read_write_domain.br

用于GPU通用计算的编程语言BrookGPU 0.4

// read_write_domain.br

// confirm that streamRead and streamWrite
// interact properly with domain

#include <stdio.h>

typedef struct Pair_t {
	float field1;
	float field2;
} Pair;

#define SIZE_X1 333

#define START_X1 127
#define END_X1 305
#define DOMAIN_X1 (END_X1 - START_X1)

#define SIZE_X2 427
#define SIZE_Y2 987

#define START_X2 0
#define END_X2 199
#define DOMAIN_X2 (END_X2 - START_X2)

#define START_Y2 576
#define END_Y2 987
#define DOMAIN_Y2 (END_Y2 - START_Y2)

kernel void clear2( out float2 result<> ) {
	result = -1;
}

kernel void clearPair( out Pair result<> ) {
	result.field1 = -1;
	result.field2 = -1;
}

void check( const char* name, const char* dim, int xidx, int yidx, int expected, float actual )
{
	if( (float)expected == actual ) return;
	
	printf("fail: %s[%d][%d].%s = %f, expected %d\n", name, yidx, xidx, dim, actual, expected );
	exit(-1);
}

void check1( const char* name, const float2* data )
{
	int i;
	int dx, dy;
	for( i = 0; i < SIZE_X1; i++ )
	{
		dx = -1;
		dy = -1;
		
		if( i >= START_X1 && i < END_X1 )
		{
			dx = i - START_X1;
			dy = i - START_X1;
		}
	
		check( name, "x", i, 0, dx, data[i].x );
		check( name, "y", i, 0, dy, data[i].y );
	}
}

void check2( const char* name, const float2* data )
{
	int i, j;
	int dx, dy;
	for( j = 0; j < SIZE_Y2; j++ )
	for( i = 0; i < SIZE_X2; i++ )
	{
		dx = -1;
		dy = -1;
		
		if( i >= START_X2 && i < END_X2
			&& j >= START_Y2 && j < END_Y2 )
		{
			dx = i - START_X2;
			dy = j - START_Y2;
		}

		check( name, "x", i, j, dx, data[j*SIZE_X2+i].x );
		check( name, "y", i, j, dy, data[j*SIZE_X2+i].y );
	}
}

int main()
{
	int i, j;
	float2 a1< SIZE_X1 >;
	Pair b1< SIZE_X1 >;
	float2 a2< SIZE_Y2, SIZE_X2 >;
	Pair b2< SIZE_Y2, SIZE_X2 >;
	
	float2* input1 = (float2*) malloc( DOMAIN_X1 * sizeof(float2) );
	float2* input2 = (float2*) malloc( DOMAIN_Y2 * DOMAIN_X2 * sizeof(float2) );

	float2* a1_data = (float2*) malloc( SIZE_X1 * sizeof(float2) );
	float2* b1_data = (float2*) malloc( SIZE_X1 * sizeof(float2) );
	float2* a2_data = (float2*) malloc( SIZE_Y2 * SIZE_X2 * sizeof(float2) );
	float2* b2_data = (float2*) malloc( SIZE_Y2 * SIZE_X2 * sizeof(float2) );
	
	for( i = 0; i < DOMAIN_X1; i++ )
	{
		input1[ i ].x = (float)i;
		input1[ i ].y = (float)i;
	}

	for( j = 0; j < DOMAIN_Y2; j++ )
	for( i = 0; i < DOMAIN_X2; i++ )
	{
		input2[ j*DOMAIN_X2 + i ].x = (float)i;
		input2[ j*DOMAIN_X2 + i ].y = (float)j;
	}
	
	clear2( a1 );
	clear2( a2 );
	
	clearPair( b1 );
	clearPair( b2 );
	
	streamRead( a1.domain( START_X1, END_X1 ), input1 );
	streamRead( b1.domain( START_X1, END_X1 ), input1 );
	streamRead( a2.domain( int2( START_X2, START_Y2 ), int2( END_X2, END_Y2 ) ), input2 );
	streamRead( b2.domain( int2( START_X2, START_Y2 ), int2( END_X2, END_Y2 ) ), input2 );
	
	streamWrite( a1, a1_data );
	streamWrite( b1, b1_data );
	streamWrite( a2, a2_data );
	streamWrite( b2, b2_data );
	
	check1( "a1", a1_data );
	check1( "b1", b1_data );
	check2( "a2", a2_data );
	check2( "b2", b2_data );

	printf( "pass\n" );
	return 0;
}