COMPMID-1271: New system for GEMM heuristics

This patch implements a system for separating the "validity" from
"preferred" aspect of the current heuristics in gemm_*.cpp.

Now, each gemm_*.cpp defines a list of candidate implementations,
each of which supplies an is_valid() function (to check for
validity), an is_preferred() function (the "heuristic" part), and an
instantiate() function which actually produces the GemmCommon object
pointer.

The actual gemm() function is now templated and uses this list to
select an implementation.  This patch also implements a mechanism to
identify the preferred implementation, and override it via the
GemmConfig structure.

Change-Id: Id49ab7af8bf2e3e9fd951a9698883ade234d40e1
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/139120
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

1 files changed, 35 insertions, 15 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
index 04803eb81a..f50b399de5 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_int8.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_s8_4x4.hpp"
 #include "kernels/a64_gemm_s16_12x8.hpp"
 #include "kernels/a64_gemm_s8_12x8.hpp"
+#include "kernels/a64_gemm_s8_4x4.hpp"
 
 namespace arm_gemm {
 
-template<>
-UniqueGemmCommon<int8_t, int32_t> gemm<int8_t, int32_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 int32_t alpha, const int32_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<int8_t, int32_t>(new GemmInterleaved<gemm_s8_12x8, int8_t, int32_t>(&ci, M, N, K, nbatches, nmulti, trA, trB, alpha, beta, maxthreads, pretransposed_hint));
+class GemmImpl_gemm_s8_interleaved_dot : public GemmImplementation<int8_t, int32_t> {
+public:
+    bool is_supported(const GemmArgs<int32_t> &args) override {
+        return args._ci->has_dotprod();
     }
 
-    return UniqueGemmCommon<int8_t, int32_t>(new GemmInterleaved<gemm_s8_4x4, int8_t, int32_t>(&ci, M, N, K, nbatches, nmulti, trA, trB, alpha, beta, maxthreads, pretransposed_hint));
+    UniqueGemmCommon<int8_t, int32_t> instantiate(const GemmArgs<int32_t> &args) override {
+        return UniqueGemmCommon<int8_t, int32_t>(new GemmInterleaved<gemm_s8_12x8, int8_t, int32_t>(args));
+    }
+
+    GemmImpl_gemm_s8_interleaved_dot() : GemmImplementation<int8_t, int32_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_s8_interleaved : public GemmImplementation<int8_t, int32_t> {
+public:
+    UniqueGemmCommon<int8_t, int32_t> instantiate(const GemmArgs<int32_t> &args) override {
+        return UniqueGemmCommon<int8_t, int32_t>(new GemmInterleaved<gemm_s8_4x4, int8_t, int32_t>(args));
+    }
+
+    GemmImpl_gemm_s8_interleaved() : GemmImplementation<int8_t, int32_t>(GemmMethod::GEMM_INTERLEAVED) { }
+};
+
+static std::vector<GemmImplementation<int8_t, int32_t> *> gemm_s8_methods = {
+    new GemmImpl_gemm_s8_interleaved_dot(),
+    new GemmImpl_gemm_s8_interleaved()
+};
+
+template<>
+std::vector<GemmImplementation<int8_t, int32_t> *> &gemm_implementation_list<int8_t, int32_t>() {
+    return gemm_s8_methods;
 }
 
+/* Explicitly instantiate the external functions for these types. */
+template UniqueGemmCommon<int8_t, int32_t> gemm<int8_t, int32_t>(GemmArgs<int32_t> &args, GemmConfig *cfg);
+template GemmMethod get_gemm_method<int8_t, int32_t>(GemmArgs<int32_t> &args);
+template bool method_is_compatible<int8_t, int32_t>(GemmMethod method, GemmArgs<int32_t> &args);
+
 } // namespace arm_gemm
 
-#endif // aarch64
+#endif // __aarch64__