at_indexof_output.br

用于GPU通用计算的编程语言BrookGPU 0.4

// at_indexof_output.br

// test address-translation code for
// determining output index

#include <stdio.h>

// arbitrary large stream sizes
#define X_1D 10000

#define X_2D 5000
#define Y_2D 4

#define X_3D 40
#define Y_3D 10
#define Z_3D 7

#define X_4D 3
#define Y_4D 6
#define Z_4D 8
#define W_4D 5

kernel void getIndexStream1( out float result<> ) {
	result = (indexof result).x;
}

kernel void getIndexStream2( out float2 result<> ) {
	result = (indexof result).xy;
}

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

kernel void getIndexStream4( out float4 result<> ) {
	result = indexof result;
}

void check( int dim, int expected, float actual, const char* dimName )
{
	if( (int)actual == expected ) return;
	
	printf("FAIL: %dD.%s : expected %d got %f\n", dim, dimName, expected, actual );
	exit(0);
}

int main( int argc, char** argv )
{
	float stream_1d< X_1D >;
	float2 stream_2d< Y_2D, X_2D >;
	float3 stream_3d< Z_3D, Y_3D, X_3D >;
	float4 stream_4d< W_4D, Z_4D, Y_4D, X_4D >;
	
	float* data_1d = (float*)malloc( X_1D*sizeof(float) );
	float* data_2d = (float*)malloc( Y_2D*X_2D*sizeof(float2) );
	float* data_3d = (float*)malloc( Z_3D*Y_3D*X_3D*sizeof(float3) );
	float* data_4d = (float*)malloc( W_4D*Z_4D*Y_4D*X_4D*sizeof(float4) );
	
	float* temp;
	int x, y, z, w;
	
	getIndexStream1( stream_1d );
	getIndexStream2( stream_2d );
	getIndexStream3( stream_3d );
	getIndexStream4( stream_4d );
	
	streamWrite( stream_1d, data_1d );
	streamWrite( stream_2d, data_2d );
	streamWrite( stream_3d, data_3d );
	streamWrite( stream_4d, data_4d );
	
	temp = data_1d;
	for( x = 0; x < X_1D; x++ )
	{
		check( 1, x, *temp++, "x" );
	}
	
	temp = data_2d;
	for( y = 0; y < Y_2D; y++ )
	for( x = 0; x < X_2D; x++ )
	{
		check( 2, x, *temp++, "x" );
		check( 2, y, *temp++, "y" );
	}
	
	temp = data_3d;
	for( z = 0; z < Z_3D; z++ )
	for( y = 0; y < Y_3D; y++ )
	for( x = 0; x < X_3D; x++ )
	{
		check( 3, x, *temp++, "x" );
		check( 3, y, *temp++, "y" );
		check( 3, z, *temp++, "z" );
	}

	temp = data_4d;
	for( w = 0; w < W_4D; w++ )
	for( z = 0; z < Z_4D; z++ )
	for( y = 0; y < Y_4D; y++ )
	for( x = 0; x < X_4D; x++ )
	{
		check( 4, x, *temp++, "x" );
		check( 4, y, *temp++, "y" );
		check( 4, z, *temp++, "z" );
		check( 4, w, *temp++, "w" );
	}
	printf( "pass\n" );

	return 0;
}