replication_stride.br

用于GPU通用计算的编程语言BrookGPU 0.4

// replication_stride.br
// Tests automatic replication and stride of kernel arguments
// The output stream always determines the number of operations
// performed, and elements from the input streams are either
// skipped or repeated to compensate.
// This test does not demonstrate it, but it is an error
// to use stream arguments that are not an integer multiple
// or divisor of the output stream's extents.

#include <stdio.h>

kernel void copy( float input<>, out float output<> ) {
	output = input;
}

void test1( int extentIn, int extentOut )
{
	float input< extentIn >;
	float output< extentOut >;
	float* input_data = 0;
	float* output_data = 0;
	int i;
	
	input_data = (float*)malloc( extentIn*sizeof(float) );
	output_data = (float*)malloc( extentOut*sizeof(float) );
	
	for( i = 0; i < extentIn; i++ )
		input_data[i] = (float)i;
	
	streamRead( input, input_data );
	
	copy( input, output );
	
	streamWrite( output, output_data );
	
	for( i = 0; i < extentOut; i++ )
		printf( "%6.3f ", output_data[i] );
	printf( "\n" );
}

void test2( int extentInX, int extentInY, int extentOutX, int extentOutY )
{
	float input< extentInY, extentInX >;
	float output< extentOutY, extentOutX >;
	float* input_data = 0;
	float* output_data = 0;
	int i, j;
	
	input_data = (float*)malloc( extentInX*extentInY*sizeof(float) );
	output_data = (float*)malloc( extentOutX*extentOutY*sizeof(float) );
	
	for( i = 0; i < extentInY; i++ )
		for( j = 0; j < extentInX; j++ )
			input_data[i*extentInX+j] = (float)(100*i+j);
	
	streamRead( input, input_data );
	
	copy( input, output );
	
	streamWrite( output, output_data );
	
	for( i = 0; i < extentOutY; i++ )
	{
		for( j = 0; j < extentOutX; j++ )
			printf( "%6.3f ", output_data[i*extentOutX+j] );
		printf("\n");
	}
}

int main()
{
	// 1-dimensional streams
	printf("5 -> 10\n");
	test1( 5, 10 );
	printf("10 -> 5\n");
	test1( 10, 5 );
	printf("7 -> 21\n");
	test1( 7, 21 );
	printf("21 -> 7\n");
	test1( 21, 7 );

	// 2-dimensional streams
	printf("15 -> 5, 2 ->4\n");
	test2( 15, 2, 5, 4 );
	printf("3 -> 21, 10 ->2\n");
	test2( 3, 10, 21, 2 );
	
	return 0;
}