1 files changed, 36 insertions, 16 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
index 1ca92f9d4e..d97dd5c3de 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
@@ -25,32 +25,52 @@
 
 #include "arm_gemm.hpp"
 #include "gemm_common.hpp"
+#include "gemm_implementation.hpp"
 #include "gemm_interleaved.hpp"
 
-#include "kernels/a64_gemm_u8_4x4.hpp"
+#include "kernels/a64_gemm_u16_12x8.hpp"
 #include "kernels/a64_gemm_u8_12x8.hpp"
+#include "kernels/a64_gemm_u8_4x4.hpp"
 
 namespace arm_gemm {
 
-template<>
-UniqueGemmCommon<uint8_t, uint32_t> gemm<uint8_t, uint32_t>(const CPUInfo &ci, const unsigned int M, const unsigned int N, const unsigned int K,
-                                                            const unsigned int nbatches, const unsigned int nmulti,
-                                                            const bool trA, const bool trB, const uint32_t alpha, const uint32_t beta,
-                                                            const int maxthreads, const bool pretransposed_hint) {
-    if (ci.has_dotprod()) {
-        // Dot product supporting CPUs.  This family has a special version for A55r1.
-        return UniqueGemmCommon<uint8_t, uint32_t>(new GemmInterleaved<gemm_u8_12x8, uint8_t, uint32_t>(&ci, M, N, K, nbatches, nmulti, trA, trB, alpha, beta, maxthreads, pretransposed_hint));
+class GemmImpl_gemm_u8_interleaved_dot : public GemmImplementation<uint8_t, uint32_t> {
+public:
+    bool is_supported(const GemmArgs<uint32_t> &args) override {
+        return args._ci->has_dotprod();
     }
 
-    // Non dot-product code.
-    return UniqueGemmCommon<uint8_t, uint32_t>(new GemmInterleaved<gemm_u8_4x4, uint8_t, uint32_t>(&ci, M, N, K, nbatches, nmulti, trA, trB, alpha, beta, maxthreads, pretransposed_hint));
+    UniqueGemmCommon<uint8_t, uint32_t> instantiate(const GemmArgs<uint32_t> &args) override {
+        return UniqueGemmCommon<uint8_t, uint32_t>(new GemmInterleaved<gemm_u8_12x8, uint8_t, uint32_t>(args));
+    }
+
+    GemmImpl_gemm_u8_interleaved_dot() : GemmImplementation<uint8_t, uint32_t>(GemmMethod::GEMM_INTERLEAVED_DOT) { }
+};
 
-    // TODO: There's a better approach for A53, but it doesn't work
-    // well on heterogeneous systems as the required data formats
-    // are different.  Figure out how to enable this:
-    // gemm = new GemmInterleaved<gemm_s16_12x8, int8_t, int32_t>(ci, M, N, K, trA, trB);
+class GemmImpl_gemm_u8_interleaved : public GemmImplementation<uint8_t, uint32_t> {
+public:
+    UniqueGemmCommon<uint8_t, uint32_t> instantiate(const GemmArgs<uint32_t> &args) override {
+        return UniqueGemmCommon<uint8_t, uint32_t>(new GemmInterleaved<gemm_u8_4x4, uint8_t, uint32_t>(args));
+    }
+
+    GemmImpl_gemm_u8_interleaved() : GemmImplementation<uint8_t, uint32_t>(GemmMethod::GEMM_INTERLEAVED) { }
+};
+
+static std::vector<GemmImplementation<uint8_t, uint32_t> *> gemm_u8_methods = {
+    new GemmImpl_gemm_u8_interleaved_dot(),
+    new GemmImpl_gemm_u8_interleaved()
+};
+
+template<>
+std::vector<GemmImplementation<uint8_t, uint32_t> *> &gemm_implementation_list<uint8_t, uint32_t>() {
+    return gemm_u8_methods;
 }
 
+/* Explicitly instantiate the external functions for these types. */
+template UniqueGemmCommon<uint8_t, uint32_t> gemm<uint8_t, uint32_t>(GemmArgs<uint32_t> &args, GemmConfig *cfg);
+template GemmMethod get_gemm_method<uint8_t, uint32_t>(GemmArgs<uint32_t> &args);
+template bool method_is_compatible<uint8_t, uint32_t>(GemmMethod method, GemmArgs<uint32_t> &args);
+
 } // namespace arm_gemm
 
-#endif // aarch64
+#endif // __aarch64__