fix header bugs and update submodule

Author: devshgraphicsprogramming

examples_tests

@@ -194,7 +194,7 @@ struct num_traits : num_base<T>
 {
     // have to be weird like that to avoid a warning
     NBL_CONSTEXPR_STATIC_INLINE T min            = T(num_base<T>::is_signed)<<(num_base<T>::is_signed ? num_base<T>::digits:0);
-    NBL_CONSTEXPR_STATIC_INLINE T max            = ~(T(num_base<T>::is_signed)<<num_base<T>::digits);
+    NBL_CONSTEXPR_STATIC_INLINE T max            = ~min;
     NBL_CONSTEXPR_STATIC_INLINE T denorm_min     = T(0);
     NBL_CONSTEXPR_STATIC_INLINE T quiet_NaN      = T(0);
     NBL_CONSTEXPR_STATIC_INLINE T signaling_NaN  = T(0);

include/nbl/builtin/hlsl/limits.hlsl

@@ -50,8 +50,8 @@ uint16_t BallotDWORDCount(const uint16_t itemCount)
 template<class Accessor>
 void ballot(const bool value, NBL_REF_ARG(Accessor) accessor)
 {
-    const uint32_t index = SubgroupContiguousIndex();
-    const bool initialize = index<BallotDWORDCount(Volume());
+    const uint16_t index = SubgroupContiguousIndex();
+    const bool initialize = index<impl::BallotDWORDCount(Volume());
     if (initialize)
         accessor.set(index,0u);
 
@@ -64,7 +64,7 @@ void ballot(const bool value, NBL_REF_ARG(Accessor) accessor)
 }
 
 template<class Accessor>
-bool ballotBitExtract(const uint32_t index, NBL_REF_ARG(Accessor) accessor)
+bool ballotBitExtract(const uint16_t index, NBL_REF_ARG(Accessor) accessor)
 {
     return bool(accessor.get(impl::getDWORD(index))&(1u<<(index&31u)));
 }

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

@@ -17,25 +17,25 @@ namespace workgroup
 static const uint32_t MaxWorkgroupSizeLog2 = 11;
 static const uint32_t MaxWorkgroupSize = 0x1u<<MaxWorkgroupSizeLog2;
 
-uint32_t Volume()
+uint16_t Volume()
 {
-    const uint32_t3 dims = glsl::gl_WorkGroupSize();
+    const uint16_t3 dims = uint16_t3(glsl::gl_WorkGroupSize());
     return dims.x*dims.y*dims.z;
 }
 
-uint32_t SubgroupContiguousIndex()
+uint16_t SubgroupContiguousIndex()
 {
-    return glsl::gl_SubgroupID()*glsl::gl_SubgroupSize()+glsl::gl_SubgroupInvocationID();
+    return uint16_t(glsl::gl_SubgroupID()*glsl::gl_SubgroupSize()+glsl::gl_SubgroupInvocationID());
 }
 
 bool Elect()
 {
     return glsl::gl_SubgroupID()==0 && glsl::gl_SubgroupInvocationID()==0;
 }
 
-uint32_t ElectedSubgroupContiguousIndex()
+uint16_t ElectedSubgroupContiguousIndex()
 {
-    return glsl::subgroupBroadcastFirst<uint32_t>(SubgroupContiguousIndex());
+    return glsl::subgroupBroadcastFirst<uint16_t>(SubgroupContiguousIndex());
 }
 
 }

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

@@ -6,6 +6,7 @@
 
 
 #include "nbl/builtin/hlsl/type_traits.hlsl"
+#include "nbl/builtin/hlsl/subgroup/basic.hlsl"
 
 
 namespace nbl
@@ -37,7 +38,7 @@ struct scratch_size_ballot
 };
 
 // you're only writing one element
-NBL_CONSTEXPR_STATIC scratch_size_broadcast = 1u;
+NBL_CONSTEXPR uint32_t scratch_size_broadcast = 1u;
 
 // if you know better you can use the actual subgroup size
 template<uint16_t ContiguousItemCount, uint16_t SubgroupSize=subgroup::MinSubgroupSize>