Subgroup strided FFTs (#142)

* Subgroup strided FFTs

* tidying

* add more variety to the sizes chosen for testing

* update log line

* remove unneeded +1

* moved private memory definition to make AMD happy

* comment

* use initialised private array

* format

* add comment to explain AMD hack

Author: FMarno

README.md

@@ -96,7 +96,7 @@ portFFT is still in early development. The supported configurations are:
   * size in each dimension must be supported by 1D transforms
 * Arbitrary forward and backward scales
 * Arbitrary forward and backward offsets
-* Arbitrary strides and distance where the problem size + auxilary data fits in the registers of a single work-item.
+* Arbitrary strides and distance where the problem size + auxilary data fits in the registers of a single subgroup.
 
 Any 1D arbitrarily large input size that fits in global memory is supported, with a restriction that large input sizes should not have large prime factors.
 The largest prime factor depend on the device and the values set by `PORTFFT_REGISTERS_PER_WI` and `PORTFFT_SUBGROUP_SIZES`.

src/portfft/descriptor_validation.hpp

@@ -24,7 +24,7 @@
 #include <string_view>
 
 #include "common/exceptions.hpp"
-#include "common/workitem.hpp"
+#include "common/subgroup.hpp"
 #include "enums.hpp"
 #include "utils.hpp"
 
@@ -65,10 +65,17 @@ inline void validate_layout(const std::vector<std::size_t>& lengths, portfft::de
     }
   }
   if (forward_layout == portfft::detail::layout::UNPACKED || backward_layout == portfft::detail::layout::UNPACKED) {
-    if (!portfft::detail::fits_in_wi<Scalar>(lengths.back())) {
+    bool fits_subgroup = false;
+    for (auto sg_size : {PORTFFT_SUBGROUP_SIZES}) {
+      fits_subgroup =
+          fits_subgroup || portfft::detail::fits_in_sg<Scalar>(static_cast<IdxGlobal>(lengths.back()), sg_size);
+      if (fits_subgroup) {
+        break;
+      }
+    }
+    if (!fits_subgroup) {
       throw unsupported_configuration(
-          "Arbitrary strides and distances are only supported for sizes that fit in the registers of a single "
-          "work-item");
+          "Arbitrary strides and distances are only supported for sizes that fit in the registers of a subgroup");
     }
   }
 }