at_stream_domain.br

用于GPU通用计算的编程语言BrookGPU 0.4

// at_stream_fetch.br

// test address-translation code for
// fetching from a stream domain

#include <stdio.h>

// arbitrary weird stream sizes:

#define RESULT_X 50
#define RESULT_Y 24
#define RESULT_Z 13

#define INPUT_X 51
#define INPUT_Y 40
#define INPUT_Z 19

#define INPUT_X_START 1
#define INPUT_X_END 51
#define INPUT_Y_START 5
#define INPUT_Y_END 29
#define INPUT_Z_START 0
#define INPUT_Z_END 13

// TIM: not guaranteed, but for now:
#define INPUT_TEXTURE_WIDTH 256
#define INPUT_TEXTURE_HEIGHT 152

kernel void clear3( out float3 result<> ) { result = -1; }

kernel void getIndexStream3( out float3 result<> ) {
	result = (indexof result).xyz;
}

kernel void copy3( float3 input<>, out float3 result<> ) {
	result = input;
}

void checkItem( const char* dimName, int expected, float actual, int x, int y, int z, int xd, int yd, int zd, const char* streamName )
{
	if( (int)actual != expected )
	{
		printf("FAIL: %s[%d][%d][%d].%s ~ <%d,%d,%d> : expected %d got %f\n", streamName, z, y, x, dimName, zd, yd, xd, expected, actual );
		exit(0);
	}
//	else
//		printf("%s[%d][%d][%d].%s ~ <%d,%d,%d> : %f\n", streamName, z, y, x, dimName, zd, yd, xd, actual );
}

void checkInput( float* inData )
{
	float* data = inData;
	int x, y, z;
	int linear;
	int tx, ty;
	
	for( z = 0; z < INPUT_Z; z++ )
	for( y = 0; y < INPUT_Y; y++ )
	for( x = 0; x < INPUT_X; x++ )
	{	
		linear = x + y*INPUT_X + z*INPUT_X*INPUT_Y;
		tx = linear / INPUT_TEXTURE_WIDTH;
		ty = linear % INPUT_TEXTURE_WIDTH;
	
		checkItem( "x", x, *data++, x, y, z, tx, ty, 0, "input" );
		checkItem( "y", y, *data++, x, y, z, tx, ty, 0, "input" );
		checkItem( "z", z, *data++, x, y, z, tx, ty, 0, "input" );		
	}
}

void check( float* inData )
{
	float* data = inData;
	int x, y, z;
	int xd, yd, zd;
	
	for( z = 0; z < RESULT_Z; z++ )
	for( y = 0; y < RESULT_Y; y++ )
	for( x = 0; x < RESULT_X; x++ )
	{
		xd = x + INPUT_X_START;
		yd = y + INPUT_Y_START;
		zd = z + INPUT_Z_START;
		
		checkItem( "x", xd, *data++, x, y, z, xd, yd, zd, "result" );
		checkItem( "y", yd, *data++, x, y, z, xd, yd, zd, "result" );
		checkItem( "z", zd, *data++, x, y, z, xd, yd, zd, "result" );
		
//		if( y > 5 ) exit(0);
	}
}

int main( int argc, char** argv )
{
	float3 result< RESULT_Z, RESULT_Y, RESULT_X >;
	float3 input< INPUT_Z, INPUT_Y, INPUT_X >;
	
	float* data_input = (float*)malloc( INPUT_Z*INPUT_Y*INPUT_X*sizeof(float3) );
	float* data_result = (float*)malloc( RESULT_Z*RESULT_Y*RESULT_X*sizeof(float3) );
	
	getIndexStream3( input );
	streamWrite( input, data_input );
	checkInput( data_input );
	
	clear3( result );
	copy3( input.domain(
		int3( INPUT_X_START, INPUT_Y_START, INPUT_Z_START ),
		int3( INPUT_X_END, INPUT_Y_END, INPUT_Z_END ) ), result );
	streamWrite( result, data_result );
	check( data_result );

	printf( "pass\n" );

	return 0;
}