Update Neon™ depthwise kernel
- Reduce duplication and simplify overall structure.
- Improve multi-threaded performance by sharing more data
in lower-level caches.
Partially Resolves: COMPMID-5054
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: Iac747f39b21c540122fa75218762631c4d787911
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7449
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Andrew Mundy
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_implementation_constraints.hpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_implementation_constraints.hpp
index 4198727..78b6aec 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_implementation_constraints.hpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_implementation_constraints.hpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -109,6 +109,12 @@
return args.channel_multiplier == 1;
}
+bool has_channel_multiplier(const DepthwiseArgs &args, const void *) __attribute__ ((unused));
+bool has_channel_multiplier(const DepthwiseArgs &args, const void *)
+{
+ return args.channel_multiplier > 1;
+}
+
bool qp_has_no_left_shift(const DepthwiseArgs &args, const void *_qp) __attribute__ ((unused));
bool qp_has_no_left_shift(const DepthwiseArgs &, const void *_qp)
{
@@ -118,6 +124,21 @@
(qp->per_layer_left_shift == 0);
}
+bool qp_zero_a_offset(const DepthwiseArgs &args, const void *_qp) __attribute__ ((unused));
+bool qp_zero_a_offset(const DepthwiseArgs &, const void *_qp)
+{
+ const auto qp = static_cast<const arm_gemm::Requantize32 *>(_qp);
+ return qp->a_offset == 0;
+}
+
+template <typename T> bool qp_skip_clamp(const DepthwiseArgs &args, const void *_qp) __attribute__ ((unused));
+template <typename T> bool qp_skip_clamp(const DepthwiseArgs &, const void *_qp)
+{
+ const auto qp = static_cast<const arm_gemm::Requantize32 *>(_qp);
+ return (qp->minval == std::numeric_limits<T>::min() &&
+ qp->maxval == std::numeric_limits<T>::max());
+}
+
} // namespace
} // namespace depthwise
} // namespace arm_conv