- Differentiate concepts for FFT based on ElementsPerInvocationLog2,

  this allows the offset accessor passed in the bigger FFTs to pass the
  concept requirements
- Minor change to the offset accessor, I think it follows the original
  intent

Author: Fletterio

include/nbl/builtin/hlsl/concepts/accessors/fft.hlsl

@@ -21,23 +21,45 @@ namespace fft
 #define NBL_CONCEPT_TPLT_PRM_KINDS (typename)
 #define NBL_CONCEPT_TPLT_PRM_NAMES (T)
 #define NBL_CONCEPT_PARAM_0 (accessor, T)
-#define NBL_CONCEPT_PARAM_1 (index_t, uint32_t)
-#define NBL_CONCEPT_PARAM_2 (value_t, uint32_t)
+#define NBL_CONCEPT_PARAM_1 (index, uint32_t)
+#define NBL_CONCEPT_PARAM_2 (val, uint32_t)
 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
+#define index NBL_CONCEPT_PARAM_T NBL_CONCEPT_PARAM_1
+#define val NBL_CONCEPT_PARAM_T NBL_CONCEPT_PARAM_2
 NBL_CONCEPT_END(
-    ((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.set(index, val)), is_same_v, void))
+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((accessor.get(index, val)), is_same_v, void))
     ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((accessor.workgroupExecutionAndMemoryBarrier()), is_same_v, void))
 );
-#undef value_t
-#undef index_t
+#undef val
+#undef index
 #undef accessor
 #include <nbl/builtin/hlsl/concepts/__end.hlsl>
 
 
+// The Accessor (for a small FFT) 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);
+
+#define NBL_CONCEPT_NAME SmallFFTAccessor
+#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, uint32_t)
+#define NBL_CONCEPT_PARAM_2 (val, complex_t<Scalar>)
+NBL_CONCEPT_BEGIN(3)
+#define accessor NBL_CONCEPT_PARAM_T NBL_CONCEPT_PARAM_0
+#define index NBL_CONCEPT_PARAM_T NBL_CONCEPT_PARAM_1
+#define val NBL_CONCEPT_PARAM_T NBL_CONCEPT_PARAM_2
+NBL_CONCEPT_END(
+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((accessor.set(index, val)), is_same_v, void))
+    ((NBL_CONCEPT_REQ_EXPR_RET_TYPE)((accessor.get(index, val)), is_same_v, void))
+);
+#undef val
+#undef index
+#undef accessor
+#include <nbl/builtin/hlsl/concepts/__end.hlsl>
 
 
 // The Accessor MUST provide the following methods:
@@ -49,19 +71,11 @@ NBL_CONCEPT_END(
 #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<Scalar>)
-NBL_CONCEPT_BEGIN(3)
+NBL_CONCEPT_BEGIN(1)
 #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_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))
-);
-#undef value_t
-#undef index_t
+) && SmallFFTAccessor<T, Scalar>;
 #undef accessor
 #include <nbl/builtin/hlsl/concepts/__end.hlsl>
 

include/nbl/builtin/hlsl/memory_accessor.hlsl

@@ -205,14 +205,15 @@ template<class BaseAccessor, typename IndexType=uint32_t, typename _Offset=void>
 struct Offset : impl::OffsetBase<IndexType,_Offset>
 {
     using base_t = impl::OffsetBase<IndexType,_Offset>;
+    using index_t = IndexType;
 
     BaseAccessor accessor;
 
     template <typename T>
-    void set(uint32_t idx, T value) {accessor.set(idx+base_t::offset,value); }
+    void set(index_t idx, T value) {accessor.set(idx+base_t::offset,value); }
 
     template <typename T> 
-    void get(uint32_t idx, NBL_REF_ARG(T) value) {accessor.get(idx+base_t::offset,value);}
+    void get(index_t idx, NBL_REF_ARG(T) value) {accessor.get(idx+base_t::offset,value);}
 
     template<typename S=BaseAccessor>
     enable_if_t<

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

@@ -274,19 +274,18 @@ template<bool Inverse, typename consteval_params_t, class device_capabilities=vo
 struct FFT;
 
 // For the FFT methods below, we assume:
-//      - 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).
+//      - Accessor is an accessor to an array fitting ElementsPerInvocation * WorkgroupSize elements of type complex_t<Scalar>, used to get inputs / set outputs of the FFT. 
+//        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.
 //        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.
+//        If using `ConstevalParameters::ElementsPerInvocationLog2 == 1` the Accessor IS ALLOWED TO not provide this last method.
+//        If 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`.
+//      - SharedMemoryAccessor accesses a workgroup-shared memory array of size `WorkgroupSize` elements of type complex_t<Scalar>.
 //        The SharedMemoryAccessor MUST provide the following methods:
 //             * void get(uint32_t index, inout uint32_t value);  
 //             * void set(uint32_t index, in uint32_t value); 
@@ -308,7 +307,7 @@ struct FFT<false, fft::ConstevalParameters<1, WorkgroupSizeLog2, Scalar>, device
     }
 
 
-    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::FFTAccessor<Accessor, Scalar> && fft::FFTSharedMemoryAccessor<SharedMemoryAccessor>)
+    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::SmallFFTAccessor<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;
@@ -374,7 +373,7 @@ struct FFT<true, fft::ConstevalParameters<1, WorkgroupSizeLog2, Scalar>, device_
     }
 
 
-    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::FFTAccessor<Accessor, Scalar> && fft::FFTSharedMemoryAccessor<SharedMemoryAccessor>)
+    template<typename Accessor, typename SharedMemoryAccessor NBL_FUNC_REQUIRES(fft::SmallFFTAccessor<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;