GPU transforms.

CUDA transforms working on GPU memory, without transfers, but NOT thread-safe. More...

Functions
void	cushtns_profiling (shtns_cfg, int on)
double	cushtns_profiling_read_time (shtns_cfg, double *time_1, double *time_2)
const char *	cushtns_get_cfg_info (shtns_cfg)

Transforms
void	cu_spat_to_SH (shtns_cfg shtns, double *Vr, cplx *Qlm, int ltr)
	Same as spat_to_SH, but working on data residing on the GPU.
void	cu_spat_to_SH_float (shtns_cfg shtns, float *Vr, cplx_f *Qlm, int ltr)
	Same as cu_spat_to_SH, but working on single precision data.
void	cu_SH_to_spat (shtns_cfg shtns, cplx *Qlm, double *Vr, int ltr)
	Same as SH_to_spat, but working on data residing on the GPU.
void	cu_SH_to_spat_float (shtns_cfg shtns, cplx_f *Qlm, float *Vr, int ltr)
	Same as cu_SH_to_spat, but working on single precision data.
void	cu_spat_to_SHsphtor (shtns_cfg, double *Vt, double *Vp, cplx *Slm, cplx *Tlm, int ltr)
	Same as spat_to_SHsphtor, but working on data residing on the GPU.
void	cu_spat_to_SHsphtor_float (shtns_cfg, float *Vt, float *Vp, cplx_f *Slm, cplx_f *Tlm, int ltr)
void	cu_SHsphtor_to_spat (shtns_cfg, cplx *Slm, cplx *Tlm, double *Vt, double *Vp, int ltr)
	Same as SHsphtor_to_spat, but working on data residing on the GPU.
void	cu_SHsphtor_to_spat_float (shtns_cfg, cplx_f *Slm, cplx_f *Tlm, float *Vt, float *Vp, int ltr)
void	cu_SHsph_to_spat (shtns_cfg, cplx *Slm, double *Vt, double *Vp, int ltr)
	Same as SHsph_to_spat, but working on data residing on the GPU.
void	cu_SHsph_to_spat_float (shtns_cfg, cplx_f *Slm, float *Vt, float *Vp, int ltr)
void	cu_SHtor_to_spat (shtns_cfg, cplx *Tlm, double *Vt, double *Vp, int ltr)
	Same as SHtor_to_spat, but working on data residing on the GPU.
void	cu_SHtor_to_spat_float (shtns_cfg, cplx_f *Tlm, float *Vt, float *Vp, int ltr)
void	cu_spat_to_SHqst (shtns_cfg, double *Vr, double *Vt, double *Vp, cplx *Qlm, cplx *Slm, cplx *Tlm, int ltr)
	Same as spat_to_SHqst, but working on data residing on the GPU.
void	cu_spat_to_SHqst_float (shtns_cfg, float *Vr, float *Vt, float *Vp, cplx_f *Qlm, cplx_f *Slm, cplx_f *Tlm, int ltr)
void	cu_SHqst_to_spat (shtns_cfg, cplx *Qlm, cplx *Slm, cplx *Tlm, double *Vr, double *Vt, double *Vp, int ltr)
	Same as SHqst_to_spat, but working on data residing on the GPU.
void	cu_SHqst_to_spat_float (shtns_cfg, cplx_f *Qlm, cplx_f *Slm, cplx_f *Tlm, float *Vr, float *Vt, float *Vp, int ltr)

Initialization
int	cushtns_init_gpu (shtns_cfg shtns)
	Initialize given config to work on the current (or default) GPU, allowing to call GPU transforms cu_* above, working on data residing in the memory of this GPU.
shtns_cfg	cushtns_clone (shtns_cfg shtns, shtns_gpu_stream_t compute_stream, shtns_gpu_stream_t transfer_stream)
	Clone a gpu-enabled shtns config, and assign it to different streams (to allow compute overlap and/or usage from multiple threads).
void	cushtns_set_streams (shtns_cfg shtns, shtns_gpu_stream_t compute_stream, shtns_gpu_stream_t transfer_stream)
	Set user-specified streams for compute (including fft) and transfer.
void	cushtns_release_gpu (shtns_cfg)
	Release resources needed for GPU transforms, which won't work after this call.

Detailed Description

CUDA transforms working on GPU memory, without transfers, but NOT thread-safe.

Warning

These transforms are NOT thread-safe. Use one distinct shtns_cfg per thread. Use cushtns_clone to clone them. The transforms are Non-blocking, working with their own streams. Each clone has distinct streams.

See also

Using SHTns with GPU (nvida and AMD)

Function Documentation

cushtns_clone()

shtns_cfg cushtns_clone	(	shtns_cfg	shtns,
		shtns_gpu_stream_t	compute_stream,
		shtns_gpu_stream_t	transfer_stream )

Clone a gpu-enabled shtns config, and assign it to different streams (to allow compute overlap and/or usage from multiple threads).

This implies allocation of memory on the GPU and other limited resources, and thus may fail.

Parameters

[in]	shtns	is a valid shtns configuration created with shtns_create and with an associated grid (see shtns_set_grid_auto )
[in]	compute_stream	is a cuda Stream that will be used for transforms. If 0, the default (0) stream will be used.
[in]	transfer_stream	is a cuda Stream that will be used for data transfers between host and device for auto-offload mode. If 0, a new stream will be created and used.

Returns

a new shtns_cfg that can safely be used concurrently with the original one.

cushtns_init_gpu()

int cushtns_init_gpu

(

shtns_cfg

shtns

)

Initialize given config to work on the current (or default) GPU, allowing to call GPU transforms cu_* above, working on data residing in the memory of this GPU.

This does not enable auto-offload. Use cudaSetDevice() or hipSetDevice() to set the target GPU before calling this function. Note that it is the user's responsibility to ensure the current device will be the same for subsequent calls to transform functions with this configuration.

Parameters

[in]

shtns

is a valid shtns configuration created with shtns_create and with an associated grid (see shtns_set_grid_auto )

Returns

device_id on success, or -1 on failure.