Bring fuzzy matching support into master #5508
Conversation
Signed-off-by: Nathan Hjelm <[email protected]>
This commit updates the new custom matching code in pml/ob1 so it can not be enabled with a configure option. This commit also renames the fuzzy-matching headers to avoid potential name conflicts and removes the use of C reserved identifiers. Signed-off-by: Nathan Hjelm <[email protected]>
|
|
||
| typedef struct custom_match_prq_node | ||
| { | ||
| int32_t tags[PRQ_SIZE]; |
There was a problem hiding this comment.
These fields seems to always be accessed together, but laying out the structure this way guarantee 5 cache misses per access which is rather expensive for such a time critical operation.
There was a problem hiding this comment.
I agree that restructuring would be pertinent to allow for performant PRQ_SIZE and UMQ_SIZE resizing. The current implementation of 'pml_ob1_custom_match_arrays.h' matching engine structure is explicitly sized to fit each prq/umq node into a single cache line.
|
|
||
| typedef struct custom_match_umq_node | ||
| { | ||
| int32_t tags[UMQ_SIZE]; |
There was a problem hiding this comment.
same comment as for the prq struct.
| result = _mm512_cmpeq_epi8_mask(_mm512_and_epi32(elem->keys, elem->mask), _mm512_and_epi32(search, elem->mask)); | ||
| if(result) | ||
| { | ||
| for(i = elem->start; i <= elem->end; i++) |
There was a problem hiding this comment.
I would think that looping around the set bits in result will be faster as it saves few branches.
| { | ||
| for(i = elem->start; i <= elem->end; i++) | ||
| { | ||
| if((0x1 << i & result) && elem->value[i]) |
There was a problem hiding this comment.
is 0x1 really promoted to __mmask64 ?
Signed-off-by: Nathan Hjelm <[email protected]>
|
I have error trying to compile with new matching on Intel Xeon processor with AVX and AVX2 instruction set. GCC 7.1. Is there something wrong on the configure or am I missing something? Same error with --with-pml-ob1-matching=vector or fuzzy-*. |
|
@thananon The vector code here is writen for AVX-512 and is not currently compatible with AVX or AVX2 (or the limited AVX-512 implementation on Knights Corner). |
|
Ah, I see. Thank you. |
This PR brings the fuzzy matching support developed by @mdosanjh at Sandia into pml/ob1 on master.
The fuzzy matching code is disabled by default and can be enabled on the appropriate platforms by specifying the --with-pml-ob1-matching configure option.