[GPU] Optimize out Gather by converting to implicit crop #17743
Conversation
byungilm
force-pushed
the
opt_out_gather_crop
branch
from
7d9919b to
8848ebd
Compare
| if (usr->get_preferred_impl_type() == impl_types::onednn) | ||
| return; | ||
| } | ||
| if (is_optimizable_padding_for_crop(node)) { |
There was a problem hiding this comment.
what about style below:
if (!is_optimizable_padding_for_crop)())
return;
Other code uses such style and it keeps indent small.
| if (is_optimizable_padding_for_crop(node)) { | ||
| const auto& crop_layout = node.get_output_layout(); | ||
| const auto& crop_size = crop_layout.get_tensor(); | ||
| const auto& out_padd = crop_layout.data_padding; |
There was a problem hiding this comment.
Modified. Also applied to old code.
| auto new_padding = out_padd; | ||
| new_padding.lower_size().batch[0] = opt_lower_pad; | ||
| new_padding.upper_size().batch[0] = opt_upper_pad; | ||
| node.set_output_padding(new_padding); |
There was a problem hiding this comment.
Please add a unittest for this.
byungilm
force-pushed
the
opt_out_gather_crop
branch
3 times, most recently
from
4c370b6 to
6b96b85
Compare
byungilm
force-pushed
the
opt_out_gather_crop
branch
4 times, most recently
from
2e6186d to
356bbdc
Compare
|
|
||
| std::vector<int8_t> expected_results = { | ||
| 5, 6, 7, 8 | ||
| }; |
There was a problem hiding this comment.
what is the purpose of this test? It is just generated as gather, not crop..
There was a problem hiding this comment.
As I explained shortly, I wanted to check processing of B axis cropping by both gather and crop.
I also added similar test into crop_gpu_test.
Expected results of can_be_optimized are different.
byungilm
force-pushed
the
opt_out_gather_crop
branch
from
28c0955 to
9d4d38e
Compare
|
|
||
| std::vector<int8_t> expected_results = { | ||
| 5, 6, 7, 8 | ||
| }; |
There was a problem hiding this comment.
could you add crop test for opt-out?
There was a problem hiding this comment.
This test also pass through the logic.
[ RUN ] crop_single_axis.simple_Baxis
-- prepare_buffer_fusing::run
-- can_crop_be_optimized_along_batch()
There was a problem hiding this comment.
Added test into 'prepare_buffer_fusing'
byungilm
force-pushed
the
opt_out_gather_crop
branch
from
9d4d38e to
28dc744
Compare
+ Changed Gather if it divides input tensor along batch axis + Converted Gather to cldnn Crop in CreateGatherOpBase + Added implicit Crop condition for batch axis Signed-off-by: Min, Byungil <[email protected]>
byungilm
force-pushed
the
opt_out_gather_crop
branch
from
28dc744 to
8f35c81
Compare
…lkit#17743) + Changed Gather if it divides input tensor along batch axis + Converted Gather to cldnn Crop in CreateGatherOpBase + Added implicit Crop condition for batch axis Signed-off-by: Min, Byungil <[email protected]>
Details:
Tickets: