Integrate improved CPU depthwise convolution kernels

* Replace assembly kernels for depthwise convolution with more optimized
  ones.
* Add int8 assembly kernels.
* Fix implicit padding on optimized kernels

Resolves: COMPMID-3867, COMPMID-4361

Change-Id: I0b0867e05f61be4f368f62190d55e14d0ab3ebf2
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5622
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 105 insertions, 0 deletions

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
new file mode 100644
index 0000000000..b4814bef92
--- /dev/null
+++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_implementation_constraints.hpp
@@ -0,0 +1,105 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+/* Utilities for constructing functions which constrain which kernels are
+ * selected for a given depthwise problem.
+ *
+ * It is expected that this will be included in the files which list the
+ * available kernels. To avoid multiple definitions, an anonymous namespace is
+ * used.
+ */
+
+#pragma once
+
+#include "arm_gemm.hpp"
+#include "depthwise.hpp"
+
+namespace arm_conv
+{
+namespace depthwise
+{
+namespace
+{
+
+template <class OutputStage>
+using ConstraintFn = std::function<bool(const DepthwiseArgs &, const OutputStage &)>;
+
+using GenericConstraintFn = std::function<bool(const DepthwiseArgs &, const void *)>;
+
+GenericConstraintFn make_constraint(const GenericConstraintFn &f) __attribute__ ((unused));
+GenericConstraintFn make_constraint(const GenericConstraintFn &f)
+{
+  return f;
+}
+
+template <typename ... Fs>
+GenericConstraintFn make_constraint(const GenericConstraintFn &f, Fs ... fs)
+{
+  return [f, fs...] (const DepthwiseArgs &args, const void *os) -> bool {
+    return f(args, os) && make_constraint(fs...)(args, os);
+  };
+}
+
+template <typename OutputStage=Nothing, typename ... Fs>
+ConstraintFn<OutputStage> constraint(Fs ... fs)
+{
+  return [fs...] (const DepthwiseArgs &args, const OutputStage &os) -> bool {
+    return make_constraint(fs...)(args, &os);
+  };
+}
+
+// Some useful constraints
+template <class Strategy>
+bool is_supported(const DepthwiseArgs &args, const void *)
+{
+  return ((args.kernel_rows == Strategy::kernel_rows) &&
+          (args.kernel_cols == Strategy::kernel_cols) &&
+          (args.stride_rows == Strategy::stride_rows) &&
+          (args.stride_cols == Strategy::stride_cols));
+}
+
+bool cpu_has_dot_product(const DepthwiseArgs &args, const void *) __attribute__ ((unused));
+bool cpu_has_dot_product(const DepthwiseArgs &args, const void *)
+{
+  return args.cpu_info->has_dotprod();
+}
+
+bool has_no_channel_multiplier(const DepthwiseArgs &args, const void *) __attribute__ ((unused));
+bool has_no_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)
+{
+  const auto qp = static_cast<const arm_gemm::Requantize32 *>(_qp);
+  return qp->per_channel_requant ?
+    (qp->per_channel_left_shifts == nullptr) :
+    (qp->per_layer_left_shift == 0);
+}
+
+}  // namespace
+}  // namespace depthwise
+}  // namespace arm_conv