Integrate SME2 kernels

* Add SME/SME2 detection.
* Integrate SME2 implementation for:
  - Normal convolution
  - Winograd
  - Depthwise convolution
  - Pooling

Resolves: COMPMID-5700
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: I2f1ca1d05f8cfeee9309ed1c0a36096a4a6aad5c
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8692
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 21 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
index 78b6aec388..1ba7694f1e 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
@@ -85,6 +85,18 @@ bool cpu_has_dot_product(const DepthwiseArgs &args, const void *)
   return args.cpu_info->has_dotprod();
 }
 
+bool cpu_has_sme(const DepthwiseArgs &args, const void *) __attribute__ ((unused));
+bool cpu_has_sme(const DepthwiseArgs &args, const void *)
+{
+  return args.cpu_info->has_sme();
+}
+
+bool cpu_has_sme2(const DepthwiseArgs &args, const void *) __attribute__ ((unused));
+bool cpu_has_sme2(const DepthwiseArgs &args, const void *)
+{
+  return args.cpu_info->has_sme2();
+}
+
 bool cpu_has_sve(const DepthwiseArgs &args, const void *) __attribute__ ((unused));
 bool cpu_has_sve(const DepthwiseArgs &args, const void *)
 {
@@ -115,6 +127,15 @@ bool has_channel_multiplier(const DepthwiseArgs &args, const void *)
   return args.channel_multiplier > 1;
 }
 
+// Planar kernels require a "priming" step before the main processing loop.  The kernels can prime with left padding
+// or input data, but not right padding - which could be needed in some extreme cases such as a 5x5 kernel, width 1
+// padding 2.  These are rare enough and can be handled with other kernels anyway, so filter them out with this.
+bool no_prime_right_pad(const DepthwiseArgs &args, const void *) __attribute__ ((unused));
+bool no_prime_right_pad(const DepthwiseArgs &args, const void *)
+{
+  return (args.input_cols + args.padding.left) >= (args.kernel_cols - 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)
 {