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_gemm/gemm_interleaved.hpp b/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
index 470cee1557..e7346e8039 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_interleaved.hpp
@@ -30,6 +30,7 @@
 #include "bfloat.hpp"
 #include "convolver.hpp"
 #include "kernel_weight_format.hpp"
+#include "kernel_traits.hpp"
 #include "mergeresults.hpp"
 #include "performance_parameters.hpp"
 #include "quantized.hpp"
@@ -576,6 +577,26 @@ class GemmInterleaved : public GemmCommon<To, Tr> {
             return get_ktotal(args);
         }
 
+        // Special blocking for SME
+        if (is_sme<strategy>::value) {
+            // Don't bother to block below this size threshold, experimentally determined to be 320 for FP32
+            unsigned int scaling_threshold = 1280 / sizeof(Toi);
+
+            if (get_ktotal(args) <= scaling_threshold) {
+                return get_ktotal(args);
+            }
+
+            // Once we are blocking, this (lower) threshold determines when we should use more blocks
+            // NOTE: Could be that some factor-based solution would work better here.
+            unsigned int max_block_size = 1024 / sizeof(Toi);
+
+            unsigned int num_k_blocks = iceildiv(get_ktotal(args), max_block_size);
+
+            unsigned int k_block = roundup(iceildiv(get_ktotal(args), num_k_blocks), strategy::k_unroll());
+
+            return k_block;
+        }
+
         const unsigned int L1_size = args._ci->get_L1_cache_size();
         unsigned int k_block;