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, 98 insertions, 1 deletions

diff --git a/arm_compute/core/NEON/kernels/assembly/arm_gemm.hpp b/arm_compute/core/NEON/kernels/assembly/arm_gemm.hpp
index 8d1433dd24..162cbc5c46 100644
--- a/arm_compute/core/NEON/kernels/assembly/arm_gemm.hpp
+++ b/arm_compute/core/NEON/kernels/assembly/arm_gemm.hpp
@@ -30,9 +30,100 @@
 
 namespace arm_gemm {
 
+enum class GemmMethod
+{
+    DEFAULT,
+    GEMV_BATCHED,
+    GEMV_PRETRANSPOSED,
+    GEMV_NATIVE_TRANSPOSED,
+    GEMM_NATIVE,
+    GEMM_INTERLEAVED,
+    GEMM_INTERLEAVED_FP16,
+    GEMM_INTERLEAVED_DOT
+};
+
+struct GemmConfig
+{
+    GemmMethod method             = GemmMethod::DEFAULT;
+    unsigned int inner_block_size = 0;
+    unsigned int outer_block_size = 0;
+
+    GemmConfig(GemmMethod method) : method(method) { }
+};
+
+template<typename T>
+struct GemmArgs
+{
+public:
+    const CPUInfo *_ci;
+    unsigned int   _Msize;
+    unsigned int   _Nsize;
+    unsigned int   _Ksize;
+    unsigned int   _nbatches;
+    unsigned int   _nmulti;
+    bool           _trA;
+    bool           _trB;
+    T              _alpha;
+    T              _beta;
+    int            _maxthreads;
+    bool           _pretransposed_hint;
+
+    GemmArgs(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 T alpha, const T beta, const int maxthreads,
+             const bool pretransposed_hint) :
+             _ci(ci), _Msize(M), _Nsize(N), _Ksize(K), _nbatches(nbatches), _nmulti(nmulti),
+             _trA(trA), _trB(trB), _alpha(alpha), _beta(beta), _maxthreads(maxthreads),
+             _pretransposed_hint(pretransposed_hint)
+    {
+    }
+};
+
 template<typename Top, typename Tret>
 using UniqueGemmCommon = std::unique_ptr<GemmCommon<Top, Tret> >;
 
+/* Low level API calls.
+ * These are implemented as 'GemmArgs' versions, or with the arguments explicitly listed. */
+
+/* method_is_compatible(): Can a GEMM of the templated types with the
+ * provided parameters be provided using the supplied method?  */
+
+template<typename Top, typename Tret>
+bool method_is_compatible(GemmMethod method, GemmArgs<Tret> &args);
+
+template<typename Top, typename Tret>
+bool method_is_compatible(GemmMethod method, 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 Tret alpha, const Tret beta,
+                          const int maxthreads, const bool pretransposed_hint)
+{
+    GemmArgs<Tret> args(&ci, M, N, K, nbatches, nmulti, trA, trB, alpha, beta, maxthreads, pretransposed_hint);
+
+    return method_is_compatible<Top, Tret>(method, args);
+}
+
+/* get_gemm_method(): Given the templated types and provided parameters,
+ * which is the preferred method to implement this GEMM?  */
+template<typename Top, typename Tret>
+GemmMethod get_gemm_method(GemmArgs<Tret> &args);
+
+template<typename Top, typename Tret>
+GemmMethod get_gemm_method(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 Tret alpha, const Tret beta,
+                           const int maxthreads, const bool pretransposed_hint)
+{
+    GemmArgs<Tret> args(&ci, M, N, K, nbatches, nmulti, trA, trB, alpha, beta, maxthreads, pretransposed_hint);
+
+    return get_gemm_method<Top, Tret>(args);
+}
+
+template<typename Top, typename Tret>
+UniqueGemmCommon<Top, Tret> gemm(GemmArgs<Tret> &args, GemmConfig *cfg);
+
 /** Request an object to process a GEMM.
  *
  * @param[in]  ci                 Describes CPU properties.
@@ -47,12 +138,18 @@ using UniqueGemmCommon = std::unique_ptr<GemmCommon<Top, Tret> >;
  * @param[in]  beta               Scalar multiplier to apply to input C matrix before adding product.
  * @param[in]  maxthreads         Maximum (and default) number of threads that will call execute method.
  * @param[in]  pretransposed_hint Can the B tensor can be pretransposed (ie shared across invocations)?
+ * @param[in]  cfg                (optional) configuration parameters
  */
 template<typename Top, typename Tret>
 UniqueGemmCommon<Top, Tret> gemm(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 Tret alpha, const Tret beta,
-                                 const int maxthreads, const bool pretransposed_hint);
+                                 const int maxthreads, const bool pretransposed_hint, GemmConfig *cfg=nullptr)
+{
+    GemmArgs<Tret> args(&ci, M, N, K, nbatches, nmulti, trA, trB, alpha, beta, maxthreads, pretransposed_hint);
+
+    return gemm<Top, Tret>(args, cfg);
+}
 
 } // namespace arm_gemm