This is ArmV82_Keras_CNN_Inference, a Arm machine learning prototyping software project, optimized for Arm Cortex v8 CPU arch.

1. Based on Keras mnist training database
2. Will leverage Arm neon instruction set, http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0489e/CJAJIIGG.html
3. Will apply dot product
4. Executing CA73x2 + CA53x4 FVP
5. Enable DS5 profiling tools: streamline

Step:

1. Training model
2. Translate to C code
3. DS5 compiler
4. Simualtion: A) Debug script: Load training model into DDR (json file or binary) Load data into FVP DDR space Some data pre-operation B) Start simulation C) Display result

DDR:

Test memeory map:

0x0			Parameter
	0x00  ~ 0040: Conv-1
	0x680 ~ 0700: Conv-2
	0xCF00 ~ D100: FullConn-1
	0xD100 ~ D128: FullConn-2
	
0x50000 mnist test image
	0x50000 : image 0
	0x51000 : image 1
	0x52000 : image 2
	0x53000 : image 3
	0x54000 : image 4
	0x55000 : image 5

0x60~0xEF	workspace
	0x60000 ~ 0x77000
	0x78000 ~ 0x8F000
	0x90000 ~ 0xA7000
	0xA8000 ~ 0xBF000
	0xC0000 ~ 0xD7000
	0xD8000 ~ 0xEF000

Workspace memory layout map: (a75_a55)

1. 0x80370000: Test data, 0x80371000: Test data, 0x80372000: Test data, 0x80373000: Test data, 0x80374000: Test data, 0x80375000: Test data,

   work space

2. 0x80350000: Normalized, unsign int / 255 --> float

3. 0x80351000: Conv-1, [28,28,1] --> [24,24,1] x16 [5,5] filter 16 biase Relu

4. 0x8035B000: MaxPool-1 [24,24,1] x16 --> [12,12,1] x16 [2,2] filter

5. 0x8035E000: Conv-2 [12,12,1] x16 --> [8,8,1] x32 [5,5] filter 32 biase Relu

6. 0x80360000: MaxPool-2 [8,8,1] x32 --> [4,4,1] x32 [2,2] filter

7. 0x80361000: FullConn-1 [512] --> [128]

8. 0x80361300: FullConn-2 [128] --> [10]

API: PMU

ToDo: NEON Cifar-10