Work buffer size too big for the rocfft backend

[tpadioleau]

I think we allocate too much memory in the rocfft backend. The function rocfft_plan_get_work_buffer_size https://rocm.docs.amd.com/projects/rocFFT/en/latest/reference/api.html#_CPPv432rocfft_plan_get_work_buffer_sizeK11rocfft_planP6size_t returns a size in bytes.

If I am not mistaken we allocate a buffer of type Kokkos::complex<float_type*> in

kokkos-fft/fft/src/KokkosFFT_ROCM_plans.hpp

Lines 184 to 195 in 95bd3b3

	std::size_t workbuffersize = 0;
	status = rocfft_plan_get_work_buffer_size(*plan, &workbuffersize);
	KOKKOSFFT_THROW_IF(status != rocfft_status_success,
	"rocfft_plan_get_work_buffer_size failed");
	if (workbuffersize > 0) {
	buffer = BufferViewType("work_buffer", workbuffersize);
	status = rocfft_execution_info_set_work_buffer(
	execution_info, (void*)buffer.data(), workbuffersize);
	KOKKOSFFT_THROW_IF(status != rocfft_status_success,
	"rocfft_execution_info_set_work_buffer failed");
	}

because BufferViewType is given by

kokkos-fft/fft/src/KokkosFFT_Plans.hpp

Lines 101 to 102 in 95bd3b3

	using BufferViewType =
	Kokkos::View<Kokkos::complex<float_type>*, layout_type, execSpace>;