1 files changed, 31 insertions, 8 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
index 9e49df1c28..53749fd4f6 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
@@ -34,9 +34,12 @@
 #include "kernels/a64_gemm_s8_12x8.hpp"
 #include "kernels/a64_gemm_s8_4x4.hpp"
 #include "kernels/a64_hybrid_s8s32_dot_16x4.hpp"
+#include "kernels/a64_smallK_hybrid_s8s32_dot_4x6.hpp"
+#include "kernels/a64_smallK_hybrid_s8s32_dot_4x8.hpp"
 #include "kernels/sve_hybrid_s8s32_dot_4VLx4.hpp"
 #include "kernels/sve_interleaved_s8s32_dot_3VLx8.hpp"
 #include "kernels/sve_native_s8s32_dot_4VLx4.hpp"
+#include "kernels/sve_smallK_hybrid_s8s32_dot_1VLx8.hpp"
 
 namespace arm_gemm {
 
@@ -44,8 +47,15 @@ static const GemmImplementation<int8_t, int32_t> gemm_s8_methods[] = {
 #ifdef __ARM_FEATURE_SVE
 {
     GemmMethod::GEMM_HYBRID,
+    "smallK_hybrid_s8s32_dot_1VLx8",
+    [](const GemmArgs<int32_t> &args) { return args._Ksize<=64 && args._alpha==1 && args._beta==0 && !args._trA && args._pretransposed_hint; },
+    nullptr,
+    [](const GemmArgs<int32_t> &args) { return new GemmHybrid<smallK_hybrid_s8s32_dot_1VLx8, int8_t, int32_t>(args); }
+},
+{
+    GemmMethod::GEMM_HYBRID,
     "hybrid_s8s32_dot_4VLx4",
-    [](const GemmArgs<int32_t> &args) { return args._Ksize>=16 && args._alpha==1 && !args._trA && !args._trB && args._pretransposed_hint; },
+    [](const GemmArgs<int32_t> &args) { return args._Ksize>=16 && args._alpha==1 && !args._trA && args._pretransposed_hint; },
     [](const GemmArgs<int32_t> &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); },
     [](const GemmArgs<int32_t> &args) { return new GemmHybrid<hybrid_s8s32_dot_4VLx4, int8_t, int32_t>(args); }
 },
@@ -59,15 +69,29 @@ static const GemmImplementation<int8_t, int32_t> gemm_s8_methods[] = {
 {
     GemmMethod::GEMM_INTERLEAVED,
     "interleaved_s8s32_dot_3VLx8",
-    [](const GemmArgs<int32_t> &args) { return (args._Ksize>4); },
+    [](const GemmArgs<int32_t> &args) { return (args._Ksize>4) && args._alpha==1 && (args._beta == 0 || args._beta==1); },
     nullptr,
     [](const GemmArgs<int32_t> &args) { return new GemmInterleaved<interleaved_s8s32_dot_3VLx8, int8_t, int32_t>(args); }
 },
 #endif
 {
     GemmMethod::GEMM_HYBRID,
+    "smallK_hybrid_s8s32_dot_4x8",
+    [](const GemmArgs<int32_t> &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && args._alpha==1 && args._beta==0 && !args._trA && args._pretransposed_hint; },
+    nullptr,
+    [](const GemmArgs<int32_t> &args) { return new GemmHybrid<smallK_hybrid_s8s32_dot_4x8, int8_t, int32_t>(args); }
+},
+{
+    GemmMethod::GEMM_HYBRID,
+    "smallK_hybrid_s8s32_dot_4x6",
+    [](const GemmArgs<int32_t> &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && args._alpha==1 && args._beta==0 && !args._trA && args._pretransposed_hint; },
+    nullptr,
+    [](const GemmArgs<int32_t> &args) { return new GemmHybrid<smallK_hybrid_s8s32_dot_4x6, int8_t, int32_t>(args); }
+},
+{
+    GemmMethod::GEMM_HYBRID,
     "hybrid_s8s32_dot_16x4",
-    [](const GemmArgs<int32_t> &args) { return args._ci->has_dotprod() && args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; },
+    [](const GemmArgs<int32_t> &args) { return args._ci->has_dotprod() && args._Ksize>=16 && args._alpha==1 && !args._trA && !args._trB && args._pretransposed_hint; },
     [](const GemmArgs<int32_t> &args) { return args._Nsize<=256 && args._Ksize>128; },
     [](const GemmArgs<int32_t> &args) { return new GemmHybrid<hybrid_s8s32_dot_16x4, int8_t, int32_t>(args); }
 },
@@ -100,11 +124,10 @@ const GemmImplementation<int8_t, int32_t> *gemm_implementation_list<int8_t, int3
 }
 
 /* Explicitly instantiate the external functions for these types. */
-template UniqueGemmCommon<int8_t, int32_t> gemm<int8_t, int32_t>(const GemmArgs<int32_t> &args);
-template KernelDescription get_gemm_method<int8_t, int32_t>(const GemmArgs<int32_t> &args);
-template bool method_is_compatible<int8_t, int32_t>(GemmMethod method, const GemmArgs<int32_t> &args);
-template std::vector<KernelDescription> get_compatible_kernels<int8_t, int32_t> (const GemmArgs<int32_t> &args);
+template UniqueGemmCommon<int8_t, int32_t> gemm<int8_t, int32_t, Nothing>(const GemmArgs<int32_t> &args, const Nothing &);
+template KernelDescription get_gemm_method<int8_t, int32_t, Nothing>(const GemmArgs<int32_t> &args, const Nothing &);
+template std::vector<KernelDescription> get_compatible_kernels<int8_t, int32_t, Nothing> (const GemmArgs<int32_t> &args, const Nothing &);
 
 } // namespace arm_gemm
 
-#endif // __aarch64__
\ No newline at end of file
+#endif // __aarch64__