Non-standard reshape fail during reshape_lazy lowering

[attila-dusnoki-htec]

AWSRN-BAM fails with the following:

migraphx-driver: /code/AMDMIGraphX/src/targets/gpu/lowering.cpp:76: void migraphx::gpu::miopen_apply::check_shape(shape, instruction_ref): Assertion 'x == i->get_shape()' failed.

This script can be used for creating the model.

The issue occures during lowering:

BEFORE
@619 = gpu::precompile_op[op=pointwise,additional_args=1,ignore_modules=0](@571,@617,@599,@615,@145,@618), [main:pointwise127] -> float_type, {1, 3, 85, 2, 85, 2}, {86700, 28900, 85, 14450, 1, 7225}, target_id=0
@620 = reshape[dims={1, 3, 170, 170}](@619) -> float_type, {1, 3, 170, 170}, {86700, 28900, 14450, 7225}, target_id=0
@621 = @return(@620), target_id=0

AFTER
@619 = gpu::precompile_op[op=pointwise,additional_args=1,ignore_modules=0](@571,@617,@599,@615,@145,@618), [main:pointwise127] -> float_type, {1, 3, 85, 2, 85, 2}, {86700, 28900, 85, 14450, 1, 7225}, target_id=0
@620 = allocate[shape=float_type, {1, 3, 85, 2, 85, 2}, {86700, 28900, 85, 14450, 1, 7225},buf_type=nullopt] -> float_type, {1, 3, 85, 2, 85, 2}, {86700, 28900, 85, 14450, 1, 7225}, target_id=0
@621 = gpu::contiguous(@619,@620) -> float_type, {1, 3, 85, 2, 85, 2}, {86700, 28900, 340, 170, 2, 1}, target_id=0
@622 = allocate[shape=float_type, {1, 3, 170, 170}, {86700, 28900, 170, 1},buf_type=nullopt] -> float_type, {1, 3, 170, 170}, {86700, 28900, 170, 1}, target_id=0
@623 = reshape_lazy[dims={1, 3, 170, 170}](@621) -> float_type, {1, 3, 170, 170}, {86700, 28900, 170, 1}, target_id=0
@624 = gpu::contiguous(@623,@622) -> float_type, {1, 3, 170, 170}, {86700, 28900, 170, 1}, target_id=0
@625 = @return(@624), target_id=0

The original reshape is in non-standard format. The gpu::contiguous which is added by add_reshape_lazy_op is in standard format.

{1, 3, 170, 170}, {86700, 28900, 14450, 7225} // before
{1, 3, 170, 170}, {86700, 28900, 170, 1} // after

cc @causten @pfultz2

[gyulaz-htec]

Could be related to: #2487

[causten]

@bpickrel see if this is a duplicate

[bpickrel]

Not a clear duplicate--the same assertion fails but with two different model files, and a different sequence of operators. I'll wait and see if they both have the same underlying cause, though.

[bpickrel]

The code branch for #2038 seems to fix this problem. It runs all 3 of these models successfully, and without needing the MIGRAPHX_DISABLE_MLIR setting.