1 files changed, 43 insertions, 0 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
index 58e4861bc0..515d55c73b 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
@@ -48,10 +48,20 @@
 #include "kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL.hpp"
 #include "kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp"
 #endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS
+
+#ifdef ARM_COMPUTE_ENABLE_SVE
+#ifdef ARM_COMPUTE_ENABLE_SME2
+#include "kernels/sme2_gemv_bf16fp32_dot_16VL.hpp"
+#include "kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL.hpp"
+#include "kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL.hpp"
+#include "kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL.hpp"
+#endif // ARM_COMPUTE_ENABLE_SME2
+
 #include "kernels/sve_hybrid_bf16fp32_dot_6x4VL.hpp"
 #include "kernels/sve_hybrid_bf16fp32_mmla_6x4VL.hpp"
 #include "kernels/sve_interleaved_bf16fp32_dot_8x3VL.hpp"
 #include "kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp"
+#endif // ARM_COMPUTE_ENABLE_SVE
 
 namespace arm_gemm {
 
@@ -60,6 +70,39 @@ static const GemmImplementation<bfloat16, float> gemm_bf16_methods[] =
 #ifdef __aarch64__
 #ifdef ARM_COMPUTE_ENABLE_BF16
 #ifdef ARM_COMPUTE_ENABLE_SVE
+#ifdef ARM_COMPUTE_ENABLE_SME2
+// SME kernels
+{
+    GemmMethod::GEMM_HYBRID,
+    "sme2_gemv_bf16fp32_dot_16VL",
+    [](const GemmArgs &args) { return args._ci->has_sme2() && args._Msize==1 && args._nbatches==1 && !args._indirect_input; },
+    nullptr,
+    [](const GemmArgs &args) { return new GemvPretransposed<cls_sme2_gemv_bf16fp32_dot_16VL, bfloat16, float>(args); }
+},
+{
+    GemmMethod::GEMM_INTERLEAVED,
+    "sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL",
+    [](const GemmArgs &args) { return args._ci->has_sme2(); },
+    [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>();
+                               return args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); },
+    [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL, bfloat16, float>(args); }
+},
+{
+    GemmMethod::GEMM_INTERLEAVED,
+    "sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL",
+    [](const GemmArgs &args) { return args._ci->has_sme2(); },
+    [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>();
+                               return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); },
+    [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL, bfloat16, float>(args); }
+},
+{
+    GemmMethod::GEMM_INTERLEAVED,
+    "sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL",
+    [](const GemmArgs &args) { return args._ci->has_sme2(); },
+    nullptr,
+    [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL, bfloat16, float>(args); }
+},
+#endif // ARM_COMPUTE_ENABLE_SME2
 // gemm_bf16_interleaved
 GemmImplementation<bfloat16, float>::with_estimate(
     GemmMethod::GEMM_INTERLEAVED,