Add binop::identity() as well as attempt at SPIR-V implementations of subgroup barrier ops

Author: kpentaris

include/nbl/builtin/hlsl/binops.hlsl

@@ -17,6 +17,11 @@ struct bitwise_and
     {
         return lhs&rhs;
     }
+	
+	T identity()
+	{
+		return ~0;
+	}
 };
 
 template<typename T>
@@ -26,6 +31,11 @@ struct bitwise_or
     {
         return lhs|rhs;
     }
+	
+	T identity()
+	{
+		return 0;
+	}
 };
 
 template<typename T>
@@ -35,6 +45,11 @@ struct bitwise_xor
     {
         return lhs^rhs;
     }
+	
+	T identity()
+	{
+		return 0;
+	}
 };
 
 template<typename T>
@@ -44,6 +59,11 @@ struct add
     {
         return lhs+rhs;
     }
+	
+	T identity()
+	{
+		return 0;
+	}
 };
 
 template<typename T>
@@ -53,6 +73,11 @@ struct mul
     {
         return lhs*rhs;
     }
+	
+	T identity()
+	{
+		return 1;
+	}
 };
 
 template<typename T, class Comparator>
@@ -72,6 +97,11 @@ struct min
 		comparator_lt_t comp;
         return bitwise_min(lhs, rhs, comp);
     }
+	
+	T identity()
+	{
+		return ~0;
+	}
 };
 
 template<typename T, class Comparator>
@@ -91,6 +121,11 @@ struct max
 		comparator_gt_t comp;
         return bitwise_max(lhs, rhs, comp);
     }
+	
+	T identity()
+	{
+		return 0; // REVIEW: This assumes T = unsigned but what if we got T = signed ?
+	}
 };
 
 template<typename T>

include/nbl/builtin/hlsl/subgroup/arithmetic_portability_impl.hlsl

@@ -265,9 +265,10 @@ struct ScratchAccessorAdaptor {
 struct scan_base
 {
    // even if you have a `const uint nbl::hlsl::subgroup::Size` it wont work I think, so `#define` needed
-   static const uint HalfSubgroupSize = WaveGetLaneCount()>>1u; // TODO (PentaKon): Replace with nbl_hlsl_SubgroupSize or nbl::hlsl::subgroup::Size
-   static const uint LoMask = WaveGetLaneCount()-1u; // TODO (PentaKon): Replace with nbl_hlsl_SubgroupSize
-   static const uint LastWorkgroupInvocation = _NBL_HLSL_WORKGROUP_SIZE_-1; // TODO (PentaKon): Where should this be defined?
+   static const uint SubgroupSize = nbl::hlsl::subgroup::subgroupSize();
+   static const uint HalfSubgroupSize = SubgroupSize>>1u; // REVIEW: Is this ok?
+   static const uint LoMask = SubgroupSize-1u;
+   static const uint LastWorkgroupInvocation = _NBL_HLSL_WORKGROUP_SIZE_-1; // REVIEW: Where should this be defined?
    static const uint pseudoSubgroupInvocation = localInvocationIndex&LoMask; // Also used in substructs, thus static const
 
     static inclusive_scan<Binop,ScratchAccessor> create()
@@ -283,7 +284,7 @@ struct scan_base
        retval.scanStoreOffset = paddingMemoryEnd+pseudoSubgroupInvocation;
 
 		uint reductionResultOffset = paddingMemoryEnd;
-		if ((LastWorkgroupInvocation>>firstbithigh(WaveGetLaneCount()))!=nbl::hlsl::subgroup::ID()) // TODO (PentaKon): Replace with nbl_hlsl_SubgroupSizeLog2
+		if ((LastWorkgroupInvocation>>nbl::hlsl::subgroup::subgroupSizeLog2())!=nbl::hlsl::subgroup::subgroupInvocationID())
            retval.reductionResultOffset += LastWorkgroupInvocation&LoMask;
 		else
            retval.reductionResultOffset += LoMask;
@@ -319,7 +320,7 @@ struct inclusive_scan : scan_base
            nbl::hlsl::subgroupMemoryBarrierShared();
            scratchAccessor.set(scanStoreOffset ,value);
            if (scan_base::pseudoSubgroupInvocation<scan_base::HalfSubgroupSize)
-               scratchAccessor.set(lastLoadOffset,Binop::identity());
+               scratchAccessor.set(lastLoadOffset,op::identity());
        }
        nbl::hlsl::subgroupBarrier();
        nbl::hlsl::subgroupMemoryBarrierShared();
@@ -405,7 +406,7 @@ struct reduction
        nbl::hlsl::subgroupBarrier();
        nbl::hlsl::subgroupMemoryBarrierShared();
        uint reductionResultOffset = impl.paddingMemoryEnd;
-       if ((scan_base::LastWorkgroupInvocation>>firstbithigh(WaveGetLaneCount()))!=nbl::hlsl::subgroup::ID()) // TODO (PentaKon): Replace with nbl_hlsl_SubgroupSizeLog2
+       if ((scan_base::LastWorkgroupInvocation>>nbl::hlsl::subgroup::subgroupSizeLog2())!=nbl::hlsl::subgroup::ID())
            reductionResultOffset += scan_base::LastWorkgroupInvocation & scan_base::LoMask;
        else
            reductionResultOffset += scan_base::LoMask;

include/nbl/builtin/hlsl/subgroup/basic_portability.hlsl

@@ -7,6 +7,9 @@ namespace hlsl
 {
 	static const uint MaxWorkgroupSizeLog2 = 11;
 	static const uint MaxWorkgroupSize = 0x1 << MaxWorkgroupSizeLog2;
+	
+namespace subgroup
+{
 	static const uint MinSubgroupSizeLog2 = 2;
 	static const uint MinSubgroupSize = 0x1 << MinSubgroupSizeLog2;
 
@@ -60,9 +63,27 @@ namespace hlsl
 
 	// WAVE BARRIERS
 
+	// REVIEW: Review everything related to subgroup barriers and SPIR-V injection
+	
+	[[vk::ext_instruction(/* subgroupBarrier */ 224)]] // https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#OpControlBarrier
+	void spirv_subgroupBarrier(uint executionScope, uint memoryScope, uint memorySemantics);
+	
 	void subgroupBarrier() {
+		// https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_scope_id
+		// Subgroup scope is number 3
 
+		// https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#_memory_semantics_id
+		// By providing memory semantics None we do both control and memory barrier as is done in GLSL
+		spirv_subgroupBarrier(3, 3, 0x0);
 	}
+	
+	[[vk::ext_instruction(/* subgroupMemoryBarrierShared */ 225)]] // https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#OpControlBarrier
+	void spirv_subgroupMemoryBarrierShared(uint memoryScope, uint memorySemantics);
+	
+	void subgroupMemoryBarrierShared() {
+		spirv_subgroupMemoryBarrierShared(3, 0x0); // REVIEW: Need advice on memory semantics. Would think SubgroupMemory(0x80) but have no idea
+	}
+}
 }
 }