COMPMID-2707: add keep_dims parameter to Reduction Operation
The added parameter is used to decide whether or not to keep
the target dimension of reduction operation. ArgMinMax operations
will always remove the reduced dimension. Following things
are updated to support the parameter.
- [CL/NEON] functions and reference kernel
- [CL/NEON] ArgMinMax function to use ReductionOperation function
- [CL/NEON] validation test suite for Reduction and ArgMinMax operations
to validate the added parameter
- ReductionOperationFixture is modified NOT to pre-populate output
tensor and now relies on underlying kernel/function.
- Adjust CL validation test suite for Reduction operation to remove
excessive test cases with axis values beyond input tensor's
dimension.
Change-Id: I3e24d276ed469a4201f323001708f0c525f11c4f
Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2167
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/src/core/Utils.cpp b/src/core/Utils.cpp
index 7e1af0e..fa335d7 100644
--- a/src/core/Utils.cpp
+++ b/src/core/Utils.cpp
@@ -427,6 +427,16 @@
return std::make_pair(w, h);
}
+bool arm_compute::needs_serialized_reduction(ReductionOperation op, DataType dt, unsigned int axis)
+{
+ const bool is_arg_min_max = (op == ReductionOperation::ARG_IDX_MAX || op == ReductionOperation::ARG_IDX_MIN);
+ const bool is_min_max = (op == ReductionOperation::MAX || op == ReductionOperation::MIN);
+ const bool is_quantized_type = is_data_type_quantized(dt);
+ const bool is_first_dim = (axis == 0);
+
+ return !is_first_dim || is_arg_min_max || is_min_max || is_quantized_type;
+}
+
#ifdef ARM_COMPUTE_ASSERTS_ENABLED
void arm_compute::print_consecutive_elements(std::ostream &s, DataType dt, const uint8_t *ptr, unsigned int n, int stream_width, const std::string &element_delim)
{