[CPU] Enable compressed FC via oneDNN Matmul primitive #27459
Conversation
| @@ -64,11 +64,11 @@ endif() | |||
|
|
|||
| # find all the source files with the name of a class file | |||
| if(X86_64) | |||
| file(GLOB_RECURSE LIST_OF_TEST_ARCH_INSTANCES ${TEST_DIR}/instances/x64/${TEST_CLASS_FILE_NAME}) | |||
| file(GLOB_RECURSE LIST_OF_TEST_ARCH_INSTANCES ${TEST_DIR}/x64/${TEST_CLASS_FILE_NAME}) | |||
There was a problem hiding this comment.
TODO: remove before merge
| @@ -42,7 +42,7 @@ namespace { | |||
| if (!ftz) { | |||
| return; | |||
| } | |||
| if (src.getDesc().getPrecision() != ov::element::f32 || dst.getDesc().getPrecision() == ov::element::bf16) { | |||
| if (src.getDesc().getPrecision() != ov::element::f32 || dst.getDesc().getPrecision() != ov::element::f32) { | |||
There was a problem hiding this comment.
@maxnick Please take a look. No sure about idea behind previous logic. Maybe you remember smt.
Anyway previous condition was incorrect for cases like src (fp32) -> dst (i32); it tried to apply ftz on i32 data.
There was a problem hiding this comment.
The original idea is to skip denormal nullifying for all the non float source types (as it doesn't make sense for integer types and bf16, as for the latter case HW implicitly assumes FTZ and DAZ flags are set when performing bf16 specific operations) also for bf16 dst type, as HW performs flushing denormal to zeros while converting float to bf16 using specific instructions.
|
@EgorDuplensky Please review the PR |
dmitry-gorokhov
force-pushed
the
feature/fc_matmul_with_decompression_executor
branch
from
7479c09 to
4137826
Compare
dmitry-gorokhov
force-pushed
the
feature/fc_matmul_with_decompression_executor
branch
2 times, most recently
from
7ffe60a to
370e2b2
Compare
dmitry-gorokhov
force-pushed
the
feature/fc_matmul_with_decompression_executor
branch
from
370e2b2 to
9a336e2
Compare
Details:
-- Build OV with: -DENABLE_DEBUG_CAPS=ON cmake option
-- export OV_CPU_ENABLE_DNNL_MAMTUL_FOR_FC=1