Enabling RNG based on oneMKL `generate_single`#4

Open

L30nardoSV wants to merge 29 commits intoemascarenhas:developfrom

LeoCollab:onemkl-rng-single

Commits on Oct 12, 2022

Preliminary creation of oneMKL RNG object in kernel4 (not using it yet).

Leonardo Solis
authored and
L30nardoSV
committed

Commits on Oct 14, 2022

Declaring two RNG distribution objects: discrete and continuous.
L30nardoSV
committed
Redefining rng-engine for 64 elements + passing rng engine and distrs to kernel code + calling rng <generate_single>.
L30nardoSV
committed
Enabling and formatting printfs.
L30nardoSV
committed
Enabling and formatting printfs.
L30nardoSV
committed
Enabling and formatting printfs.
L30nardoSV
committed
Enabling and formatting printfs.
L30nardoSV
committed
Enabling and formatting printfs.
L30nardoSV
committed
Adding local wi id.
L30nardoSV
committed
Removing temp. private array + writing generated rngs directly into shared array.
L30nardoSV
committed
Removing unused CUDA printfs + commenting out equivalent SYCL printfs.
L30nardoSV
committed
Replacing call to gpu_randf() with call to oneapi generate_single().
L30nardoSV
committed
Replacing gpu_randf() calls with oneapi generate_single() ones each.
L30nardoSV
committed
Fixing seed and offset of RNG engine. Otherwise, some resulting energies and overall clustering are wrong.
L30nardoSV
committed
Removing unused discrete RNG distribution object.
L30nardoSV
committed
Adding RNG engine and distribution to LS kernel (not yet calling RNG generatefrom kernel).
L30nardoSV
committed
In LS: replacing gpu_randf() calls with oneapi generate_single() ones each.
L30nardoSV
committed
Commenting out again DOCK_TRACE (as it was initially).
L30nardoSV
committed
In AD: replacing gpu_randf() call with oneapi generate_single() call.
L30nardoSV
committed
In ADAM: replacing gpu_randf() call with oneapi generate_single() call.
L30nardoSV
committed
Correcting RNG VecSize for philox from 64 (incorrect) to 16 (maximum possible).
L30nardoSV
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Enabling RNG based on oneMKL `generate_single`#4

Enabling RNG based on oneMKL `generate_single`#4
L30nardoSV wants to merge 29 commits intoemascarenhas:developfrom
LeoCollab:onemkl-rng-single

Commits on Oct 12, 2022

Preliminary creation of oneMKL RNG object in kernel4 (not using it yet).

Commits on Oct 14, 2022

Declaring two RNG distribution objects: discrete and continuous.

Redefining rng-engine for 64 elements + passing rng engine and distrs to kernel code + calling rng <generate_single>.

Enabling and formatting printfs.

Enabling and formatting printfs.

Enabling and formatting printfs.

Enabling and formatting printfs.

Enabling and formatting printfs.

Adding local wi id.

Removing temp. private array + writing generated rngs directly into shared array.

Removing unused CUDA printfs + commenting out equivalent SYCL printfs.

Replacing call to gpu_randf() with call to oneapi generate_single().

Replacing gpu_randf() calls with oneapi generate_single() ones each.

Fixing seed and offset of RNG engine. Otherwise, some resulting energies and overall clustering are wrong.

Removing unused discrete RNG distribution object.

Adding RNG engine and distribution to LS kernel (not yet calling RNG generatefrom kernel).

In LS: replacing gpu_randf() calls with oneapi generate_single() ones each.

Commenting out again DOCK_TRACE (as it was initially).

In AD: replacing gpu_randf() call with oneapi generate_single() call.

In ADAM: replacing gpu_randf() call with oneapi generate_single() call.

Correcting RNG VecSize for philox from 64 (incorrect) to 16 (maximum possible).

Commits on Oct 19, 2022

In Kernel4: adding preprocessor directives to select RNG engine.

Defining <RNG_ONEMKL_TYPE> for different engine types.

Adding macro <RNG_ORIGINAL> to enable original adhoc/manual RNG code.

In SW: adding macro <RNG_ORIGINAL> to enable original adhoc/manual RNG code.

In AD and ADAM: adding macro <RNG_ORIGINAL>.

Redefining macro name for oneMKL engine type.

Using <RNG_ONEMKL_DISTRIBUTION_TYPE> macro.

Minor comment reordering.