Implements `top_k` in dpctl.tensor#1921

Merged

ndgrigorian merged 30 commits into

masterfrom

feature/topk

Jan 5, 2025

Commits on Jan 5, 2025

Implements `top_k`
ndgrigorian
committed
Add `top_k` to doc sources
ndgrigorian
committed
Use `std::size_t` for `k` instead of `py::ssize_t`
ndgrigorian
committed
Add implementation of `top_k` using radix sort
ndgrigorian
committed
Reduce code duplication by using std::optional in py_topk directly
ndgrigorian
committed
Clean up top_k docstring
ndgrigorian
committed
Move top_k into _sorting.py
ndgrigorian
committed
top_k raises when k > 1 and input is a scalar
ndgrigorian
committed
Fix bug in top_k partial merge sort implementation
ndgrigorian
committed
Add test file for top_k functionality

oleksandr-pavlyk
authored and
ndgrigorian
committed
Add sort_utils.hpp with iota_impl

oleksandr-pavlyk
authored and
ndgrigorian
committed
Simplify write-out kernels in topk implementation (avoid recomputing gid)

oleksandr-pavlyk
authored and
ndgrigorian
committed
Add explicit read into registers

oleksandr-pavlyk
authored and
ndgrigorian
committed
Use unique_ptr as temporary owner of USM allocation

oleksandr-pavlyk
authored and
ndgrigorian
committed
Factor out write-out kernel into separate detail function

oleksandr-pavlyk
authored and
ndgrigorian
committed
Typo: smart_malloc_jost -> smart_malloc_host

oleksandr-pavlyk
authored and
ndgrigorian
committed
Reimplemented map_back_impl to process few elements per work-item

oleksandr-pavlyk
authored and
ndgrigorian
committed
Replace use of std::log2(size_t_value)

oleksandr-pavlyk
authored and
ndgrigorian
committed
Remove dead code disable via preprocessor conditional

oleksandr-pavlyk
authored and
ndgrigorian
committed
Add information displayed on failure, renamed variables

oleksandr-pavlyk
authored and
ndgrigorian
committed
Add comment to explain ceil_log2 algo

oleksandr-pavlyk
authored and
ndgrigorian
committed
Add static assers to async_smart_free

oleksandr-pavlyk
authored and
ndgrigorian
committed
Skip top_k test known to fail for some CPU architectures
ndgrigorian
committed
Fixed blunder in work-item id to data_id computation

oleksandr-pavlyk
authored and
ndgrigorian
committed
Replace map_back_impl in sort_utils

oleksandr-pavlyk
authored and
ndgrigorian
committed
Use get_global_linear_id instead of get_global_id and rely on implicit conversion

oleksandr-pavlyk
authored and
ndgrigorian
committed
Counters in one-workgroup kernel to use uint16_t from uint32_t

oleksandr-pavlyk
authored and
ndgrigorian
committed
Apply work-around for failing tests with CPU device and short sub-groups

oleksandr-pavlyk
authored and
ndgrigorian
committed
Remove skipping of tests for i1/i2 dtypes since work-around

oleksandr-pavlyk
authored and
ndgrigorian
committed
Add entry to changelog for top_k
ndgrigorian
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implements `top_k` in dpctl.tensor#1921

Implements `top_k` in dpctl.tensor#1921
ndgrigorian merged 30 commits into
masterfrom
feature/topk

Commits on Jan 5, 2025

Implements `top_k`

Add `top_k` to doc sources

Use `std::size_t` for `k` instead of `py::ssize_t`

Add implementation of `top_k` using radix sort

Reduce code duplication by using std::optional in py_topk directly

Clean up top_k docstring

Move top_k into _sorting.py

top_k raises when k > 1 and input is a scalar

Fix bug in top_k partial merge sort implementation

Add test file for top_k functionality

Add sort_utils.hpp with iota_impl

Simplify write-out kernels in topk implementation (avoid recomputing gid)

Add explicit read into registers

Use unique_ptr as temporary owner of USM allocation

Factor out write-out kernel into separate detail function

Typo: smart_malloc_jost -> smart_malloc_host

Reimplemented map_back_impl to process few elements per work-item

Replace use of std::log2(size_t_value)

Remove dead code disable via preprocessor conditional

Add information displayed on failure, renamed variables

Add comment to explain ceil_log2 algo

Add static assers to async_smart_free

Skip top_k test known to fail for some CPU architectures

Fixed blunder in work-item id to data_id computation

Replace map_back_impl in sort_utils

Use get_global_linear_id instead of get_global_id and rely on implicit conversion

Counters in one-workgroup kernel to use uint16_t from uint32_t

Apply work-around for failing tests with CPU device and short sub-groups

Remove skipping of tests for i1/i2 dtypes since work-around

Add entry to changelog for top_k