Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[CPU]Fix crash issue: RuntimeError: Primitive descriptor was not found for… #16186

Merged
Show file tree
Hide file tree
Changes from 4 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 8 additions & 2 deletions src/plugins/intel_cpu/src/nodes/matmul.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -624,15 +624,21 @@ void MatMul::prepareParams() {
primitive_desc_iterator itpd = desc.createPrimitiveDescriptorIterator(engine, key.attr);
matmul::primitive_desc prim_desc;

auto itpd_first = itpd;
while (static_cast<bool>(itpd)) {
impl_desc_type impl_type = parse_impl_name(itpd.impl_info_str());

if (impl_type == key.implType) {
prim_desc = itpd.get();
break;
}
if (!itpd.next_impl())
return matmul();
if (!itpd.next_impl()) {
// implType(AMX) will be not in list when input dynamic shape for AMX.
// Bcause AMX will be available when it assumes batch 16 for dynamic shape in compile model.
// However if we only pass batch 1 in runtime, AMX will be not available(Performance is low).
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
// implType(AMX) will be not in list when input dynamic shape for AMX.
// Bcause AMX will be available when it assumes batch 16 for dynamic shape in compile model.
// However if we only pass batch 1 in runtime, AMX will be not available(Performance is low).
// implType(AMX) may not be in the list when the input shapes are dynamic in the case of AMX.
// AMX is available on the compile model stage when the dummy shape is used.
// However, if we pass batch 1 at runtime, AMX will not be available (due to the oneDNN implementation selection algorithm).

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

My version would be 😄:
"In case of dynamic shapes an implementation type chosen as optimal for a primitive_desc with undefined input shapes, is not necessarily available for the primitive_desc with defined shape.
Example: brgemm_avx512_amx is available for a primitive with undefined input shapes but not available for primitive_desc with input batch 1"

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@EgorDuplensky gave the root cause. It will be great.

prim_desc = itpd_first.get();
break;
}
}
return matmul(prim_desc);
};
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -1277,6 +1277,13 @@ INSTANTIATE_TEST_SUITE_P(nightly_MM_Brgemm_Amx_Static, MatMulLayerCPUTest, testB


const std::vector<ShapeRelatedParams> IS_Brgemm_Dynamic = {
{
{
{{-1, 256}, {{1, 256}}},
{{256, 384}, {{256, 384}}}
},
{false, false}
},
{
{
{{-1, -1}, {{55, 12}, {33, 7}}},
Expand Down