LTO Examples

The cuFFTDx provides multiple samples which demonstrate how to use the features enabled by LTO.

Group	Subgroup	Example	Description
Introduction Examples	09_introduction_lto_example	introduction_lto_example	(offline) cuFFTDx LTO introduction
	04_nvrtc_fft	nvrtc_fft_block_lto	(online) cuFFTDx LTO introduction
	10_cufft_device_api_example	cufft_device_api_example	(offline) cuFFT Device API introduction
Simple FFT Examples	01_simple_fft_thread	simple_fft_thread_lto	(offline) Complex-to-complex (C2C) thread FFT using LTO
	02_simple_fft_block	simple_fft_block_c2r_lto	(offline) Complex-to-real block FFT using LTO
NVRTC Examples	04_nvrtc_fft	nvrtc_fft_thread_lto	(online) Complex-to-complex thread FFT using LTO
		nvrtc_fft_block_lto	(online) Complex-to-complex block FFT using LTO
FFT Performance	03_block_fft_performance	block_fft_lto_ptx_performance	(offline) Benchmark for C2C block FFT (LTO vs PTX)
Runtime Database Query	09_introduction_lto_example	lto_query_database_example	(offline) List all matching implementations from an offline LTO database
	04_nvrtc_fft	nvrtc_query_database_fft_block_lto	(offline) C2C block FFT using offline LTO database to obtain runtime traits
	04_nvrtc_fft	nvrtc_query_autotune_lto	(online) Autotune across all online LTO implementations via runtime query

For more information about “online” vs “offline” annotated in the table above, see Augment cuFFTDx with LTO.

Introduction Examples

• introduction_lto_example

• nvrtc_fft_block_lto

• cufft_device_api_example

Examples used in the documentation demonstrating how to adopt LTO features for existing cuFFTDx projects. introduction_lto_example and nvrtc_fft_block_lto are used in the Augment cuFFTDx with LTO section to showcase the steps for offline and online kernel generation, respectively. In cufft_device_api_example, used in the Custom LTO Helper section, we demonstrate how to write a custom LTO database helper tool using the cuFFT Device API, as an alternative to the LTO Helper.

Simple FFT Examples

• simple_fft_thread_lto

• simple_fft_block_c2r_lto

The LTO version of the simple_fft_thread and simple_fft_block_c2r samples, respectively. See Simple FFT Examples for more details.

NVRTC Examples

• nvrtc_fft_thread_lto

• nvrtc_fft_block_lto

The LTO versions of the existing NVRTC examples presenting how to use cuFFTDx on thread and block level with NVRTC and nvJitLink. See NVRTC Examples for more details.

FFT Performance

• block_fft_lto_ptx_performance

The example compares the performance of cuFFTDx device functions calculating FFT using the inlined-PTX implementation vs. the LTOIR implementation.

Fig. 7 Performance comparison between LTOIR and PTX implementations of a single-precision complex-to-complex forward FFT. Tests were performed on H100 80GB with maximum clocks set.

Runtime Database Query Examples

• lto_query_database_example

• nvrtc_query_database_fft_block_lto

• nvrtc_query_autotune_lto

The samples listed above demonstrate how to use the Runtime Database Query API together with LTO. They cover the two LTO query modes (cufftdx::experimental::query_code_type::ltoir_offline and cufftdx::experimental::query_code_type::ltoir_online) introduced in cuFFTDx 1.7.1. As noted in Runtime Database Query, the LTO query modes require cuFFT 12.3 or later (from CUDA Toolkit 13.3).

All three examples have #define CUFFTDX_ENABLE_RUNTIME_DATABASE before including the cuFFTDx headers, since the query_database and get_all_implementations APIs are gated behind this macro. The experimental::query_code_type::ltoir_online mode (used by nvrtc_query_autotune_lto) additionally requires CUFFTDX_ENABLE_CUFFT_DEPENDENCY to be defined and the application to be linked against cuFFT.

lto_query_database_example is a CUDA C++ example that includes an offline LTO database header (generated ahead of time by the LTO helper) and calls cufftdx::experimental::utils::get_all_implementations(...) with experimental::query_code_type::ltoir_offline to enumerate every implementation that matches a given FFT configuration in the offline LTO database.

nvrtc_query_database_fft_block_lto is the LTO counterpart of the non-LTO nvrtc_query_database_fft_block example (see NVRTC Examples). It uses cufftdx::experimental::utils::query_database(...) with experimental::query_code_type::ltoir_offline to look up the optimal implementation for a given FFT inside a pre-generated offline LTO database, then NVRTC-compiles a kernel that links against the matching pre-generated LTOIR artifacts via nvJitLink.

nvrtc_query_autotune_lto shows how to autotune across every online LTO implementation that matches a given FFT configuration. It calls cufftdx::experimental::utils::get_all_implementations(...) with experimental::query_code_type::ltoir_online (which delegates to query_all_cufft_implementations) to enumerate the candidates, benchmarks each one, and then calls cufftdx::utils::get_database_and_ltoir() to retrieve the database string and LTOIRs for the selected configuration.