Fixed accessor concepts for FFT

Author: Fletterio

include/nbl/builtin/hlsl/workgroup/fft/accessor_concepts.hlsl renamed to include/nbl/builtin/hlsl/concepts/accessors/fft.hlsl

@@ -1,9 +1,8 @@
-#ifndef _NBL_BUILTIN_HLSL_WORKGROUP_FFT_ACCESSOR_CONCEPTS_INCLUDED_
-#define _NBL_BUILTIN_HLSL_WORKGROUP_FFT_ACCESSOR_CONCEPTS_INCLUDED_
+#ifndef _NBL_BUILTIN_HLSL_CONCEPTS_ACCESSORS_FFT_INCLUDED_
+#define _NBL_BUILTIN_HLSL_CONCEPTS_ACCESSORS_FFT_INCLUDED_
 
-#include <nbl/builtin/hlsl/cpp_compat.hlsl>
-#include <nbl/builtin/hlsl/concepts.hlsl>
-#include <nbl/builtin/hlsl/fft/common.hlsl>
+#include "nbl/builtin/hlsl/concepts.hlsl"
+#include "nbl/builtin/hlsl/fft/common.hlsl"
 
 namespace nbl
 {
@@ -40,24 +39,23 @@ NBL_CONCEPT_END(
 
 
 
-// The Accessor MUST provide a typename `Accessor::scalar_t`
+
 // The Accessor MUST provide the following methods:
 //     * void get(uint32_t index, inout complex_t<Scalar> value);
 //     * void set(uint32_t index, in complex_t<Scalar> value);
 //     * void memoryBarrier();
 
 #define NBL_CONCEPT_NAME FFTAccessor
-#define NBL_CONCEPT_TPLT_PRM_KINDS (typename)
-#define NBL_CONCEPT_TPLT_PRM_NAMES (T)
+#define NBL_CONCEPT_TPLT_PRM_KINDS (typename)(typename)
+#define NBL_CONCEPT_TPLT_PRM_NAMES (T)(Scalar)
 #define NBL_CONCEPT_PARAM_0 (accessor, T)
 #define NBL_CONCEPT_PARAM_1 (index_t, uint32_t)
-#define NBL_CONCEPT_PARAM_2 (value_t, complex_t<typename T::scalar_t>)
-NBL_CONCEPT_BEGIN(4)
+#define NBL_CONCEPT_PARAM_2 (value_t, complex_t<Scalar>)
+NBL_CONCEPT_BEGIN(3)
 #define accessor NBL_CONCEPT_PARAM_T NBL_CONCEPT_PARAM_0
 #define index_t NBL_CONCEPT_PARAM_T NBL_CONCEPT_PARAM_1
 #define value_t NBL_CONCEPT_PARAM_T NBL_CONCEPT_PARAM_2
 NBL_CONCEPT_END(
-    ((NBL_CONCEPT_REQ_TYPE)(T::scalar_t))
     ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((accessor.set(index_t, value_t)), is_same_v, void))
     ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((accessor.get(index_t, value_t)), is_same_v, void))
     ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((accessor.memoryBarrier()), is_same_v, void))

include/nbl/builtin/hlsl/workgroup/fft.hlsl

@@ -83,7 +83,7 @@ inline OptimalFFTParameters optimalFFTParameters(const uint32_t maxWorkgroupSize
 #include "nbl/builtin/hlsl/mpl.hlsl"
 #include "nbl/builtin/hlsl/memory_accessor.hlsl"
 #include "nbl/builtin/hlsl/bit.hlsl"
-#include "nbl/builtin/hlsl/workgroup/fft/accessor_concepts.hlsl"
+#include "nbl/builtin/hlsl/concepts/accessors/fft.hlsl"
 
 // Caveats
 // - Sin and Cos in HLSL take 32-bit floats. Using this library with 64-bit floats works perfectly fine, but DXC will emit warnings
@@ -274,14 +274,17 @@ template<bool Inverse, typename consteval_params_t, class device_capabilities=vo
 struct FFT;
 
 // For the FFT methods below, we assume:
-//      - Accessor is a global memory accessor to an array fitting 2 * WorkgroupSize elements of type complex_t<Scalar>, used to get inputs / set outputs of the FFT,
-//        that is, one "lo" and one "hi" complex numbers per thread, essentially 4 Scalars per thread. The arrays it accesses with `get` and `set` can optionally be
-//        different, if you don't want the FFT to be done in-place. 
-//        The Accessor MUST provide a typename `Accessor::scalar_t`, and this type MUST be the same as the `Scalar` template parameter of the FFT struct's consteval parameters
+//      - Accessor is an accessor to an array fitting 2 * WorkgroupSize elements of type complex_t<Scalar>, used to get inputs / set outputs of the FFT,
+//        that is, one "lo" and one "hi" complex numbers per thread, essentially 4 Scalars per thread. If `ConstevalParameters::ElementsPerInvocationLog2 == 1`, 
+//        the arrays it accesses with `get` and `set` can optionally be different, if you don't want the FFT to be done in-place. Otherwise, you MUST make it in-place
+//        (this is because if using more than 2 elements per invocation, we use the same array to store intermediate operations).
 //        The Accessor MUST provide the following methods:
 //            * void get(uint32_t index, inout complex_t<Scalar> value);
 //            * void set(uint32_t index, in complex_t<Scalar> value);
 //            * void memoryBarrier();
+//        For it to work correctly, this memory barrier must use `AcquireRelease` semantics, with the proper flags set for the memory type.
+//        If using `ConstevalParameters::ElementsPerInvocationLog2 == 1` or otherwise not needing it (such as when using preloaded accessors) we still require the method to exist
+//        but you can just make it do nothing.
 
 //      - SharedMemoryAccessor accesses a workgroup-shared memory array of size `2 * sizeof(Scalar) * WorkgroupSize`.
 //        The SharedMemoryAccessor MUST provide the following methods:
@@ -305,7 +308,7 @@ struct FFT<false, fft::ConstevalParameters<1, WorkgroupSizeLog2, Scalar>, device
     }
 
 
-    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::FFTAccessor<Accessor> && fft::FFTSharedMemoryAccessor<SharedMemoryAccessor>)
+    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::FFTAccessor<Accessor, Scalar> && fft::FFTSharedMemoryAccessor<SharedMemoryAccessor>)
     static void __call(NBL_REF_ARG(Accessor) accessor, NBL_REF_ARG(SharedMemoryAccessor) sharedmemAccessor)
     {
         NBL_CONSTEXPR_STATIC_INLINE uint16_t WorkgroupSize = consteval_params_t::WorkgroupSize;
@@ -371,7 +374,7 @@ struct FFT<true, fft::ConstevalParameters<1, WorkgroupSizeLog2, Scalar>, device_
     }
 
 
-    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::FFTAccessor<Accessor> && fft::FFTSharedMemoryAccessor<SharedMemoryAccessor>)
+    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::FFTAccessor<Accessor, Scalar> && fft::FFTSharedMemoryAccessor<SharedMemoryAccessor>)
     static void __call(NBL_REF_ARG(Accessor) accessor, NBL_REF_ARG(SharedMemoryAccessor) sharedmemAccessor)
     {
         NBL_CONSTEXPR_STATIC_INLINE uint16_t WorkgroupSize = consteval_params_t::WorkgroupSize;
@@ -432,7 +435,7 @@ struct FFT<false, fft::ConstevalParameters<ElementsPerInvocationLog2, WorkgroupS
     using consteval_params_t = fft::ConstevalParameters<ElementsPerInvocationLog2, WorkgroupSizeLog2, Scalar>;
     using small_fft_consteval_params_t = fft::ConstevalParameters<1, WorkgroupSizeLog2, Scalar>;
 
-    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::FFTAccessor<Accessor> && fft::FFTSharedMemoryAccessor<SharedMemoryAccessor>)
+    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::FFTAccessor<Accessor, Scalar> && fft::FFTSharedMemoryAccessor<SharedMemoryAccessor>)
     static void __call(NBL_REF_ARG(Accessor) accessor, NBL_REF_ARG(SharedMemoryAccessor) sharedmemAccessor)
     {
         NBL_CONSTEXPR_STATIC_INLINE uint16_t WorkgroupSize = consteval_params_t::WorkgroupSize;
@@ -481,7 +484,7 @@ struct FFT<true, fft::ConstevalParameters<ElementsPerInvocationLog2, WorkgroupSi
     using consteval_params_t = fft::ConstevalParameters<ElementsPerInvocationLog2, WorkgroupSizeLog2, Scalar>;
     using small_fft_consteval_params_t = fft::ConstevalParameters<1, WorkgroupSizeLog2, Scalar>;
 
-    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::FFTAccessor<Accessor> && fft::FFTSharedMemoryAccessor<SharedMemoryAccessor>)
+    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::FFTAccessor<Accessor, Scalar> && fft::FFTSharedMemoryAccessor<SharedMemoryAccessor>)
     static void __call(NBL_REF_ARG(Accessor) accessor, NBL_REF_ARG(SharedMemoryAccessor) sharedmemAccessor)
     {
         NBL_CONSTEXPR_STATIC_INLINE uint16_t WorkgroupSize = consteval_params_t::WorkgroupSize;

src/nbl/builtin/CMakeLists.txt

@@ -310,7 +310,6 @@ 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/fft.hlsl")
-LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/workgroup/fft/accessor_concepts.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/shuffle.hlsl")
@@ -331,5 +330,7 @@ LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/concepts/accessors/anisotropi
 LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/concepts/accessors/loadable_image.hlsl")
 LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/concepts/accessors/mip_mapped.hlsl")
 LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/concepts/accessors/storable_image.hlsl")
+LIST_BUILTIN_RESOURCE(NBL_RESOURCES_TO_EMBED "hlsl/concepts/accessors/fft.hlsl")
+
 
 ADD_CUSTOM_BUILTIN_RESOURCES(nblBuiltinResourceData NBL_RESOURCES_TO_EMBED "${NBL_ROOT_PATH}/include" "nbl/builtin" "nbl::builtin" "${NBL_ROOT_PATH_BINARY}/include" "${NBL_ROOT_PATH_BINARY}/src" "STATIC" "INTERNAL")