package shared;

import ml.Cell;

/** 
 * Protects normal classes from seeing SIMD in case it doesn't compile or is absent.
 * ...in theory.
 * @author Brian Bushnell
 * @date Sep 12, 2013
 *
 */
public final class Vector {
	
	public static void main(String[] args) {
		int[] array=new int[19999];
		for(int i=0; i<array.length; i++) {
			array[i]=(byte)(i&127);
		}
		Timer t=new Timer();
		
		int loops=80000;
		
		long sum=0, sum2=0, max=0;
		for(int outer=0; outer<4; outer++) {
			t.start();
			sum=0; sum2=0;
			Shared.SIMD=false;
			for(int i=0; i<loops; i++) {
				sum=sum(array);
				sum2+=sum;
			}
			System.err.println(sum2);
			t.stop("Scalar: ");

			t.start();
			sum=0; sum2=0;
			Shared.SIMD=true;
			for(int i=0; i<loops; i++) {
				sum=sum(array);
				sum2+=sum;
			}
			System.err.println(sum2);
			t.stop("Vector: ");
		}
		
		for(int outer=0; outer<4; outer++) {
			t.start();
			max=0; sum=0; sum2=0;
			Shared.SIMD=false;
			for(int i=0; i<loops; i++) {
				max=max(array);
				sum2+=max;
			}
			System.err.println(max+", "+sum2);
			t.stop("Scalar: ");

			t.start();
			max=0; sum=0; sum2=0;
			Shared.SIMD=true;
			for(int i=0; i<loops; i++) {
				max=max(array);
				sum2+=max;
			}
			System.err.println(max+", "+sum2);
			t.stop("Vector: ");
		}
	}

	/** 
	 * Returns "c+=a[i]*b[i]" where a and b are equal-length arrays.
	 * @param a A vector to multiply.
	 * @param b A vector to multiply.
	 * @return Sum of products of vector elements.
	 */
	public static final float fma(final float[] a, final float[] b){
		assert(a.length==b.length);
		if(Shared.SIMD && a.length>=MINLEN32) {return SIMD.fma(a, b);}
		float c=0;
		for(int i=0; i<a.length; i++) {c+=a[i]*b[i];}
		return c;
	}

	/** 
	 * Returns "c+=a[i]*b[bSet[i]]".
	 * @param a A vector to multiply.
	 * @param b A vector to multiply.
	 * @param bSet Subset of B's indices.
	 * @param blockSize bSet should be in sets of consecutive indices of this length,
	 * for example, blockSize=8 would allow AVX256 vector operations.
	 * @return Sum of products of vector elements.
	 */
	public static final float fma(final float[] a, final float[] b, final int[] bSet, 
			final int blockSize, boolean allowSimd){
		assert(a.length==bSet.length);
		if(Shared.SIMD && a.length>=MINLEN32 && a.length==b.length) {return SIMD.fma(a, b);}
		if(Shared.SIMD && a.length>=MINLEN32 && allowSimd && ((blockSize&7)==0)) {//This ensures length-8 blocks
			return SIMD.fmaSparse(a, b, bSet);
		}
		float c=0;
		for(int i=0; i<a.length; i++) {c+=a[i]*b[bSet[i]];}
		return c;
	}
	
	public static final void feedForward(Cell[] layer, float[] valuesIn){
//		assert(layer.length==valuesIn.length);
		if(Shared.SIMD && valuesIn.length>=MINLEN32) {
			SIMD.feedForward(layer, valuesIn);
			return;
		}
		
		for(int cnum=0; cnum<layer.length; cnum++) {
			Cell c=layer[cnum];
			float[] weights=c.weights;
			float sum=c.bias;
			assert(valuesIn.length==weights.length) : valuesIn.length+", "+weights.length;
			sum+=Vector.fma(valuesIn, weights);
			c.sum=sum;
			final float v=(float)c.activation(sum);
			c.setValue(v);
		}
	}
	
	public static final void feedForwardDense(Cell[] layer, float[] valuesIn){
//		assert(layer.length==valuesIn.length);
		if(false && Shared.SIMD && valuesIn.length>=MINLEN32) {//Discovered anomaly here; very different results
			SIMD.feedForward(layer, valuesIn);
			return;
		}
		
		for(int cnum=0; cnum<layer.length; cnum++) {
			Cell c=layer[cnum];
			float[] weights=c.weights;
			float sum=c.bias;
			assert(valuesIn.length==weights.length) : valuesIn.length+", "+weights.length;
			sum+=Vector.fma(valuesIn, weights);
			c.sum=sum;
			final float v=(float)c.activation(sum);
			c.setValue(v);
		}
	}
	
	public static void backPropFma(Cell[] layer, float[] eOverNetNext, float[][] weightsOutLnum) {
		if(Shared.SIMD && eOverNetNext.length>=MINLEN32) {
			SIMD.backPropFma(layer, eOverNetNext, weightsOutLnum);
			return;
		}
		for(int i=0; i<layer.length; i++){
			Cell cell=layer[i];
			cell.eTotalOverOut=Vector.fma(weightsOutLnum[i], eOverNetNext);
		}
	}
	
	/** 
	 * Performs "a+=incr" where a and incr are equal-length arrays.
	 * @param a A vector to increment.
	 * @param incr Increment amount.
	 */
	public static final void add(final float[] a, final float[] incr){
		assert(a.length==incr.length);
		if(Shared.SIMD && a.length>=MINLEN32) {SIMD.add(a, incr); return;}
		for(int i=0; i<a.length; i++) {a[i]+=incr[i];}
	}
	
	/** 
	 * Performs "a[i]+=b[i]*mult" where a and b are equal-length arrays.
	 * @param a A vector to increment.
	 * @param b Increment amount.
	 * @param mult Increment multiplier.
	 */
	public static final void addProduct(final float[] a, final float[] b, final float mult){
		assert(a.length==b.length);
		if(Shared.SIMD && a.length>=MINLEN32) {SIMD.addProduct(a, b, mult); return;}
		for(int i=0; i<a.length; i++) {a[i]+=b[i]*mult;}
	}
	
	/** 
	 * Performs "a[i]+=b[bSet[i]]*mult".
	 * @param a A vector to increment.
	 * @param b Increment amount.
	 * @param bSet Subset of B's indices.
	 * @param mult Increment multiplier.
	 */
	public static final void addProduct(final float[] a, final float[] b, int[] bSet, final float mult, int blockSize){
		assert(a.length==bSet.length);
		if(Shared.SIMD && a.length>=MINLEN32 && a.length==b.length) {SIMD.addProduct(a, b, mult); return;}
		if(Shared.SIMD && a.length>=MINLEN32 && SIMD_MULT_SPARSE && ((blockSize&7)==0)) {SIMD.addProductSparse(a, b, bSet, mult); return;}
		for(int i=0; i<a.length; i++) {a[i]+=b[bSet[i]]*mult;}
	}
	
	public static void copy(float[] dest, float[] source) {
//		assert(a.length==b.length);
		if(SIMDCOPY && Shared.SIMD && dest.length>=MINLEN32) {SIMD.copy(dest, source); return;}
		for(int i=0, max=Tools.min(dest.length, source.length); i<max; i++) {dest[i]=source[i];}
	}
	
	public static void copy(int[] dest, int[] source) {
//		assert(a.length==b.length);
//		if(SIMDCOPY && Shared.SIMD && dest.length>=MINLEN32) {SIMD.copy(dest, source); return;}//TODO
		for(int i=0, max=Tools.min(dest.length, source.length); i<max; i++) {dest[i]=source[i];}
	}
	
	/** Returns number of matches */
	public static final int countMatches(final byte[] s1, final byte[] s2, int a1, int b1, int a2, int b2){
		assert(b1-a1==b2-a2) : a1+"-"+b1+", "+a2+"-"+b2+", len="+s1.length+", "+(b1-a1)+"!="+(b2-a2);
		if(Shared.SIMD && b1-a1+1>=MINLEN8) {return SIMD.countMatches(s1, s2, a1, b1, a2, b2);}
		int matches=0;
		for(int i=a1, j=a2; j<=b2; i++, j++) {
			final byte x=s1[i], y=s2[j];
			final int m=((x==y) ? 1 : 0);//Does not take into account capitalization or undefined bases
			matches+=m;
		}
		assert(matches>=0 && matches<=b1-a1+1);
		return matches;
	}
	
	public static final int countMismatches(final byte[] s1, final byte[] s2, int a1, int b1, int a2, int b2){
		return (b1-a1+1)-countMatches(s1, s2, a1, b1, a2, b2);
	}
	
	public static final int find(final byte[] a, final byte symbol, final int from, final int to){
//		if(Shared.SIMD && to-from>=MINLEN8) {return SIMD.find(a, symbol, from, to);}
		int len=from;
		while(len<to && a[len]!=symbol){len++;}
		return len;
	}
	
	
	public static double sum(float[] array){//
		if(array==null){return 0;}
		if(Shared.SIMD && array.length>=MINLEN32) {return SIMD.sum(array, 0, array.length-1);}
		double x=0;
		for(float y : array){x+=y;}
		return x;
	}

	public static long sum(byte[] array){
		if(array==null){return 0;}
		if(Shared.SIMD && array.length>=MINLEN8) {return SIMD.sum(array, 0, array.length-1);}
		long x=0;
		for(byte y : array){x+=y;}
		return x;
	}

	public static long sum(char[] array){
		if(array==null){return 0;}
//		if(Shared.SIMD && array.length>=SMINLEN) {return SIMD.sum(array, 0, array.length-1);}
		long x=0;
		for(char y : array){x+=y;}
		return x;
	}
	
	public static long sum(short[] array){
		if(array==null){return 0;}
//		if(Shared.SIMD && array.length>=SMINLEN) {return SIMD.sum(array, 0, array.length-1);}
		long x=0;
		for(short y : array){x+=y;}
		return x;
	}
	
	public static long sum(int[] array){
		if(array==null){return 0;}
		if(Shared.SIMD && array.length>=MINLEN32) {return SIMD.sum(array, 0, array.length-1);}
		long x=0;
		for(int y : array){x+=y;}
		return x;
	}

	public static double sum(double[] array){
		if(array==null){return 0;}
		if(Shared.SIMD && array.length>=MINLEN64) {return SIMD.sum(array, 0, array.length-1);}
		double x=0;
		for(double y : array){x+=y;}
		return x;
	}
	
	public static long sum(long[] array){
		if(array==null){return 0;}
		if(Shared.SIMD && array.length>=MINLEN64) {return SIMD.sum(array, 0, array.length-1);}
		long x=0;
		for(long y : array){x+=y;}
		return x;
	}
	
	public static long sum(int[] array, int from, int to){
		if(array==null){return 0;}
		if(Shared.SIMD && array.length>=MINLEN32) {return SIMD.sum(array, 0, array.length-1);}
		long x=0;
		for(int i=from; i<=to; i++){x+=array[i];}
		return x;
	}
	
	public static long sum(long[] array, int from, int to){
		if(array==null){return 0;}
		if(Shared.SIMD && array.length>=MINLEN64) {return SIMD.sum(array, 0, array.length-1);}
		long x=0;
		for(int i=from; i<=to; i++){x+=array[i];}
		return x;
	}
	
	public static final int max(int[] array){
		if(array==null){return 0;}
		if(Shared.SIMD && array.length>=MINLEN32) {return SIMD.max(array, 0, array.length-1);}
		int max=array[0];
		for(int i=1; i<array.length; i++){
			int x=array[i];
			max=(x>max ? x : max);
		}
		return max;
	}
	
	public static final float max(float[] array){
		if(array==null){return 0;}
		if(Shared.SIMD && array.length>=MINLEN32) {return SIMD.max(array, 0, array.length-1);}
		float max=array[0];
		for(int i=1; i<array.length; i++){
			float x=array[i];
			max=(x>max ? x : max);
		}
		return max;
	}
	
	public static final long max(long[] array){
		if(array==null){return 0;}
		if(Shared.SIMD && array.length>=MINLEN32) {return SIMD.max(array, 0, array.length-1);}
		long max=array[0];
		for(int i=1; i<array.length; i++){
			long x=array[i];
			max=(x>max ? x : max);
		}
		return max;
	}
	
	public static final int MINLEN8=32;
	public static final int MINLEN16=16;
	public static final int MINLEN32=16;//16 or 32 are optimal; 0, 24, and 48 are worse.
	public static final int MINLEN64=8;
	public static boolean SIMDCOPY=false;//Does not seem to affect speed, but could increase power usage.
	public static boolean SIMD_MULT_SPARSE=true;//Grants a speedup, and same results (but currently broken at ebs=1)
	public static boolean SIMD_FMA_SPARSE=true;//Grants a speedup, slightly different results
}