/*
 * test.cl
 *
 *  Created on: Jan 22, 2012
 *      Author: scannon
 */


#ifndef TEST_CL
#define TEST_CL

//#include "ranluxcl.cl"
//#include "Perceptron.cl"


 char test_PerceptronNode(ranluxcl_state_t *rst)
{
	PerceptronNode pNode1;
	float16 w1 = (float16)(0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	float16 i1 = (float16)(1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f);
	PerceptronNode_init(&pNode1, w1, -7.0f);
	if(PerceptronNode_calc(&pNode1, i1) >= 0.001f)
		return 1;
 
	PerceptronNode pNode2;
	float16 w2 = (float16)(1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f);
	float16 i2 = (float16)(0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	PerceptronNode_init(&pNode2, w2, -7.0f);
	if(PerceptronNode_calc(&pNode2, i2) >= 0.001f)
		return 2;
 
	PerceptronNode pNode3;
	float16 w3 = (float16)(1.0f, 1.0f, 1.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	float16 i3 = (float16)(0.75f, 0.75f, 0.75f, 0.75f, 0.75f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	PerceptronNode_init(&pNode3, w3, -2.0f);
	if(PerceptronNode_calc(&pNode3, i3) - 0.731f > 0.001f ||
	   PerceptronNode_calc(&pNode3, i3) - 0.731f < -0.001f)
		return 3;
 
	PerceptronNode pNode4;
	float16 i4 = (float16)(0.75f, 0.75f, 0.75f, 0.75f, 0.75f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	PerceptronNode_init2(&pNode4, 4);
	if(PerceptronNode_calc(&pNode4, i4) - 0.731f > 0.001f ||
	   PerceptronNode_calc(&pNode4, i4) - 0.731f < -0.001f)
		return 4;

	PerceptronNode pNode5;
	float16 w5 = (float16)(1.0f, 3.0f, 2.0f, 4.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	float16 i5 = (float16)(1.0f, 3.0f, 5.0f, 7.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	PerceptronNode_init(&pNode5, w5, -7.0f);
	if(PerceptronNode_calc(&pNode5, i5) - 0.999f > 0.001f ||
	   PerceptronNode_calc(&pNode5, i5) - 0.999f < -0.001f)
		return 5;
	
	PerceptronNode pNode6;
	PerceptronNode pNode7;
	PerceptronNode_init3(&pNode6, 7, rst);
	PerceptronNode_init3(&pNode7, 7, rst);
	if(PerceptronNode_calc(&pNode6, i5) > 1 || PerceptronNode_calc(&pNode6, i5) < 0)
		return 6;
	if(PerceptronNode_calc(&pNode7, i5) > 1 || PerceptronNode_calc(&pNode7, i5) < 0)
		return 6;
	if(pNode6.w.s0 == pNode7.w.s0 &&
		pNode6.w.s1 == pNode7.w.s1 &&
		pNode6.w.s2 == pNode7.w.s2 &&
		pNode6.w.s3 == pNode7.w.s3 &&
		pNode6.w.s4 == pNode7.w.s4 &&
		pNode6.w.s5 == pNode7.w.s5 &&
		pNode6.w.s6 == pNode7.w.s6)
		return 6;

	PerceptronNode pNode8;
	float16 i8 = (float16)(0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	PerceptronNode_init2(&pNode8, 0);
	if(PerceptronNode_calc(&pNode8, i8) - 0.5f > 0.001f ||
	   PerceptronNode_calc(&pNode8, i8) - 0.5f < -0.001f)
		return 7;
	
	return 0;
}
 
char test_Perceptron(ranluxcl_state_t *rst)
{
	int i = 0;
	
	Perceptron p1;
	float16 i1 = (float16)(1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	float y1[48];
	Perceptron_init2(&p1, 10);
	Perceptron_calc2(&p1, i1, y1);
	if(y1[0] - 0.999f > 0.001f ||
	   y1[0] - 0.999f < -0.001f)
		return 11;
	if(y1[47] - 0.999f > 0.001f ||
	   y1[47] - 0.999f < -0.001f)
		return 11;
	
	Perceptron p2;
	PerceptronNode hiddenNodes[16];
	PerceptronNode outputNodes[48];
	float16 w2 = (float16)(0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	float16 wh2a = (float16)(11.0f, 11.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	float16 wh2b = (float16)(5.0f, 5.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	float16 w02 = (float16)(11.0f, -15.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	PerceptronNode_init(&hiddenNodes[0], wh2a, -7.0f);
	PerceptronNode_init(&hiddenNodes[1], wh2b, -7.0f);
	for(i = 2; i < 16; i++)
		PerceptronNode_init(&hiddenNodes[i], w2, -7.0f);
	PerceptronNode_init(&outputNodes[0], w02, -7.0f);
	for(i = 1; i < 48; i++)
		PerceptronNode_init(&outputNodes[i], w2, -7.0f);
	Perceptron_init(&p2, hiddenNodes, outputNodes);
	float y[48];
	float16 i2a = (float16)(0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	float16 i2b = (float16)(1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	float16 i2c = (float16)(0.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	float16 i2d = (float16)(1.0f, 1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	Perceptron_calc2(&p2, i2a, y);
	if(y[0] >= 0.15f)
		return 12;
	Perceptron_calc2(&p2, i2b, y);
	if(y[0] <= 0.85f)
		return 13;
	Perceptron_calc2(&p2, i2c, y);
	if(y[0] <= 0.85f)
		return 14;
	Perceptron_calc2(&p2, i2d, y);
	if(y[0] >= 0.15f)
		return 15;
	
	return 0;
}

char test_GMM(ranluxcl_state_t *rst)
{
	int numSamples = 1000;
	int i = 0;
	float mean = 0;
	float sumX2 = 0;
	float var = 0;
	Perceptron p1;
	Perceptron_init2(&p1, 0);
	GMM node1;
	GMM__init(&node1, &p1, 0, 0);
	for(i = 0; i < numSamples; i++)
	{
		float sample = GMM_sample(&node1, rst);
		mean += sample;
		sumX2 += sample*sample;
		node1.sampled = false;
	}
	mean /= numSamples;
	var = sumX2/numSamples - mean*mean;
	if(mean - 0.5f > 0.05f || mean - 0.5f < -0.05f)
		return 21;
	if(var - 0.25f > 0.05f || var - 0.25f < -0.05f)
		return 22;

	PerceptronNode hiddenNodes[16];
	for(i = 0; i < 16; i++)
		PerceptronNode_init2(&hiddenNodes[i], 0);
	PerceptronNode outputNodes[48];
	float16 w1 = (float16)(0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f);
	float bias[3] = {0.0f, -4.0f, 0.0f};
	for(i = 0; i < 48; i++)
		PerceptronNode_init(&outputNodes[i], w1, bias[i%3]);
	Perceptron p2;
	Perceptron_init(&p1, hiddenNodes, outputNodes);
	Perceptron_init(&p2, hiddenNodes, outputNodes);
	GMM node2;
	GMM__init(&node1, &p1, 0, 0);
	GMM__init(&node2, &p2, 0, 0);
	Perceptron p3;
	PerceptronNode hiddenNodes2[16];
	for(i = 0; i < 16; i++)
		PerceptronNode_init2(&hiddenNodes2[i], 2);
	PerceptronNode outputNodes2[48];
	float16 w2 = (float16)(1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f);
	float bias2[3] = {-7.0f, -12.0f, 0.0f};
	for(i = 0; i < 48; i++)
		PerceptronNode_init(&outputNodes2[i], w2, bias2[i%3]);
	Perceptron_init(&p3, hiddenNodes2, outputNodes2);
	GMM node3;
	GMM * parents[16] = {&node1, &node2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
	GMM__init(&node3, &p3, parents, 2);
	mean = 0;
	sumX2 = 0;
	for(i = 0; i < numSamples; i++)
	{
		float sample = GMM_sample(&node3, rst);
		mean += sample;
		sumX2 += sample*sample;
		node3.sampled = false;
	}
	mean /= numSamples;
	var = sumX2/numSamples - mean*mean;
	if(mean - 0.73f > 0.05f || mean - 0.73f < -0.05f)
		return 23;
	if(var - 0.0003f > 0.0005f || var - 0.0003f < -0.0005f)
		return 10000*var;//24;

	return 0;
}

kernel void Kernel_Ranluxcl_Init(private uint ins,
                                 global ranluxcl_state_t *ranluxcltab)
{
	ranluxcl_initialization(ins, ranluxcltab);
}
        
__kernel void test_PerceptronNode_kernel(global ranluxcl_state_t *ranluxcltab,
                                         __global char *result)
{
	int gid = get_global_id(0);
	ranluxcl_state_t ranluxclstate;

	// Download state
	ranluxcl_download_seed(&ranluxclstate, ranluxcltab);

	// Conduct tests
	result[gid] = 0;
	result[gid] = test_PerceptronNode(&ranluxclstate);
		
	// Upload state again
	ranluxcl_upload_seed(&ranluxclstate, ranluxcltab);
}

__kernel void test_Perceptron_kernel(global ranluxcl_state_t *ranluxcltab,
                                     __global char *result)
{
	int gid = get_global_id(0);
	ranluxcl_state_t ranluxclstate;

	// Download state
	ranluxcl_download_seed(&ranluxclstate, ranluxcltab);

	// Conduct tests
	result[gid] = 0;
	result[gid] = test_Perceptron(&ranluxclstate);
		
	// Upload state again
	ranluxcl_upload_seed(&ranluxclstate, ranluxcltab);
}

__kernel void test_GMM_kernel(global ranluxcl_state_t *ranluxcltab,
                              __global char *result)
{
	int gid = get_global_id(0);
	ranluxcl_state_t ranluxclstate;

	// Download state
	ranluxcl_download_seed(&ranluxclstate, ranluxcltab);

	// Conduct tests
	result[gid] = 0;
	result[gid] = test_GMM(&ranluxclstate);
		
	// Upload state again
	ranluxcl_upload_seed(&ranluxclstate, ranluxcltab);
}

#endif	// TEST_CL