[Hexagon] Add fix for vtcm allocation searches #13197
Conversation
|
Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment. Generated by tvm-bot |
nverke
force-pushed
the
vtcm_alloc_fix
branch
from
2f9b770 to
8991ca1
Compare
nverke
changed the title
| ptr4 = vtcm_pool->Allocate(max_bytes); | ||
| vtcm_pool->Free(ptr4, max_bytes); | ||
| } | ||
|
|
There was a problem hiding this comment.
Suggest:
TEST_F(HexagonVtcmPoolTest, find_allocation) {
void* ptr1;
void* ptr2;
void* ptr3;
ptr1 = vtcm_pool->Allocate(two_k_block);
ptr2 = vtcm_pool->Allocate(two_k_block);
vtcm_pool->Free(ptr1, two_k_block);
ptr3 = vtcm_pool->Allocate(four_k_block);
vtcm_pool->Free(ptr2, two_k_block);
vtcm_pool->Free(ptr3, four_k_block);
// Make sure at the end we have the full amount available again
ptr1 = vtcm_pool->Allocate(max_bytes);
vtcm_pool->Free(ptr1, max_bytes);
}
There was a problem hiding this comment.
Confirmed this fails without the fix, works great with the fix.
There was a problem hiding this comment.
Looks great - I like the new test cases!
| auto entry_to_allocate = free_.begin(); | ||
| for (auto it = free_.begin(); it != free_.end(); it++) { | ||
| if ((it->second < entry_to_allocate->second) && (it->second >= nbytes)) { | ||
| if (((entry_to_allocate->second < nbytes) || (it->second < entry_to_allocate->second)) && (it->second >= nbytes)) { |
There was a problem hiding this comment.
The groups should be slightly adjusted:
if ((entry_to_allocate->second < nbytes) || ((it->second < entry_to_allocate->second) && (it->second >= nbytes))) {
There was a problem hiding this comment.
Doesnt this lead to us updating entry_to_allocate for iterator values that may not fit the desired allocation?
There was a problem hiding this comment.
But, that's ok because it will still find a free block that works if there is one. But on reflection, I like your grouping better - I think it makes it more clear for future editors.
| ptr4 = vtcm_pool->Allocate(max_bytes); | ||
| vtcm_pool->Free(ptr4, max_bytes); | ||
| } | ||
|
|
There was a problem hiding this comment.
Looks great - I like the new test cases!
| vtcm_pool->Free(ptr1, two_k_block); | ||
|
|
||
| // Allocate a new larger block to initiate search and ensure | ||
| // it succeeds despite there not being a matching block. |
There was a problem hiding this comment.
nit: "despite there not being a match in the first free block."
* [Hexagon] Add fix for VTCM allocation search and new test case to cover the issue. * Add another test for a specific case. * [Hexagon] Fix tests. * Remove unused var * Change to comments to make more clear.
* [Hexagon] Add fix for VTCM allocation search and new test case to cover the issue. * Add another test for a specific case. * [Hexagon] Fix tests. * Remove unused var * Change to comments to make more clear.
4 participants
Previously we were failing allocations in an attempt to find the smallest allocation possible but were failing to find an allocation that worked if it was available.