fix bugs (false assert triggers)

Author: devshgraphicsprogramming

include/nbl/builtin/hlsl/macros.h

@@ -6,6 +6,7 @@
 
 #ifdef __HLSL_VERSION 
 #define static_assert(...) _Static_assert(__VA_ARGS__)
+// TODO: switch between enabled and disabled depending on Nabla build config
 #define assert(expr) \
 { \
     bool con = (expr); \

include/nbl/builtin/hlsl/workgroup/ballot.hlsl

@@ -20,7 +20,6 @@ namespace impl
 uint16_t getDWORD(uint16_t invocation)
 {
     uint16_t dword = invocation>>5;
-    assert(dword<((Volume()+31)>>5));
     return dword; // log2 of sizeof(uint32_t)*8
 }
 
@@ -29,6 +28,10 @@ uint16_t BallotDWORDCount(const uint16_t itemCount)
 {
     return getDWORD(itemCount+31); // round up, in case all items don't fit in even number of DWORDs
 }
+
+// this silly thing exists only because we can't make the above `constexpr`
+template<uint16_t ItemCount>
+struct ballot_dword_count : integral_constant<uint16_t,((ItemCount+31)>>5)> {};
 }
 
 /**
@@ -55,17 +58,15 @@ void ballot(const bool value, NBL_REF_ARG(Accessor) accessor)
     if (initialize)
         accessor.set(index,0u);
 
-    accessor.ballot.workgroupExecutionAndMemoryBarrier();
+    accessor.workgroupExecutionAndMemoryBarrier();
     if(value)
-    {
-        uint32_t dummy;
-        accessor.atomicOr(impl::getDWORD(index),1u<<(index&31u),dummy);
-    }
+        accessor.atomicOr(impl::getDWORD(index),1u<<(index&31u));
 }
 
 template<class Accessor>
 bool ballotBitExtract(const uint16_t index, NBL_REF_ARG(Accessor) accessor)
 {
+    assert(index<Volume());
     return bool(accessor.get(impl::getDWORD(index))&(1u<<(index&31u)));
 }
 

include/nbl/builtin/hlsl/workgroup/basic.hlsl

@@ -25,7 +25,9 @@ uint16_t Volume()
 
 uint16_t SubgroupContiguousIndex()
 {
-    return uint16_t(glsl::gl_SubgroupID()*glsl::gl_SubgroupSize()+glsl::gl_SubgroupInvocationID());
+    const uint16_t retval = (uint16_t(glsl::gl_SubgroupID())<<glsl::gl_SubgroupSizeLog2())+uint16_t(glsl::gl_SubgroupInvocationID());
+    assert(retval<Volume());
+    return retval;
 }
 
 bool Elect()

src/nbl/builtin/CMakeLists.txt

@@ -276,7 +276,7 @@ LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/workgroup/basic.hlsl")
 LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/workgroup/ballot.hlsl")
 LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/workgroup/broadcast.hlsl")
 LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/workgroup/scratch_size.hlsl")
-LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/workgroup/shared_Scan.hlsl")
+LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/workgroup/shared_scan.hlsl")
 
 
 macro(NBL_ADD_BUILTIN_RESOURCES _TARGET_) # internal & Nabla only, must be added with the macro to properly propagate scope