…e-org#22275)

Weight backward convolutions have a special CHWN layout, where the
filter sizes (corresponding to output image sizes in forward
convolutions) are typically large, while the output spatial dimensions
are small. This makes the split reduction strategy particularly
effective. This PR adds support to split these convs along the input
channel dimension.

Some experimental thresholds are applied to filter out cases that won't
benefit from splitting reduction. Particular checks include:

- When the batch and output channel sizes are large, the workload tends
to distributed across many workgroups, making split reduction little to
no effect.
- When the input spatial sizes are small while the batch and output
channel sizes are relatively larger (medium size), split reduction often
has no effect or even degrades performance.

---------

Signed-off-by: yzhang93 <[email protected]>
Signed-off-by: Philipp <[email protected]>

… weight backward convs (#22491)

This PR is a follow-up for #22275.
It removes the constraint that only splitting input channel dimension,
and added support to split across multiple dimensions. The heuristics
for setting multi-dimension tile sizes is similar to what is for GEMM
#22357. More than half of the
tracked weight backward shapes are benefiting from this change.

Example runtime comparison for
`convbfp16 -n 16 -c 16 -H 225 -W 225 -k 64 -y 3 -x 3 -p 1 -q 1 -u 1 -v 1
-l 1 -j 1 -m conv -g 1 -F 4 -t 1 --in_layout NHWC --out_layout NHWC
--fil_layout NHWC --iter 100`

- Without split reduction: 19352.8 ms
- Split only the input channel dimension: 1445.1 ms
- Split multiple reduction dimensions: 371.7 ms

---------

Signed-off-by: yzhang93 <[email protected]>

… weight backward convs (iree-org#22491)

This PR is a follow-up for iree-org#22275.
It removes the constraint that only splitting input channel dimension,
and added support to split across multiple dimensions. The heuristics
for setting multi-dimension tile sizes is similar to what is for GEMM
iree-org#22357. More than half of the
tracked weight backward shapes are benefiting from this change.

Example runtime comparison for
`convbfp16 -n 16 -c 16 -H 225 -W 225 -k 64 -y 3 -x 3 -p 1 -q 1 -u 1 -v 1
-l 1 -j 1 -m conv -g 1 -F 4 -t 1 --in_layout NHWC --out_layout NHWC
--fil_layout NHWC --iter 100`

- Without split reduction: 19352.8 ms
- Split only the input channel dimension: 1445.1 ms
- Split multiple reduction dimensions: 371.7 ms

---------

Signed-off-by: yzhang93 <[email protected]>

…e-org#22275)

Weight backward convolutions have a special CHWN layout, where the
filter sizes (corresponding to output image sizes in forward
convolutions) are typically large, while the output spatial dimensions
are small. This makes the split reduction strategy particularly
effective. This PR adds support to split these convs along the input
channel dimension.

Some experimental thresholds are applied to filter out cases that won't
benefit from splitting reduction. Particular checks include:

- When the batch and output channel sizes are large, the workload tends
to distributed across many workgroups, making split reduction little to
no effect.
- When the input spatial sizes are small while the batch and output
channel sizes are relatively larger (medium size), split reduction often
has no effect or even degrades performance.

---------

Signed-off-by: yzhang93 <[email protected]>

… weight backward convs (iree-org#22491)

This PR is a follow-up for iree-org#22275.
It removes the constraint that only splitting input channel dimension,
and added support to split across multiple dimensions. The heuristics
for setting multi-dimension tile sizes is similar to what is for GEMM
iree-org#22357. More than half of the
tracked weight backward shapes are benefiting from this change.

Example runtime comparison for
`convbfp16 -n 16 -c 16 -H 225 -W 225 -k 64 -y 3 -x 3 -p 1 -q 1 -u 1 -v 1
-l 1 -j 1 -m conv -g 1 -F 4 -t 1 --in_layout NHWC --out_layout NHWC
--fil_layout NHWC --iter 100`

- Without split reduction: 19352.8 ms
- Split only the input channel dimension: 1445.1 ms
- Split multiple reduction dimensions: 371.7 ms

---------

Signed-off-by: yzhang93 <[email protected]>