Fix WavePrefixCountBits() being off by one.

It was counting bits up to the current lane included, whereas the
documentation says it should be excluded. This now matches dxc's behavior
as well.

Fix #2929

diff --git a/Test/baseResults/hlsl.waveprefix.comp.out b/Test/baseResults/hlsl.waveprefix.comp.out
index cc4737e..e4e942c 100644 (file)
--- a/Test/baseResults/hlsl.waveprefix.comp.out
+++ b/Test/baseResults/hlsl.waveprefix.comp.out
@@ -1126,7 +1126,7 @@ local_size = (32, 16, 1)
 0:54              0 (const int)
 0:54          Constant:
 0:54            0 (const int)
-0:54        subgroupBallotInclusiveBitCount ( temp uint)
+0:54        subgroupBallotExclusiveBitCount ( temp uint)
 0:54          subgroupBallot ( temp 4-component vector of uint)
 0:54            Compare Equal ( temp bool)
 0:54              direct index ( temp uint)
@@ -2289,7 +2289,7 @@ local_size = (32, 16, 1)
 0:54              0 (const int)
 0:54          Constant:
 0:54            0 (const int)
-0:54        subgroupBallotInclusiveBitCount ( temp uint)
+0:54        subgroupBallotExclusiveBitCount ( temp uint)
 0:54          subgroupBallot ( temp 4-component vector of uint)
 0:54            Compare Equal ( temp bool)
 0:54              direct index ( temp uint)
@@ -2818,7 +2818,7 @@ local_size = (32, 16, 1)
              390:      6(int) Load 389
              392:   391(bool) IEqual 390 26
              393:   13(ivec4) GroupNonUniformBallot 35 392
-             394:      6(int) GroupNonUniformBallotBitCount 35 InclusiveScan 393
+             394:      6(int) GroupNonUniformBallotBitCount 35 ExclusiveScan 393
              395:     42(ptr) AccessChain 24(data) 25 386 25 26
                               Store 395 394
                               Return

diff --git a/glslang/HLSL/hlslParseHelper.cpp b/glslang/HLSL/hlslParseHelper.cpp
index 2d0a8e9..e9369a0 100644 (file)
--- a/glslang/HLSL/hlslParseHelper.cpp
+++ b/glslang/HLSL/hlslParseHelper.cpp
@@ -5430,7 +5430,7 @@ void HlslParseContext::decomposeIntrinsic(const TSourceLoc& loc, TIntermTyped*&
         }
     case EOpWavePrefixCountBits:
         {
-            // Mapped to subgroupBallotInclusiveBitCount(subgroupBallot())
+            // Mapped to subgroupBallotExclusiveBitCount(subgroupBallot())
             // builtin
 
             // uvec4 type.
@@ -5444,7 +5444,7 @@ void HlslParseContext::decomposeIntrinsic(const TSourceLoc& loc, TIntermTyped*&
             TType uintType(EbtUint, EvqTemporary);
 
             node = intermediate.addBuiltInFunctionCall(loc,
-                EOpSubgroupBallotInclusiveBitCount, true, res, uintType);
+                EOpSubgroupBallotExclusiveBitCount, true, res, uintType);
 
             break;
         }