JNN 6.3.9

AI_Engine_6.3.7 Example/.classpath → JNN 6.3.8/.classpath

@@ -23,6 +23,5 @@
 	<classpathentry kind="lib" path="commons-math3-3.6.1/commons-math3-3.6.1-tests.jar"/>
 	<classpathentry kind="lib" path="commons-math3-3.6.1/commons-math3-3.6.1-tools.jar"/>
 	<classpathentry kind="lib" path="commons-math3-3.6.1/commons-math3-3.6.1.jar"/>
-	<classpathentry kind="lib" path="AI Engine 6.3.7.jar"/>
 	<classpathentry kind="output" path="bin"/>
 </classpath>

AI_Engine_6.3.7 Example/.project → JNN 6.3.8/.project

@@ -1,6 +1,6 @@
 <?xml version="1.0" encoding="UTF-8"?>
 <projectDescription>
-	<name>AI_Engine_6.3.7 Example</name>
+	<name>JNN 6.3.8</name>
 	<comment></comment>
 	<projects>
 	</projects>

JNN 6.3.8/LayerKernel.ptx

@@ -0,0 +1,251 @@
+//
+// Generated by NVIDIA NVVM Compiler
+//
+// Compiler Build ID: CL-22781540
+// Cuda compilation tools, release 9.0, V9.0.176
+// Based on LLVM 3.4svn
+//
+
+.version 6.0
+.target sm_30
+.address_size 64
+
+	// .globl	layer
+
+.visible .entry layer(
+	.param .u32 layer_param_0,
+	.param .u32 layer_param_1,
+	.param .u64 layer_param_2,
+	.param .u64 layer_param_3,
+	.param .u64 layer_param_4,
+	.param .u64 layer_param_5
+)
+{
+	.reg .pred 	%p<18>;
+	.reg .f32 	%f<116>;
+	.reg .b32 	%r<37>;
+	.reg .f64 	%fd<5>;
+	.reg .b64 	%rd<35>;
+
+
+	ld.param.u32 	%r17, [layer_param_0];
+	ld.param.u32 	%r18, [layer_param_1];
+	ld.param.u64 	%rd11, [layer_param_2];
+	ld.param.u64 	%rd8, [layer_param_3];
+	ld.param.u64 	%rd9, [layer_param_4];
+	ld.param.u64 	%rd10, [layer_param_5];
+	cvta.to.global.u64 	%rd1, %rd8;
+	cvta.to.global.u64 	%rd2, %rd11;
+	mov.u32 	%r1, %ntid.x;
+	mov.u32 	%r2, %ctaid.x;
+	mov.u32 	%r3, %tid.x;
+	mad.lo.s32 	%r4, %r1, %r2, %r3;
+	setp.gt.s32	%p1, %r4, %r18;
+	@%p1 bra 	BB0_19;
+
+	mov.f32 	%f19, 0f00000000;
+	setp.lt.s32	%p2, %r17, 1;
+	mov.f32 	%f113, %f19;
+	@%p2 bra 	BB0_11;
+
+	and.b32  	%r22, %r17, 3;
+	mov.f32 	%f113, 0f00000000;
+	mov.u32 	%r34, 0;
+	setp.eq.s32	%p3, %r22, 0;
+	@%p3 bra 	BB0_8;
+
+	setp.eq.s32	%p4, %r22, 1;
+	@%p4 bra 	BB0_7;
+
+	setp.eq.s32	%p5, %r22, 2;
+	@%p5 bra 	BB0_6;
+
+	ld.global.f32 	%f23, [%rd2];
+	mul.wide.s32 	%rd12, %r4, 4;
+	add.s64 	%rd13, %rd1, %rd12;
+	ld.global.f32 	%f24, [%rd13];
+	fma.rn.f32 	%f113, %f23, %f24, 0f00000000;
+	mov.u32 	%r34, 1;
+
+BB0_6:
+	mul.wide.u32 	%rd14, %r34, 4;
+	add.s64 	%rd15, %rd2, %rd14;
+	neg.s32 	%r24, %r34;
+	and.b32  	%r25, %r24, %r18;
+	add.s32 	%r26, %r25, %r4;
+	mul.wide.s32 	%rd16, %r26, 4;
+	add.s64 	%rd17, %rd1, %rd16;
+	ld.global.f32 	%f25, [%rd17];
+	ld.global.f32 	%f26, [%rd15];
+	fma.rn.f32 	%f113, %f26, %f25, %f113;
+	add.s32 	%r34, %r34, 1;
+
+BB0_7:
+	mul.wide.s32 	%rd18, %r34, 4;
+	add.s64 	%rd19, %rd2, %rd18;
+	mad.lo.s32 	%r27, %r34, %r18, %r4;
+	mul.wide.s32 	%rd20, %r27, 4;
+	add.s64 	%rd21, %rd1, %rd20;
+	ld.global.f32 	%f27, [%rd21];
+	ld.global.f32 	%f28, [%rd19];
+	fma.rn.f32 	%f113, %f28, %f27, %f113;
+	add.s32 	%r34, %r34, 1;
+
+BB0_8:
+	setp.lt.u32	%p6, %r17, 4;
+	@%p6 bra 	BB0_11;
+
+	shl.b32 	%r10, %r18, 2;
+	mul.wide.s32 	%rd22, %r34, 4;
+	add.s64 	%rd34, %rd2, %rd22;
+	mad.lo.s32 	%r35, %r34, %r18, %r4;
+
+BB0_10:
+	mul.wide.s32 	%rd23, %r35, 4;
+	add.s64 	%rd24, %rd1, %rd23;
+	ld.global.f32 	%f29, [%rd24];
+	ld.global.f32 	%f30, [%rd34];
+	fma.rn.f32 	%f31, %f30, %f29, %f113;
+	cvt.s64.s32	%rd25, %r10;
+	add.s64 	%rd26, %rd24, %rd25;
+	ld.global.f32 	%f32, [%rd26];
+	ld.global.f32 	%f33, [%rd34+4];
+	fma.rn.f32 	%f34, %f33, %f32, %f31;
+	add.s64 	%rd27, %rd26, %rd25;
+	ld.global.f32 	%f35, [%rd27];
+	ld.global.f32 	%f36, [%rd34+8];
+	fma.rn.f32 	%f37, %f36, %f35, %f34;
+	add.s64 	%rd28, %rd27, %rd25;
+	ld.global.f32 	%f38, [%rd28];
+	ld.global.f32 	%f39, [%rd34+12];
+	fma.rn.f32 	%f113, %f39, %f38, %f37;
+	add.s64 	%rd34, %rd34, 16;
+	add.s32 	%r35, %r35, %r10;
+	add.s32 	%r34, %r34, 4;
+	setp.lt.s32	%p7, %r34, %r17;
+	@%p7 bra 	BB0_10;
+
+BB0_11:
+	cvta.to.global.u64 	%rd7, %rd10;
+	cvta.to.global.u64 	%rd29, %rd9;
+	mul.wide.s32 	%rd30, %r4, 4;
+	add.s64 	%rd31, %rd29, %rd30;
+	ld.global.f32 	%f42, [%rd31];
+	add.f32 	%f10, %f113, %f42;
+	neg.f32 	%f43, %f10;
+	mov.f32 	%f41, 0f4016FC2A;
+	// inline asm
+	rcp.approx.ftz.f32 %f40,%f41;
+	// inline asm
+	mul.f32 	%f44, %f40, 0f3F37E150;
+	mul.f32 	%f45, %f44, %f44;
+	mov.f32 	%f46, 0f3C4CAF63;
+	mov.f32 	%f47, 0f3B18F0FE;
+	fma.rn.f32 	%f48, %f47, %f45, %f46;
+	mov.f32 	%f49, 0f3DAAAABD;
+	fma.rn.f32 	%f50, %f48, %f45, %f49;
+	mul.rn.f32 	%f51, %f50, %f45;
+	mul.rn.f32 	%f52, %f51, %f44;
+	mov.f32 	%f53, 0f3EB7E150;
+	sub.f32 	%f54, %f53, %f44;
+	mul.f32 	%f55, %f40, 0fBF37E150;
+	add.f32 	%f56, %f54, %f54;
+	fma.rn.f32 	%f57, %f55, %f53, %f56;
+	mul.rn.f32 	%f58, %f40, %f57;
+	add.f32 	%f59, %f44, %f52;
+	sub.f32 	%f60, %f44, %f59;
+	add.f32 	%f61, %f52, %f60;
+	add.f32 	%f62, %f58, %f61;
+	add.f32 	%f63, %f59, %f62;
+	sub.f32 	%f64, %f59, %f63;
+	add.f32 	%f65, %f62, %f64;
+	mov.f32 	%f66, 0f3F317200;
+	mov.f32 	%f67, 0f3F800000;
+	mul.rn.f32 	%f68, %f67, %f66;
+	add.f32 	%f69, %f68, %f63;
+	sub.f32 	%f70, %f68, %f69;
+	add.f32 	%f71, %f63, %f70;
+	add.f32 	%f72, %f65, %f71;
+	mov.f32 	%f73, 0f35BFBE8E;
+	mul.rn.f32 	%f74, %f67, %f73;
+	add.f32 	%f75, %f74, %f72;
+	add.f32 	%f76, %f69, %f75;
+	sub.f32 	%f77, %f69, %f76;
+	add.f32 	%f78, %f75, %f77;
+	abs.f32 	%f11, %f43;
+	setp.gt.f32	%p8, %f11, 0f77F684DF;
+	mul.f32 	%f79, %f10, 0fB9000000;
+	selp.f32	%f80, %f79, %f43, %p8;
+	mul.rn.f32 	%f81, %f80, %f76;
+	neg.f32 	%f82, %f81;
+	fma.rn.f32 	%f83, %f80, %f76, %f82;
+	fma.rn.f32 	%f84, %f80, %f78, %f83;
+	fma.rn.f32 	%f86, %f19, %f76, %f84;
+	add.rn.f32 	%f87, %f81, %f86;
+	neg.f32 	%f88, %f87;
+	add.rn.f32 	%f89, %f81, %f88;
+	add.rn.f32 	%f90, %f89, %f86;
+	mov.b32 	 %r29, %f87;
+	setp.eq.s32	%p9, %r29, 1118925336;
+	add.s32 	%r30, %r29, -1;
+	mov.b32 	 %f91, %r30;
+	add.f32 	%f92, %f90, 0f37000000;
+	selp.f32	%f93, %f91, %f87, %p9;
+	selp.f32	%f12, %f92, %f90, %p9;
+	mul.f32 	%f94, %f93, 0f3FB8AA3B;
+	cvt.rzi.f32.f32	%f95, %f94;
+	mov.f32 	%f96, 0fBF317200;
+	fma.rn.f32 	%f97, %f95, %f96, %f93;
+	mov.f32 	%f98, 0fB5BFBE8E;
+	fma.rn.f32 	%f99, %f95, %f98, %f97;
+	mul.f32 	%f100, %f99, 0f3FB8AA3B;
+	ex2.approx.ftz.f32 	%f101, %f100;
+	add.f32 	%f102, %f95, 0f00000000;
+	ex2.approx.f32 	%f103, %f102;
+	mul.f32 	%f104, %f101, %f103;
+	setp.lt.f32	%p10, %f93, 0fC2D20000;
+	selp.f32	%f105, 0f00000000, %f104, %p10;
+	setp.gt.f32	%p11, %f93, 0f42D20000;
+	selp.f32	%f115, 0f7F800000, %f105, %p11;
+	setp.eq.f32	%p12, %f115, 0f7F800000;
+	@%p12 bra 	BB0_13;
+
+	fma.rn.f32 	%f115, %f115, %f12, %f115;
+
+BB0_13:
+	add.f32 	%f106, %f11, 0f402DF854;
+	mov.b32 	 %r31, %f106;
+	setp.lt.s32	%p13, %r31, 2139095040;
+	@%p13 bra 	BB0_18;
+
+	setp.gtu.f32	%p14, %f11, 0f7F800000;
+	@%p14 bra 	BB0_17;
+	bra.uni 	BB0_15;
+
+BB0_17:
+	mov.f32 	%f107, 0f402DF854;
+	sub.f32 	%f115, %f107, %f10;
+	bra.uni 	BB0_18;
+
+BB0_15:
+	setp.neu.f32	%p15, %f11, 0f7F800000;
+	@%p15 bra 	BB0_18;
+
+	setp.gt.f32	%p16, %f10, 0f80000000;
+	selp.f32	%f115, 0f00000000, 0f7F800000, %p16;
+
+BB0_18:
+	cvt.f64.f32	%fd1, %f115;
+	add.f64 	%fd2, %fd1, 0d3FF0000000000000;
+	rcp.rn.f64 	%fd3, %fd2;
+	setp.eq.f32	%p17, %f10, 0f80000000;
+	selp.f64	%fd4, 0d3FE0000000000000, %fd3, %p17;
+	cvt.rn.f32.f64	%f108, %fd4;
+	add.s64 	%rd33, %rd7, %rd30;
+	st.global.f32 	[%rd33], %f108;
+
+BB0_19:
+	ret;
+}
+
+

JNN 6.3.8/src/CPUClasses/CPUAgent.java

@@ -0,0 +1,124 @@
+package CPUClasses;
+
+import java.util.Random;
+
+import org.apache.commons.math3.linear.MatrixUtils;
+import org.apache.commons.math3.linear.RealMatrix;
+
+import JNN.JNN;
+
+public class CPUAgent {
+
+	private RealMatrix Inputs;
+	private RealMatrix[] baises;
+	private RealMatrix[] weights;
+	private RealMatrix Outputs;
+	private float points[];
+	private Thread thread;
+	private Runnable runNN;
+
+	private double sigmoid(double x) {
+		return (1 / (1 + Math.pow(Math.E, -x)));
+	}
+
+	public CPUAgent(int Generationsteps, int[] layout) {
+		points = new float[Generationsteps];
+		this.Inputs = MatrixUtils.createRealMatrix(1, layout[0]);
+		baises = new RealMatrix[layout.length - 1];
+		weights = new RealMatrix[layout.length - 1];
+		int last_layer = layout[0];
+		Random random = new Random();
+		for (int i = 0; i < weights.length; i++) {
+			baises[i] = MatrixUtils.createRealMatrix(1, layout[i + 1]);
+			weights[i] = MatrixUtils.createRealMatrix(last_layer, layout[i + 1]);
+
+			for (int j = 0; j < layout[i + 1]; j++) {
+				baises[i].setEntry(0, j, random.nextDouble() * 2 - 1);
+				for (int k = 0; k < last_layer; k++) {
+					weights[i].setEntry(k, j, random.nextDouble() * 2 - 1);
+				}
+			}
+			last_layer = layout[i + 1];
+
+		}
+		runNN = new Runnable() {
+
+			@Override
+			public void run() {
+				RealMatrix last_layer = Inputs;
+				for (int i = 0; i < weights.length; i++) {
+					last_layer = last_layer.multiply(weights[i]);
+					last_layer = last_layer.add(baises[i]);
+					for (int j = 0; j < last_layer.getColumnDimension(); j++) {
+						last_layer.setEntry(0, j, sigmoid(last_layer.getEntry(0, j)));
+					}
+				}
+				Outputs = last_layer;
+
+			}
+		};
+	}
+
+	public void setInput(int Input, double Value) {
+		Inputs.setEntry(0, Input, Value);
+	}
+
+	public void runNN() {
+		RealMatrix last_layer = Inputs;
+		for (int i = 0; i < weights.length; i++) {
+			last_layer = last_layer.multiply(weights[i]);
+			last_layer = last_layer.add(baises[i]);
+			for (int j = 0; j < last_layer.getColumnDimension(); j++) {
+				last_layer.setEntry(0, j, sigmoid(last_layer.getEntry(0, j)));
+			}
+		}
+		Outputs = last_layer;
+	}
+
+	public void runNNMT() {
+		thread = new Thread(runNN);
+		thread.start();
+	}
+
+	public void joinThread() {
+		try {
+			thread.join();
+		} catch (InterruptedException e) {
+			// TODO Auto-generated catch block
+			e.printStackTrace();
+		}
+	}
+
+	public double getOutput(int Output) {
+		return Outputs.getEntry(0, Output);
+	}
+
+	public void setScore(float Score) {
+		points[JNN.generationStep] = Score;
+	}
+
+	float getAvarageScore() {
+		float temp = 0;
+		for (int i = 0; i < points.length; i++) {
+			temp += points[i];
+		}
+		return temp / points.length;
+	}
+
+	RealMatrix[] getWeights() {
+		return weights;
+	}
+
+	RealMatrix[] getBaises() {
+		return baises;
+	}
+
+	void setWeights(RealMatrix[] weights) {
+		this.weights = weights;
+	}
+
+	void setBaises(RealMatrix[] baises) {
+		this.baises = baises;
+	}
+
+}