@ -76,7 +80,7 @@ int main(int argc, char* argv[])
 	objectPrintInfo(stdout, fprintf);
 #if VCDOUTPUT
 	vcdfile = hdl4sesimCreateVCDFile("googlenet.vcd"); 
-
+	
 	objectCall2(vcdfile, AddSignal, "/top/net/cnn_buf_167", "wRead");
 	objectCall2(vcdfile, AddSignal, "/top/net/cnn_buf_167", "wDataValid");
 	objectCall2(vcdfile, AddSignal, "/top/net/cnn_buf_167", "bReadData");
@@ -128,6 +132,7 @@ int main(int argc, char* argv[])
 	objectCall0(vcdfile, StartRecord);
 #endif
 	clocks = 0;
+	lasttime = starttime = time(NULL);
 	do {
 		objectCall0(sim, ClkTick);
 #if VCDOUTPUT
@@ -137,8 +142,14 @@ int main(int argc, char* argv[])
 		if (clocks == 2)
 			objectCall1(sim, SetReset, 1);
 		clocks++;
-		if ((clocks & 0xFFF) == 0)
-			printf("clocks: %lld\n", clocks);
+		if ((clocks & 0xFFF) == 0) {
+			time_t thistime = time(NULL);
+			printf("clocks: %lld, TSPD=%lfcps, LSPD=%lfcps\n", 
+				clocks, 
+				(double)clocks / (double)(thistime - starttime), 
+				4096.0 / (double)(thistime - lasttime));
+			lasttime = time(NULL);
+		}
 	} while (running);
 #if VCDOUTPUT
 	objectCall0(vcdfile, StopRecord);

hdl4sesim/src/hdl4sesim.c

@@ -224,7 +224,7 @@ 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(16)
+//#pragma omp parallel for num_threads(4)
 	for (i = 0; i < pobj->multithreadmodules.itemcount; i++) {
 		hdl4se_module_ClkTick(&((IHDL4SEModuleVar*)(pobj->multithreadmodules.array[i]))->data);
 	}
@@ -243,7 +243,7 @@ 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(16)
+//#pragma omp parallel for num_threads(4)
 	for (i = 0; i < pobj->multithreadmodules.itemcount; i++) {
 		hdl4se_module_Setup(&((IHDL4SEModuleVar*)(pobj->multithreadmodules.array[i]))->data);
 	}