ReduceToSingleCS.hlsl 1.9 KB

12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576
  1. // RUN: %dxc -E main -T cs_6_0 %s | FileCheck %s
  2. // CHECK: groupId
  3. // CHECK: threadId
  4. // CHECK: flattenedThreadIdInGroup
  5. // CHECK: addrspace(3)
  6. // CHECK: barrier
  7. // CHECK: addrspace(3)
  8. // CHECK: barrier
  9. // CHECK: addrspace(3)
  10. // CHECK: barrier
  11. // CHECK: bufferStore
  12. //-----------------------------------------------------------------------------
  13. // File: ReduceToSingleCS.hlsl
  14. //
  15. // Desc: Reduce an input buffer by a factor of groupthreads
  16. //
  17. // Copyright (c) Microsoft Corporation. All rights reserved.
  18. //-----------------------------------------------------------------------------
  19. StructuredBuffer<float> Input : register( t0 );
  20. RWStructuredBuffer<float> Result : register( u0 );
  21. cbuffer cbCS : register( b0 )
  22. {
  23. uint4 g_param; // g_param.x is the actual elements contained in Input
  24. // g_param.y is the x dimension of the Dispatch call
  25. };
  26. #define groupthreads 128
  27. groupshared float accum[groupthreads];
  28. [numthreads(groupthreads,1,1)]
  29. void main( uint3 Gid : SV_GroupID, uint3 DTid : SV_DispatchThreadID, uint3 GTid : SV_GroupThreadID, uint GI : SV_GroupIndex )
  30. {
  31. if ( DTid.x < g_param.x )
  32. accum[GI] = Input[DTid.x];
  33. else
  34. accum[GI] = 0;
  35. // Parallel reduction algorithm follows
  36. GroupMemoryBarrierWithGroupSync();
  37. if ( GI < 64 )
  38. accum[GI] += accum[64+GI];
  39. GroupMemoryBarrierWithGroupSync();
  40. if ( GI < 32 )
  41. accum[GI] += accum[32+GI];
  42. GroupMemoryBarrierWithGroupSync();
  43. if ( GI < 16 )
  44. accum[GI] += accum[16+GI];
  45. GroupMemoryBarrierWithGroupSync();
  46. if ( GI < 8 )
  47. accum[GI] += accum[8+GI];
  48. GroupMemoryBarrierWithGroupSync();
  49. if ( GI < 4 )
  50. accum[GI] += accum[4+GI];
  51. GroupMemoryBarrierWithGroupSync();
  52. if ( GI < 2 )
  53. accum[GI] += accum[2+GI];
  54. GroupMemoryBarrierWithGroupSync();
  55. if ( GI < 1 )
  56. accum[GI] += accum[1+GI];
  57. if ( GI == 0 )
  58. {
  59. Result[Gid.x] = accum[0];
  60. }
  61. }