[libclc] Refine __clc_get_sub_group_size with fast full sub-group path #163149
Conversation
Add a fast path for the common case that total work-group size is multiple of max sub-group size, avoiding need to calculate number of sub-groups.
There was a problem hiding this comment.
Pull Request Overview
This PR optimizes the __clc_get_sub_group_size() function by adding a fast path for the common case where the total work-group size is a multiple of the maximum sub-group size. This optimization avoids the need to calculate the number of sub-groups in many cases.
Key changes:
- Replaced conditional logic with a more efficient calculation using modulo operation
- Simplified the function to use a single boolean condition to determine if a sub-group is full-sized
- Reduced the number of function calls and intermediate calculations
| uint linear_size = __clc_get_local_size(0) * __clc_get_local_size(1) * | ||
| __clc_get_local_size(2); |
There was a problem hiding this comment.
The function calls to __clc_get_local_size() are repeated three times. Consider caching these values in local variables to avoid potential multiple function call overhead.
| uint linear_size = __clc_get_local_size(0) * __clc_get_local_size(1) * | |
| __clc_get_local_size(2); | |
| uint local_size_0 = __clc_get_local_size(0); | |
| uint local_size_1 = __clc_get_local_size(1); | |
| uint local_size_2 = __clc_get_local_size(2); | |
| uint linear_size = local_size_0 * local_size_1 * local_size_2; |
Copilot uses AI. Check for mistakes.
| uint linear_size = __clc_get_local_size(0) * __clc_get_local_size(1) * | ||
| __clc_get_local_size(2); | ||
| uint remainder = linear_size % __clc_get_max_sub_group_size(); | ||
| bool full_sub_group = (remainder == 0) || (__clc_get_sub_group_id() < | ||
| __clc_get_num_sub_groups() - 1); | ||
|
|
||
| return full_sub_group ? __clc_get_max_sub_group_size() : remainder; |
There was a problem hiding this comment.
Type inconsistency: __clc_get_local_size() returns size_t but the result is assigned to uint. This could cause truncation on 64-bit systems where size_t is larger than uint. Consider using size_t for linear_size and remainder variables.
| uint linear_size = __clc_get_local_size(0) * __clc_get_local_size(1) * | |
| __clc_get_local_size(2); | |
| uint remainder = linear_size % __clc_get_max_sub_group_size(); | |
| bool full_sub_group = (remainder == 0) || (__clc_get_sub_group_id() < | |
| __clc_get_num_sub_groups() - 1); | |
| return full_sub_group ? __clc_get_max_sub_group_size() : remainder; | |
| size_t linear_size = __clc_get_local_size(0) * __clc_get_local_size(1) * | |
| __clc_get_local_size(2); | |
| size_t remainder = linear_size % __clc_get_max_sub_group_size(); | |
| bool full_sub_group = (remainder == 0) || (__clc_get_sub_group_id() < | |
| __clc_get_num_sub_groups() - 1); | |
| return (uint)(full_sub_group ? __clc_get_max_sub_group_size() : remainder); |
Copilot uses AI. Check for mistakes.
| return linear_size - uniform_size; | ||
| size_t linear_size = __clc_get_local_size(0) * __clc_get_local_size(1) * | ||
| __clc_get_local_size(2); | ||
| uint remainder = linear_size % __clc_get_max_sub_group_size(); |
There was a problem hiding this comment.
Not sure how this is faster? The old implementation carefully avoided division, and this introduces urem?
There was a problem hiding this comment.
Not sure how this is faster? The old implementation carefully avoided division, and this introduces urem?
in most cases __clc_get_max_sub_group_size() is a power of 2 and modulo has the same code as & (__clc_get_max_sub_group_size() - 1)
There was a problem hiding this comment.
Following is diff on nvptx64--nvidiacl.bc and diff on llc -march=nvptx64 output:
There are 3 improvements:
- return value range is tightened
- fast path for total work-group size being multiple of max sub-group size
- number of ptx register is reduced from 18 to 15
3 participants
Add a fast path for the common case that total work-group size is multiple of max sub-group size, avoiding need to calculate number of sub-groups.