Merge pull request #850 from Devsh-Graphics-Programming/full_inline_spirv_bda

Full Inline SPIR-V BDA pointers

Author: devshgraphicsprogramming

examples_tests

@@ -1,7 +1,6 @@
 // Copyright (C) 2018-2024 - DevSH Graphics Programming Sp. z O.O.
 // This file is part of the "Nabla Engine".
 // For conditions of distribution and use, see copyright notice in nabla.h
-
 #include "nbl/builtin/hlsl/type_traits.hlsl"
 #include "nbl/builtin/hlsl/bda/__ref.hlsl"
 
@@ -17,32 +16,64 @@ namespace bda
 template<typename T>
 struct __ptr
 {
-    using this_t = __ptr <T>;
-    uint64_t addr;
+    using this_t = __ptr<T>;
+    uint32_t2 addr;
 
-    static this_t create(const uint64_t _addr)
+    static this_t create(const uint32_t2 _addr)
     {
         this_t retval;
         retval.addr = _addr;
         return retval;
     }
 
-    template< uint64_t alignment=alignment_of_v<T> >
-    __ref<T,alignment,false> deref()
+    // in non-64bit mode we only support "small" arithmetic on pointers (just offsets no arithmetic on pointers)
+    __ptr operator+(uint32_t i)
+    {
+        i *= sizeof(T);
+        uint32_t2 newAddr = addr;
+        spirv::AddCarryOutput<uint32_t> lsbAddRes = spirv::addCarry<uint32_t>(addr[0],i);
+        newAddr[0] = lsbAddRes.result;
+        newAddr[1] += lsbAddRes.carry;
+        return __ptr::create(newAddr);
+    }
+    __ptr operator-(uint32_t i)
+    {
+        i *= sizeof(T);
+        uint32_t2 newAddr = addr;
+        spirv::AddCarryOutput<uint32_t> lsbSubRes = spirv::subBorrow<uint32_t>(addr[0],i);
+        newAddr[0] = lsbSubRes.result;
+        newAddr[1] -= lsbSubRes.carry;
+        return __ptr::create(newAddr);
+    }
+
+    template<uint64_t alignment=alignment_of_v<T>, bool _restrict=false>
+    __ref<T,alignment,_restrict> deref()
     {
         // TODO: assert(addr&uint64_t(alignment-1)==0);
-        using retval_t = __ref < T, alignment, false>;
-        retval_t retval;
-        retval.__init(addr);
+        __ref<T,alignment,_restrict> retval;
+        retval.__init(spirv::bitcast<spirv::bda_pointer_t<T>,uint32_t2>(addr));
         return retval;
     }
+    
+    template<uint64_t alignment=alignment_of_v<T> >
+    __ref<T,alignment,true> deref_restrict() {return deref<alignment,true>();}
 
-    __ptr operator +(int64_t i) {
-        return __ptr::create(addr + sizeof(T) * i);
+    //! Dont use these, to avoid emitting shaderUint64 capability when compiling for crappy mobile GPUs
+    static this_t create(const uint64_t _addr)
+    {
+        this_t retval;
+        retval.addr = spirv::bitcast<uint32_t2>(_addr);
+        return retval;
     }
-
-    __ptr operator-(int64_t i) {
-        return __ptr::create(addr - sizeof(T) * i);
+    __ptr operator+(int64_t i)
+    {
+        i *= sizeof(T);
+        return __ptr::create(spirv::bitcast<uint64_t>(addr)+i);
+    }
+    __ptr operator-(int64_t i)
+    {
+        i *= sizeof(T);
+        return __ptr::create(spirv::bitcast<uint64_t>(addr)-i);
     }
 };
 

include/nbl/builtin/hlsl/bda/__ptr.hlsl

@@ -4,39 +4,44 @@
 #ifndef _NBL_BUILTIN_HLSL_BDA_REF_INCLUDED_
 #define _NBL_BUILTIN_HLSL_BDA_REF_INCLUDED_
 
+// TODO: this shouldn't be included IMHO
 #include "nbl/builtin/hlsl/functional.hlsl"
+#include "nbl/builtin/hlsl/spirv_intrinsics/core.hlsl"
 
 namespace nbl
 {
 namespace hlsl
 {
 namespace bda
 {
+template<typename T, bool _restrict>
+struct __spv_ptr_t;
 template<typename T>
-using __spv_ptr_t __NBL_CAPABILITY_PhysicalStorageBufferAddresses = spirv::pointer_t<spv::StorageClassPhysicalStorageBuffer,T>;
-
+struct __spv_ptr_t<T,false>
+{
+    [[vk::ext_decorate(spv::DecorationAliasedPointer)]] spirv::bda_pointer_t<T> value;
+};
 template<typename T>
-struct __ptr;
+struct __spv_ptr_t<T,true>
+{
+    [[vk::ext_decorate(spv::DecorationRestrictPointer)]] spirv::bda_pointer_t<T> value;
+};
 
-// TODO: refactor this in terms of `nbl::hlsl::` when they fix the composite struct inline SPIR-V BDA issue
 template<typename T, uint32_t alignment, bool _restrict>
 struct __base_ref
 {
-// TODO:
-// static_assert(alignment>=alignof(T));
+    __spv_ptr_t<T,_restrict> ptr;
 
-    using spv_ptr_t = uint64_t;
-    spv_ptr_t ptr;
-
-    __spv_ptr_t<T> __get_spv_ptr()
+    void __init(const spirv::bda_pointer_t<T> _ptr)
     {
-        return spirv::bitcast < __spv_ptr_t<T> > (ptr);
+        ptr.value = _ptr;
     }
-
-    // TODO: Would like to use `spv_ptr_t` or OpAccessChain result instead of `uint64_t`
-    void __init(const spv_ptr_t _ptr)
+    
+    spirv::bda_pointer_t<T> __get_spv_ptr()
     {
-        ptr = _ptr;
+        // BUG: https://github.com/microsoft/DirectXShaderCompiler/issues/7184
+        // if I don't launder the pointer through this I get "IsNonPtrAccessChain(ptrInst->opcode())" 
+        return spirv::copyObject<spirv::bda_pointer_t<T> >(ptr.value);
     }
 
     T load()
@@ -50,16 +55,13 @@ struct __base_ref
     }
 };
 
-template<typename T, uint32_t alignment=alignment_of_v<T>, bool _restrict = false>
+// TODO: I wish HLSL had some things like C++ which would allow you to make a "stack only"/non-storable type
+// NOTE: I guess there's the Function/Private storage space variables?
+template<typename T, uint32_t alignment=alignment_of_v<T>, bool _restrict=false>
 struct __ref : __base_ref<T,alignment,_restrict>
 {
-    using base_t = __base_ref < T, alignment, _restrict>;
-    using this_t = __ref < T, alignment, _restrict>;
-
-    __spv_ptr_t<T> get_ptr()
-    {
-        return base_t::__get_spv_ptr();
-    }
+    using base_t = __base_ref< T,alignment,_restrict>;
+    using this_t = __ref<T,alignment,_restrict>;
 };
 }
 }

include/nbl/builtin/hlsl/bda/__ref.hlsl

@@ -66,15 +66,15 @@ struct BdaAccessor : impl::BdaAccessorBase
     atomicAdd(const uint64_t index, const T value)
     {
         bda::__ptr<T> target = ptr + index;
-        return glsl::atomicAdd(target.template deref().get_ptr(), value);
+        return glsl::atomicAdd(target.template deref().ptr.value, value);
     }
 
     template<typename S = T>
     enable_if_t<is_same_v<S,T> && is_integral<T>::value && (sizeof(T) == 4 || sizeof(T) == 8), T>
     atomicSub(const uint64_t index, const T value)
     {
         bda::__ptr<T> target = ptr + index;
-        return glsl::atomicSub(target.template deref().get_ptr(), value);
+        return glsl::atomicSub(target.template deref().ptr.value, value);
     }
 
     bda::__ptr<T> ptr;

include/nbl/builtin/hlsl/bda/bda_accessor.hlsl

@@ -0,0 +1,176 @@
+// Copyright (C) 2018-2024 - DevSH Graphics Programming Sp. z O.O.
+// This file is part of the "Nabla Engine".
+// For conditions of distribution and use, see copyright notice in nabla.h
+#ifndef _NBL_BUILTIN_HLSL_BDA_STRUCT_DECLARE_INCLUDED_
+#define _NBL_BUILTIN_HLSL_BDA_STRUCT_DECLARE_INCLUDED_
+
+#include "nbl/builtin/hlsl/type_traits.hlsl"
+#include "nbl/builtin/hlsl/mpl.hlsl"
+#ifdef  __HLSL_VERSION
+#include "nbl/builtin/hlsl/bda/__ptr.hlsl"
+#endif //  __HLSL_VERSION
+
+
+namespace nbl
+{
+namespace hlsl
+{
+namespace bda
+{
+// silly utility traits
+template<typename T>
+struct member_count
+{
+    NBL_CONSTEXPR_STATIC_INLINE uint32_t value = 0;
+};
+template<typename T>
+NBL_CONSTEXPR uint32_t member_count_v = member_count<T>::value;
+
+template<typename T, int32_t MemberIx>
+struct member_type;
+template<typename T, int32_t MemberIx>
+using member_type_t = typename member_type<T,MemberIx>::type;
+
+// default alignment is the alignment of the type
+template<typename T, int32_t MemberIx>
+struct member_alignment
+{
+    NBL_CONSTEXPR_STATIC_INLINE uint32_t value = alignment_of_v<member_type_t<T,MemberIx> >;
+};
+template<typename T, int32_t MemberIx>
+NBL_CONSTEXPR uint32_t member_alignment_v = member_alignment<T,MemberIx>::value;
+
+// the default specialization of the offset assumes scalar layout
+template<typename T, int32_t MemberIx>
+struct member_offset
+{
+    // TODO: assert that the custom alignment is no less than the type's natural alignment?
+    // first byte past previous member, rounded up to out alignment
+    NBL_CONSTEXPR_STATIC_INLINE uint64_t value = mpl::align_up_v<member_offset<T,MemberIx-1>::value+size_of_v<member_type_t<T,MemberIx-1> >,member_alignment_v<T,MemberIx> >;
+};
+template<typename T>
+struct member_offset<T,0>
+{
+    NBL_CONSTEXPR_STATIC_INLINE uint64_t value = 0;
+};
+template<typename T, int32_t MemberIx>
+NBL_CONSTEXPR uint64_t member_offset_v = member_offset<T,MemberIx>::value;
+
+// stuff needed to compute alignment of the struct properly
+namespace impl
+{
+template<typename T, uint32_t N>
+struct default_alignment
+{
+    NBL_CONSTEXPR_STATIC_INLINE uint32_t value = mpl::max_v<uint32_t,member_alignment_v<T,N-1>,default_alignment<T,N-1>::value>;
+};
+// le invalid values
+template<typename T>
+struct default_alignment<T,0>
+{
+    NBL_CONSTEXPR_STATIC_INLINE uint32_t value = 0;
+};
+template<typename T, typename MemberCount=member_count<T> >
+NBL_CONSTEXPR uint32_t default_alignment_v = default_alignment<T,MemberCount::value>::value;
+}
+}
+}
+}
+
+//! Need to gen identical struct in HLSL and C++, right now this tool can declare non-templated structs and full explicit specialized ones
+
+//implementation details
+#define NBL_HLSL_IMPL_DEFINE_STRUCT_GET_MEMBER_TYPE(identifier,...) __VA_ARGS__
+#define NBL_HLSL_IMPL_DEFINE_STRUCT_GET_MEMBER_NAME(identifier,...) identifier
+#define NBL_HLSL_IMPL_DEFINE_STRUCT_MEMBER_TYPE(r,IDENTIFIER,i,e) template<> \
+struct ::nbl::hlsl::bda::member_type<NBL_EVAL IDENTIFIER,i> \
+{ \
+using type = NBL_HLSL_IMPL_DEFINE_STRUCT_GET_MEMBER_TYPE e; \
+};
+
+//! TODO: handle declarations for partial template specializations and non-specializations
+#define NBL_HLSL_IMPL_DECLARE_STRUCT_MEMBER(identifier,...) __VA_ARGS__ identifier;
+#ifdef __HLSL_VERSION
+#define NBL_HLSL_IMPL_DEFINE_STRUCT_MEMBER(r,IDENTIFIER,i,e) [[vk::ext_decorate(spv::DecorationOffset,::nbl::hlsl::bda::member_offset_v<NBL_EVAL IDENTIFIER,i>)]] NBL_HLSL_IMPL_DECLARE_STRUCT_MEMBER e
+#define NBL_HLSL_IMPL_DEFINE_STRUCT_MEMBER_REFERENCE(r,unused,i,e) ::nbl::hlsl::bda::__ref< \
+	NBL_HLSL_IMPL_DEFINE_STRUCT_GET_MEMBER_TYPE e, \
+	::nbl::hlsl::mpl::min_v<uint32_t,::nbl::hlsl::bda::member_alignment_v<__referenced_t,i>,alignment>, \
+_restrict> NBL_HLSL_IMPL_DEFINE_STRUCT_GET_MEMBER_NAME e;
+#define NBL_HLSL_IMPL_INIT_STRUCT_MEMBER_REFERENCE(r,unused,i,e) NBL_HLSL_IMPL_DEFINE_STRUCT_GET_MEMBER_NAME e .__init( \
+	::nbl::hlsl::spirv::accessChain<NBL_HLSL_IMPL_DEFINE_STRUCT_GET_MEMBER_TYPE e>(base_t::ptr.value,i) \
+);
+#define NBL_HLSL_IMPL_DEFINE_STRUCT(IDENTIFIER,MEMBER_SEQ) NBL_EVAL IDENTIFIER \
+{ \
+BOOST_PP_SEQ_FOR_EACH_I(NBL_HLSL_IMPL_DEFINE_STRUCT_MEMBER,IDENTIFIER,MEMBER_SEQ) \
+}; \
+template<uint32_t alignment, bool _restrict> \
+struct ::nbl::hlsl::bda::__ref<NBL_EVAL IDENTIFIER,alignment,_restrict> : ::nbl::hlsl::bda::__base_ref<NBL_EVAL IDENTIFIER,alignment,_restrict> \
+{ \
+	using __referenced_t = NBL_EVAL IDENTIFIER; \
+    using base_t = __base_ref<__referenced_t,alignment,_restrict>; \
+    using this_t = __ref<__referenced_t,alignment,_restrict>; \
+\
+	BOOST_PP_SEQ_FOR_EACH_I(NBL_HLSL_IMPL_DEFINE_STRUCT_MEMBER_REFERENCE,dummy,MEMBER_SEQ) \
+\
+	void __init(const ::nbl::hlsl::spirv::bda_pointer_t<__referenced_t> _ptr) \
+	{ \
+		base_t::__init(_ptr); \
+		BOOST_PP_SEQ_FOR_EACH_I(NBL_HLSL_IMPL_INIT_STRUCT_MEMBER_REFERENCE,dummy,MEMBER_SEQ) \
+	} \
+}
+#else
+#define NBL_HLSL_IMPL_DEFINE_STRUCT_MEMBER(r,IDENTIFIER,i,e) alignas(::nbl::hlsl::bda::member_alignment_v<NBL_EVAL IDENTIFIER,i>) NBL_HLSL_IMPL_DECLARE_STRUCT_MEMBER e
+#define NBL_HLSL_IMPL_DEFINE_STRUCT(IDENTIFIER,MEMBER_SEQ) alignas(::nbl::hlsl::alignment_of_v<NBL_EVAL IDENTIFIER >) NBL_EVAL IDENTIFIER \
+{ \
+BOOST_PP_SEQ_FOR_EACH_I(NBL_HLSL_IMPL_DEFINE_STRUCT_MEMBER,IDENTIFIER,MEMBER_SEQ) \
+}
+#endif
+
+// some weird stuff to handle alignment
+#define NBL_HLSL_IMPL_DEFINE_STRUCT_BEGIN(IDENTIFIER,MEMBER_SEQ) template<> \
+struct ::nbl::hlsl::bda::member_count<NBL_EVAL IDENTIFIER > \
+{ \
+	NBL_CONSTEXPR_STATIC_INLINE uint32_t value = BOOST_PP_SEQ_SIZE(MEMBER_SEQ); \
+}; \
+BOOST_PP_SEQ_FOR_EACH_I(NBL_HLSL_IMPL_DEFINE_STRUCT_MEMBER_TYPE,IDENTIFIER,MEMBER_SEQ) \
+template <> \
+struct ::nbl::hlsl::alignment_of<NBL_EVAL IDENTIFIER > \
+{
+#define NBL_HLSL_IMPL_DEFINE_STRUCT_END(IDENTIFIER,MEMBER_SEQ,...) }; \
+template<> \
+struct ::nbl::hlsl::size_of<NBL_EVAL IDENTIFIER > \
+{ \
+	using type = NBL_EVAL IDENTIFIER; \
+	NBL_CONSTEXPR_STATIC_INLINE uint32_t __last_member_ix_v = ::nbl::hlsl::bda::member_count_v<type>-1; \
+	NBL_CONSTEXPR_STATIC_INLINE uint64_t __last_member_offset_v = ::nbl::hlsl::bda::member_offset_v<type, __last_member_ix_v>; \
+	NBL_CONSTEXPR_STATIC_INLINE uint64_t __last_member_size_v = ::nbl::hlsl::size_of_v<::nbl::hlsl::bda::member_type_t<type, __last_member_ix_v> >; \
+	NBL_CONSTEXPR_STATIC_INLINE uint32_t value = ::nbl::hlsl::mpl::align_up_v<__last_member_offset_v + __last_member_size_v, alignment_of_v<type > >; \
+\
+__VA_ARGS__ \
+\
+}; \
+struct NBL_HLSL_IMPL_DEFINE_STRUCT(IDENTIFIER,MEMBER_SEQ)
+
+#include <boost/preprocessor/seq/for_each_i.hpp>
+#include <boost/preprocessor/seq/size.hpp>
+// MEMBER_SEQ is to be a sequence of variable name and type (identifier0,Type0)...(identifierN,TypeN) @see NBL_HLSL_IMPL_DEFINE_STRUCT_GET_MEMBER_TYPE
+// the VA_ARGS is the struct alignment for alignas, usage example
+// ```
+// NBL_HLSL_DEFINE_STRUCT((MyStruct2),
+//	((a, float32_t))
+//	((b, int32_t))
+//	((c, int32_t2)),
+// 
+// ... block of code for the methods ...
+// 
+// );
+// ```
+#define NBL_HLSL_DEFINE_STRUCT(IDENTIFIER,MEMBER_SEQ,...) NBL_HLSL_IMPL_DEFINE_STRUCT_BEGIN(IDENTIFIER,MEMBER_SEQ) \
+	NBL_CONSTEXPR_STATIC_INLINE uint32_t value = ::nbl::hlsl::bda::impl::default_alignment_v<NBL_EVAL IDENTIFIER >; \
+NBL_HLSL_IMPL_DEFINE_STRUCT_END(IDENTIFIER,MEMBER_SEQ,__VA_ARGS__)
+// version allowing custom alignment on whole struct
+#define NBL_HLSL_DEFINE_ALIGNAS_STRUCT(IDENTIFIER,ALIGNMENT,MEMBER_SEQ,...) NBL_HLSL_IMPL_DEFINE_STRUCT_BEGIN(IDENTIFIER,MEMBER_SEQ) \
+	NBL_CONSTEXPR_STATIC_INLINE uint32_t value = ALIGNMENT; \
+NBL_HLSL_IMPL_DEFINE_STRUCT_END(IDENTIFIER,MEMBER_SEQ,__VA_ARGS__)
+
+#endif

include/nbl/builtin/hlsl/bda/struct_declare.hlsl

@@ -52,6 +52,7 @@ namespace nbl
 namespace hlsl
 {
 
+// TODO: make this BDA compatible (no unspecialized templates yet)
 template<typename Scalar>
 struct complex_t
 {