COMPMID-3776: Indirect GEMM

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I51a1b0f098bc3a8c408c50c92221e4df3061e12c
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4343
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 33 insertions, 48 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
index f3b66528a4..96b9734221 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp
@@ -25,93 +25,78 @@
 #include "bfloat.hpp"
 #include "gemm_common.hpp"
 #include "gemm_hybrid.hpp"
+#include "gemm_hybrid_indirect.hpp"
 #include "gemm_implementation.hpp"
 #include "gemm_interleaved.hpp"
 #include "gemv_batched.hpp"
 #include "gemv_pretransposed.hpp"
 
-#include "kernels/a64_interleaved_bf16fp32_dot_12x8.hpp"
-#include "kernels/a64_interleaved_bf16fp32_mmla_12x8.hpp"
-#include "kernels/a64_sgemm_12x8.hpp"
+#include "kernels/a64_hybrid_bf16fp32_dot_6x16.hpp"
+#include "kernels/a64_interleaved_bf16fp32_dot_8x12.hpp"
+#include "kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp"
+#include "kernels/a64_sgemm_8x12.hpp"
 #include "kernels/a32_sgemm_8x6.hpp"
-#include "kernels/sve_interleaved_bf16fp32_dot_3VLx8.hpp"
-#include "kernels/sve_interleaved_bf16fp32_mmla_3VLx8.hpp"
-#include "kernels/sve_hybrid_bf16fp32_dot_4VLx4.hpp"
-#include "kernels/sve_hybrid_bf16fp32_mmla_4VLx4.hpp"
-#include "kernels/sve_hybrid_bf16fp32_mmla_6VLx2.hpp"
-#include "kernels/sve_hybrid_bf16fp32_mmla_8VLx2.hpp"
+#include "kernels/sve_interleaved_bf16fp32_dot_8x3VL.hpp"
+#include "kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp"
+#include "kernels/sve_hybrid_bf16fp32_dot_6x4VL.hpp"
 
 namespace arm_gemm {
 
 static const GemmImplementation<bfloat16, float> gemm_bf16_methods[] =
 {
 #ifdef V8P6_BF
-# ifdef __ARM_FEATURE_SVE
-{
-    GemmMethod::GEMM_HYBRID,
-    "hybrid_bf16fp32_mmla_6VLx2",
-    [](const GemmArgs &args) { return (args._Ksize>=8); },
-    [](const GemmArgs &args) { return ((args._Msize <= 4) && (args._Nsize <= hybrid_bf16fp32_mmla_6VLx2::out_width())); },
-    [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_mmla_6VLx2, bfloat16, float>(args); }
-},
-{
-    GemmMethod::GEMM_HYBRID,
-    "hybrid_bf16fp32_mmla_8VLx2",
-    [](const GemmArgs &args) { return (args._Ksize>=8); },
-    [](const GemmArgs &args) { return (args._Msize <= 4); },
-    [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_mmla_8VLx2, bfloat16, float>(args); }
-},
-{
-    GemmMethod::GEMM_HYBRID,
-    "hybrid_bf16fp32_mmla_4VLx4",
-    [](const GemmArgs &args) { return (args._Ksize>=8); },
-    [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); },
-    [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_mmla_4VLx4, bfloat16, float>(args); }
-},
-{
-    GemmMethod::GEMM_HYBRID,
-    "hybrid_bf16fp32_dot_4VLx4",
-    [](const GemmArgs &args) { return (args._Ksize>=8); },
-    [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); },
-    [](const GemmArgs &args) { return new GemmHybrid<hybrid_bf16fp32_dot_4VLx4, bfloat16, float>(args); }
-},
+#ifdef __ARM_FEATURE_SVE
 { // gemm_bf16_interleaved
     GemmMethod::GEMM_INTERLEAVED,
-    "interleaved_bf16fp32_mmla_3VLx8",
+    "sve_interleaved_bf16fp32_mmla_8x3VL",
     [](const GemmArgs &args) { return (args._Ksize>4); },
     nullptr,
-    [](const GemmArgs &args) { return new GemmInterleaved<interleaved_bf16fp32_mmla_3VLx8, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, bfloat16, float>(args); }
+},
+{
+    GemmMethod::GEMM_HYBRID,
+    "sve_hybrid_bf16fp32_dot_6x4VL",
+    nullptr,
+    [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); },
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_bf16fp32_dot_6x4VL, bfloat16, float>(args); }
 },
 { // gemm_bf16_interleaved
     GemmMethod::GEMM_INTERLEAVED,
-    "interleaved_bf16fp32_dot_3VLx8",
+    "sve_interleaved_bf16fp32_dot_8x3VL",
     [](const GemmArgs &args) { return (args._Ksize>2); },
     nullptr,
-    [](const GemmArgs &args) { return new GemmInterleaved<interleaved_bf16fp32_dot_3VLx8, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_dot_8x3VL, bfloat16, float>(args); }
 },
 # endif // SVE
 { // gemm_bf16_interleaved
     GemmMethod::GEMM_INTERLEAVED,
-    "interleaved_bf16fp32_mmla_12x8",
+    "a64_interleaved_bf16fp32_mmla_8x12",
     [](const GemmArgs &args) { return (args._Ksize>4); },
     nullptr,
-    [](const GemmArgs &args) { return new GemmInterleaved<interleaved_bf16fp32_mmla_12x8, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, bfloat16, float>(args); }
+},
+{
+    GemmMethod::GEMM_HYBRID,
+    "a64_hybrid_bf16fp32_dot_6x16",
+    nullptr,
+    nullptr,
+    [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_bf16fp32_dot_6x16, bfloat16, float>(args); }
 },
 { // gemm_bf16_interleaved
     GemmMethod::GEMM_INTERLEAVED,
-    "interleaved_bf16fp32_dot_12x8",
+    "a64_interleaved_bf16fp32_dot_8x12",
     [](const GemmArgs &args) { return (args._Ksize>2); },
     nullptr,
-    [](const GemmArgs &args) { return new GemmInterleaved<interleaved_bf16fp32_dot_12x8, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_dot_8x12, bfloat16, float>(args); }
 },
 #endif // V8P6_BF
 #ifdef __aarch64__
 {
     GemmMethod::GEMM_INTERLEAVED,
-    "sgemm_12x8",
+    "a64_sgemm_8x12",
     nullptr,
     nullptr,
-    [](const GemmArgs &args) { return new GemmInterleaved<sgemm_12x8, bfloat16, float>(args); }
+    [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, bfloat16, float>(args); }
 },
 #elif defined(__arm__)
 {