struct_gather.br

用于GPU通用计算的编程语言BrookGPU 0.4

// struct_gather.br

#include <stdio.h>

typedef struct Pair_t {
  float first;
  float second;
} Pair;
/*
typedef struct traverser_t {
  float4 index;//.xy is index into the aTree  .zw is index into bTree
  float3 Translation; 
  float3 Rotationx;
  float3 Rotationy;
  float3 Rotationz;
}Traverser;
*/
typedef struct bbox_t{
  float4 Rotationx;
  float3 Rotationy;
  // float3 mRotationz  // since Rotationx and Rotationy are orthogonal
  /// cross(Rotationx,Rotationy);
  float3 Translation;
  float4 Radius;// if it's a leaf Radius.w is 1 else Radius.w = 0

  // if leaf, the Children.xy is an index to the Triangle
  // if node, the Children.xy is an index to left child
  // assert right.xy is always left + {1,0} this may require gaps in the tree
  float2 Children;
}BBox;

kernel void gatherBBox( BBox tmp [][], out float2 child<>) {
  float2 ind={2,0};
  BBox gath = tmp[ind];
  child = gath.Children;
}

kernel void gatherPairs( float index<>, Pair pairs[], out float result<> ) {
  Pair p = pairs[index];
  result = p.first + p.second;
}

int main() {
  BBox bboxes<1,4>;
  float2 bboxout<1,1>;
  BBox assign[4];
  Pair pairs< 100 >;
  float indices< 100 >;
  float output< 100 >;
  Pair pairs_data[100];
  float indices_data[100];
  float output_data[100];
  int i;
  /*Traverser baseTraverser;
    Traverser trav<1,1>;
  baseTraverser.index=float4(2,0,2,0);
  baseTraverser.Translation=float3(-0.13644515,  0.112712905,  -0.0222729594);
  baseTraverser.Rotationx=float3(-0.170251727,  0.966706932,  -0.191028923);
  baseTraverser.Rotationy=float3(0.966706932, 0.201434806,  0.15780279);
  baseTraverser.Rotationz=float3(-0.191028923,  0.15780279,  0.968816876);
  streamRead(trav,&baseTraverser);
  */
  assign[0].Children=float2(2,0);
  assign[0].Rotationx=float4( -0.0854557604, 0.99632597, -0.0056440942,-1);
  assign[0].Rotationy=float3(0.978623033, 0.0849982575, 0.187275991);
  assign[0].Radius=float4(0.0763418078,  0.100202203,  0.0545564666,  0);
  assign[1].Children=float2(0,0);
  assign[1].Rotationx=float4(1,0,0,1);
  assign[1].Rotationy=float3(0,1,0);
  assign[1].Children=float2(0,0);
  assign[1].Rotationx=float4(1,0,0,1);
  assign[1].Rotationy=float3(0,1,0);
  assign[2].Children=float2(4,0);
  assign[2].Rotationx=float4(1,0,0,1);
  assign[2].Rotationy=float3(0,1,0);
  assign[3].Children=float2(1006,1);
  assign[3].Rotationx=float4(1,0,0,1);
  assign[3].Rotationy=float3(0,1,0);
  streamRead(bboxes,assign);
  gatherBBox(bboxes,bboxout);
  streamPrint(bboxout);
  for( i = 0; i < 100; i++ )
  {
    pairs_data[i].first = (float)i;
    pairs_data[i].second = (float)(i*100);
    indices_data[i] = (float)(99-i);
  }

  streamRead( pairs, pairs_data );
  streamRead( indices, indices_data );
  gatherPairs( indices, pairs, output );
  streamWrite( output, output_data );
	
	for( i = 0; i < 100; i++ )
	{
		printf( "%6.1f ", output_data[i] );
    if( i % 10 == 9 )
      printf("\n");
	}
	
	return 0;
}