3.5 Release

CHANGELOG

@@ -1,3 +1,22 @@
+3.5
+
+Highlights of this release:
+
+- NNTool
+	- Added new quantization scheme with scaling factors (tflite quantization-like)
+	- Support import tflite (1 or 2) quantized graphs (uint8)
+	- New adjust algorithm to support more general reshape/transpose layers
+	- Added Autotiler layer for input image formatting (HWC uint8 -> CHW int8/int16)
+	- Support multi-inputs network execution
+	- Support for new tensor dumping routines in AutoTiler
+	- Support for generation of model using SQ8 AutoTiler generators
+	- Bug fixes
+
+- Autotiler
+	- New SQ8 kernel set supporting scaled quantization
+	- Tensor dumping from any memory area
+	- Several bug fixes
+
 3.4
 
 Highlights of this release:

applications/CannyEdgeDetection/CannyDetect.c

@@ -667,11 +667,9 @@ void canny_edge_detector()
 	sprintf(imageName, "../../../%s", Imagefile);
 	ImageIn_L2 = (unsigned char *) pi_l2_malloc( COL*LINE*sizeof(unsigned char));
 
-    unsigned int Wi, Hi;
-
-    if ( (ReadImageFromFile(imageName, &Wi, &Hi, ImageIn_L2, LINE*COL*sizeof(unsigned char))==0) || (Wi!=COL) || (Hi!=LINE))
+    if (ReadImageFromFile(imageName, COL,LINE, 1, ImageIn_L2, LINE*COL*sizeof(unsigned char), 0, 0))
     {
-        printf("Failed to load image %s or dimension mismatch Expects [%dx%d], Got [%dx%d]\n", imageName, COL, LINE, Wi, Hi);
+        printf("Failed to load image %s\n", imageName);
         pmsis_exit(-1);
     }
 

applications/CannyEdgeDetection/Makefile

@@ -26,10 +26,6 @@ APP_LDFLAGS +=  -flto -Wl,--gc-sections
 # in a clean way.
 APP_CFLAGS += -Wall -Werror -Wno-maybe-uninitialized -Wno-unused-but-set-variable -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wundef
 
-# Should be removed once openocd is the default bridge
-export GAP_USE_OPENOCD=1
-io=host
-
 clean::
 	rm -rf img_OUT.ppm
 

applications/FaceDetection/FaceDetGenerator.c

@@ -92,7 +92,7 @@ void GenerateResize(char *Name, int Wi, int Hi, int Wo, int Ho)
 		KernelIterSpace(1, IterTiledSpace(KER_ITER_TILE0)),
 		TILE_HOR,
 		CArgs(2, TCArg("unsigned char *", "In"), TCArg("unsigned char *", "Out")),
-		Calls(1, Call("KerResizeBilinear", LOC_INNER_LOOP,
+		Calls(1, Call("KerResizeBilinear", LOC_LOOP,
 			Bindings(8, K_Arg("In", KER_ARG_TILE),
 				        K_Arg("In", KER_ARG_W),
 				        K_Arg("In", KER_ARG_H),
@@ -123,13 +123,13 @@ void GenerateIntegralImage(char *Name,
 			TCArg("unsigned int *  __restrict__", "IntegralImage")
 		),
 		Calls(2,
-			Call("KerIntegralImagePrime", LOC_INNER_LOOP_PROLOG,
+			Call("KerIntegralImagePrime", LOC_LOOP_PROLOG,
 				Bindings(2,
 					K_Arg("KerBuffer",KER_ARG),
 					K_Arg("KerIn", KER_ARG_TILE_W)
 				)
 			),
-			Call("KerIntegralImageProcess", LOC_INNER_LOOP,
+			Call("KerIntegralImageProcess", LOC_LOOP,
 				Bindings(5,
 					K_Arg("KerIn", KER_ARG_TILE),
 					K_Arg("KerIn", KER_ARG_TILE_W),
@@ -162,13 +162,13 @@ void GenerateSquaredIntegralImage(char *Name,
 			TCArg("unsigned int *  __restrict__", "IntegralImage")
 		),
 		Calls(2,
-			Call("KerIntegralImagePrime", LOC_INNER_LOOP_PROLOG,
+			Call("KerIntegralImagePrime", LOC_LOOP_PROLOG,
 				Bindings(2,
 					K_Arg("KerBuffer",KER_ARG),
 					K_Arg("KerIn", KER_ARG_TILE_W)
 				)
 			),
-			Call("KerSquaredIntegralImageProcess", LOC_INNER_LOOP,
+			Call("KerSquaredIntegralImageProcess", LOC_LOOP,
 				Bindings(5,
 					K_Arg("KerIn", KER_ARG_TILE),
 					K_Arg("KerIn", KER_ARG_TILE_W),
@@ -208,7 +208,7 @@ void GenerateCascadeClassifier(char *Name,
 		),
 		Calls(1,
 
-			Call("KerEvaluateCascade", LOC_INNER_LOOP,
+			Call("KerEvaluateCascade", LOC_LOOP,
 				Bindings(8,
 					K_Arg("KerII", KER_ARG_TILE),
 					K_Arg("KerIISQ", KER_ARG_TILE),

applications/FaceDetection/Makefile

@@ -28,7 +28,12 @@ override config_args += --config-opt=camera/image-stream=$(CURDIR)/imgTest0.pgm
 
 APP_SRCS += $(FACE_DET_SRCS)
 APP_INC  += $(TILER_INC)
-APP_CFLAGS += -O3 -g -D__PMSIS__ -DUSE_CAMERA -DUSE_DISPLAY -DNB_FRAMES=$(NB_FRAMES)
+APP_CFLAGS += -O3 -g -D__PMSIS__ -DUSE_CAMERA -DNB_FRAMES=$(NB_FRAMES)
+
+display ?= 1
+ifeq ($(display), 1)
+APP_CFLAGS += -DUSE_DISPLAY
+endif
 
 BOARD_NAME ?= gapoc_a
 PMSIS_OS ?= freertos

applications/FaceDetection/main.c

@@ -51,9 +51,11 @@
 #define LCD_HEIGHT   240
 
 static unsigned char *imgBuff0;
+#if defined(USE_DISPLAY)
 static struct pi_device ili;
 static pi_buffer_t buffer;
 static pi_buffer_t buffer_out;
+#endif  /* USE_DISPLAY */
 static struct pi_device cam;
 
 L2_MEM unsigned char *ImageOut;
@@ -70,11 +72,9 @@ ArgCluster_T ClusterCall;
 void setCursor(struct pi_device *device,signed short x, signed short y);
 void writeFillRect(struct pi_device *device, unsigned short x, unsigned short y, unsigned short w, unsigned short h, unsigned short color);
 void writeText(struct pi_device *device,char* str,int fontsize);
-#endif  /* USE_DISPLAY */
 
 static int open_display(struct pi_device *device)
 {
-#if defined(USE_DISPLAY)
     struct pi_ili9341_conf ili_conf;
 
     pi_ili9341_conf_init(&ili_conf);
@@ -85,9 +85,9 @@ static int open_display(struct pi_device *device)
     {
         return -1;
     }
-#endif
     return 0;
 }
+#endif  /* USE_DISPLAY */
 
 #if defined(USE_CAMERA)
 #if defined(HIMAX)
@@ -166,15 +166,17 @@ void test_facedetection(void)
     {
         printf("Failed to allocate Memory for one or both Integral Images (%d bytes)\n", ImgSize*sizeof(unsigned int));
         pmsis_exit(-3);
-  }
+    }
     printf("malloc done\n");
 
+    #if defined(USE_DISPLAY)
     if (open_display(&ili))
     {
         printf("Failed to open display\n");
         pmsis_exit(-4);
     }
     printf("display done\n");
+    #endif  /* USE_DISPLAY */
 
     if (open_camera(&cam))
     {
@@ -183,6 +185,7 @@ void test_facedetection(void)
     }
     printf("Camera open success\n");
 
+    #if defined(USE_DISPLAY)
     #if defined(HIMAX)
     buffer.data = imgBuff0+CAM_WIDTH*2+2;
     buffer.stride = 4;
@@ -195,14 +198,13 @@ void test_facedetection(void)
     pi_buffer_init(&buffer, PI_BUFFER_TYPE_L2, imgBuff0);
     #endif  /* HIMAX */
 
-    #if defined(USE_DISPLAY)
     buffer_out.data = ImageOut;
     buffer_out.stride = 0;
     pi_buffer_init(&buffer_out, PI_BUFFER_TYPE_L2, ImageOut);
     pi_buffer_set_stride(&buffer_out, 0);
-    #endif /* USE_DISPLAY */
 
     pi_buffer_set_format(&buffer, CAM_WIDTH, CAM_HEIGHT, 1, PI_BUFFER_FORMAT_GRAY);
+    #endif /* USE_DISPLAY */
 
     ClusterCall.ImageIn              = imgBuff0;
     ClusterCall.Win                  = W;

applications/FaceDetection/testset.cfg

@@ -6,10 +6,10 @@ test = Test(
   name = 'face_detection',
   commands = [
     Shell('clean', 'make clean'),
-    Shell('build', 'make all NB_FRAMES=5'),
+    Shell('build', 'make all NB_FRAMES=5 display=0'),
     Shell('run',   'make run')
   ],
   timeout=1000000,
 )
-  
+
 c['tests'] = [ test ]

applications/jpeg_encoder/Makefile

@@ -1,3 +1,6 @@
+# Only Pulp-Os for now.
+PMSIS_OS=pulpos
+
 APP = test
 APP_SRCS += test.c ImgIO.c
 APP_CFLAGS += -O3 -g -Werror

configs/common.sh

@@ -14,6 +14,9 @@ export DEP_DIRS=$INSTALL_DIR
 export RULES_DIR=$GAP_SDK_HOME/tools/rules
 
 export NNTOOL_DIR=$GAP_SDK_HOME/tools/nntool
+export NNTOOL_PATH=$GAP_SDK_HOME/tools/nntool
+export NNTOOL_KERNELS_PATH=$NNTOOL_DIR/autotiler/kernels
+export NNTOOL_GENERATOR_PATH=$NNTOOL_DIR/autotiler/generators
 export PATH="$NNTOOL_DIR":$PATH
 
 # PulpOS 2
@@ -55,8 +58,8 @@ export TILER_EMU_INC=$TILER_PATH/include
 export TILER_GENERATOR_PATH=$TILER_PATH/generators
 export TILER_CNN_KERNEL_PATH=$TILER_PATH/generators/CNN
 export TILER_CNN_GENERATOR_PATH=$TILER_PATH/generators/CNN
-export TILER_CNN_KERNEL_PATH_SQ8=$TILER_PATH/CNN_Libraries_SQ8
-export TILER_CNN_GENERATOR_PATH_SQ8=$TILER_PATH/CNN_Generators_SQ8
+export TILER_CNN_KERNEL_PATH_SQ8=$TILER_PATH/generators/CNN
+export TILER_CNN_GENERATOR_PATH_SQ8=$TILER_PATH/generators/CNN
 
 
 # OpenOCD

examples/autotiler/Cifar10/Makefile

@@ -8,10 +8,26 @@ RM=rm -f
 CNN_AT_PATH = $(TILER_GENERATOR_PATH)/CNN
 
 APP = Cifar10
+
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_BiasReLULinear_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DP_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DW_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DW_DP_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Pooling_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_MatAlgebra.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_SoftMax.c
+MODEL_LIB_POW2 += $(NNTOOL_KERNELS_PATH)/norm_transpose.c
+MODEL_LIB_INCLUDE_POW2 = -I$(NNTOOL_KERNELS_PATH) -I$(TILER_CNN_KERNEL_PATH) -I$(NNTOOL_KERNELS_PATH)
+MODEL_GEN_POW2 += $(TILER_CNN_GENERATOR_PATH)/CNN_Generator_Util.c
+MODEL_GEN_POW2 += $(TILER_CNN_GENERATOR_PATH)/CNN_Generators.c
+MODEL_GEN_POW2 += $(NNTOOL_GENERATOR_PATH)/nntool_extra_generators.c
+MODEL_GEN_INCLUDE_POW2 = -I$(TILER_CNN_GENERATOR_PATH) -I$(NNTOOL_GENERATOR_PATH)
+
+
 APP_SRCS += Cifar10.c Cifar10Kernels.c \
-            $(CNN_AT_PATH)/CNN_BiasReLULinear_BasicKernels.c \
-            $(CNN_AT_PATH)/CNN_Conv_BasicKernels.c \
-            $(CNN_AT_PATH)/CNN_Pooling_BasicKernels.c
+			$(MODEL_LIB_POW2)
+
 APP_INC += $(TILER_INC) $(CNN_AT_PATH)
 
 COEF_FLAG = -DCOEF_L2
@@ -57,7 +73,7 @@ all:: model
 
 # Build the code generator
 GenTile:
-	gcc -o GenCifar10 $(COEF_FLAG) -I$(TILER_INC) -I$(CNN_AT_PATH) Cifar10Model.c $(CNN_AT_PATH)/CNN_Generators.c $(TILER_LIB)
+	gcc -o GenCifar10 $(COEF_FLAG) -I$(TILER_INC) -I$(CNN_AT_PATH) Cifar10Model.c $(MODEL_GEN_POW2) $(TILER_LIB)
 
 # Run the code generator
 Cifar10Kernels.c: GenTile

examples/autotiler/IntegralImage/Makefile

@@ -14,16 +14,12 @@ MODEL_GEN_C = $(addsuffix .c, $(MODEL_GEN))
 MODEL_GEN_CLEAN = $(MODEL_GEN_C) $(addsuffix .h, $(MODEL_GEN))
 RM=rm -f
 
-APP_SRCS   += main.c $(GAP_LIB_PATH)/img_io/ImgIO.c $(MODEL_GEN_C) $(GEN_KER_PATH)/IntegralImgBasicKernels.c
+APP_SRCS   += main.c $(MODEL_GEN_C) $(GEN_KER_PATH)/IntegralImgBasicKernels.c $(GAP_LIB_PATH)/img_io/ImgIO.c
 APP_INC     = $(GAP_LIB_PATH)/include
 
 APP_CFLAGS += -O3 -mno-memcpy -fno-tree-loop-distribute-patterns $(JENKINS_FLAGS)
 APP_CFLAGS += -I$(TILER_EMU_INC) -I$(TILER_INC) -I$(GEN_KER_PATH)
 
-#moved to new openocd bridge
-export GAP_USE_OPENOCD=1
-io=host
-
 all:: model
 
 # Build the code generator

examples/autotiler/IntegralImage/main.c

@@ -97,7 +97,7 @@ void integral_image(int argc, char *argv[])
 
 	#ifndef NO_BRIDGE
 	//Reading Image from Hyperflash
-	if ((ReadImageFromFile(ImageName, &Wi, &Hi, ImageIn, W*H*sizeof(unsigned char))==0) || (Wi!=W) || (Hi!=H)) {
+	if (ReadImageFromFile(ImageName, W, H, 1, ImageIn, W*H*sizeof(unsigned char), 0, 0)) {
 		printf("Failed to load image %s or dimension mismatch Expects [%dx%d], Got [%dx%d]\n", ImageName, W, H, Wi, Hi);
 		pmsis_exit(-6);
 	}

examples/autotiler/Mnist/Makefile

@@ -23,10 +23,23 @@ endif
 
 CNN_AT_PATH = $(TILER_GENERATOR_PATH)/CNN
 
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_BiasReLULinear_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DP_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DW_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DW_DP_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Pooling_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_MatAlgebra.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_SoftMax.c
+MODEL_LIB_POW2 += $(NNTOOL_KERNELS_PATH)/norm_transpose.c
+MODEL_LIB_INCLUDE_POW2 = -I$(NNTOOL_KERNELS_PATH) -I$(TILER_CNN_KERNEL_PATH) -I$(NNTOOL_KERNELS_PATH)
+MODEL_GEN_POW2 += $(TILER_CNN_GENERATOR_PATH)/CNN_Generator_Util.c
+MODEL_GEN_POW2 += $(TILER_CNN_GENERATOR_PATH)/CNN_Generators.c
+MODEL_GEN_POW2 += $(NNTOOL_GENERATOR_PATH)/nntool_extra_generators.c
+MODEL_GEN_INCLUDE_POW2 = -I$(TILER_CNN_GENERATOR_PATH) -I$(NNTOOL_GENERATOR_PATH)
+
 APP_SRCS += Mnist.c MnistKernels.c \
-            $(CNN_AT_PATH)/CNN_BiasReLULinear_BasicKernels.c \
-            $(CNN_AT_PATH)/CNN_Conv_BasicKernels.c \
-            $(CNN_AT_PATH)/CNN_Pooling_BasicKernels.c \
+            $(MODEL_LIB_POW2) \
             ImgIO.c
 
 APP_INC += $(TILER_INC) $(CNN_AT_PATH)
@@ -46,7 +59,7 @@ all:: model
 
 # Build the code generator
 GenTile:
-	gcc -o GenMnist -I$(TILER_INC) -I$(CNN_AT_PATH) MnistModel.c $(CNN_AT_PATH)/CNN_Generators.c $(TILER_LIB)
+	gcc -o GenMnist -I$(TILER_INC) -I$(CNN_AT_PATH) MnistModel.c $(MODEL_GEN_POW2) $(TILER_LIB)
 
 # Run the code generator
 MnistKernels.c: GenTile

examples/autotiler/MnistGraph/Makefile

@@ -24,17 +24,29 @@ io=host
 
 CNN_AT_PATH = $(TILER_GENERATOR_PATH)/CNN
 
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_BiasReLULinear_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DP_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DW_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Conv_DW_DP_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_Pooling_BasicKernels.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_MatAlgebra.c
+MODEL_LIB_POW2 += $(TILER_CNN_KERNEL_PATH)/CNN_SoftMax.c
+MODEL_LIB_POW2 += $(NNTOOL_KERNELS_PATH)/norm_transpose.c
+MODEL_LIB_INCLUDE_POW2 = -I$(NNTOOL_KERNELS_PATH) -I$(TILER_CNN_KERNEL_PATH) -I$(NNTOOL_KERNELS_PATH)
+MODEL_GEN_POW2 += $(TILER_CNN_GENERATOR_PATH)/CNN_Generator_Util.c
+MODEL_GEN_POW2 += $(TILER_CNN_GENERATOR_PATH)/CNN_Generators.c
+MODEL_GEN_POW2 += $(NNTOOL_GENERATOR_PATH)/nntool_extra_generators.c
+MODEL_GEN_INCLUDE_POW2 = -I$(TILER_CNN_GENERATOR_PATH) -I$(NNTOOL_GENERATOR_PATH)
+
+
 APP_SRCS += Mnist.c MnistKernels.c \
-            $(CNN_AT_PATH)/CNN_BiasReLULinear_BasicKernels.c \
-            $(CNN_AT_PATH)/CNN_Conv_BasicKernels.c \
-            $(CNN_AT_PATH)/CNN_Conv_DP_BasicKernels.c \
-            $(CNN_AT_PATH)/CNN_Pooling_BasicKernels.c \
-            $(CNN_AT_PATH)/CNN_SoftMax.c \
+            $(MODEL_LIB_POW2) \
             ImgIO.c
 
 APP_INC += $(TILER_INC) $(CNN_AT_PATH)
 
-APP_CFLAGS += -w -O3 -mno-memcpy -fno-tree-loop-distribute-patterns -fdata-sections -ffunction-sections
+APP_CFLAGS += -w -g3 -O3 -mno-memcpy -fno-tree-loop-distribute-patterns -fdata-sections -ffunction-sections
 APP_CFLAGS += -Wno-maybe-uninitialized -Wno-unused-but-set-variable 
 LDFLAGS    +=  -flto -Wl,--gc-sections
 
@@ -48,7 +60,7 @@ all:: model
 
 # Build the code generator
 GenTile:
-	gcc -o GenMnist -I$(TILER_INC) -I$(CNN_AT_PATH) MnistModel.c $(CNN_AT_PATH)/CNN_Generators.c $(TILER_LIB)
+	gcc -o GenMnist -I$(TILER_INC) -I$(CNN_AT_PATH) MnistModel.c $(MODEL_GEN_POW2) $(TILER_LIB)
 
 # Run the code generator
 MnistKernels.c: GenTile

examples/native/freeRTOS/periph/timer/test_timer.c

@@ -1,8 +1,6 @@
 /* PMSIS includes */
 #include "pmsis.h"
-#if defined(PMSIS_DRIVERS)
-#include "pmsis_driver/pmsis_it.h"
-#endif  /* PMSIS_DRIVERS */
+#include "pmsis/implem/drivers/pmsis_it.h"
 
 #define NB_ITER          ( 5 )