Update cpu heuristics for GEMM

If available, it prefers MMLA over DOT for smallK kernels
for int8 types for faster execution.

Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Change-Id: I190142dc66555476b4d8266d22e54ce8b45fffdb
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6251
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 12 insertions, 12 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
index cfbf66d60f..8fe1993092 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp
@@ -67,7 +67,7 @@ GemmImplementation<int8_t, int32_t>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_smallK_hybrid_s8s32_dot_8x1VL",
     [](const GemmArgs &args) { return args._ci->has_svei8mm() && args._Ksize<=64 && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args) { return new GemmHybrid<cls_sve_smallK_hybrid_s8s32_dot_8x1VL, int8_t, int32_t>(args); }
 },
 GemmImplementation<int8_t, int32_t>::with_estimate(
@@ -103,14 +103,14 @@ GemmImplementation<int8_t, int32_t>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_smallK_hybrid_s8s32_dot_8x4",
     [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_s8s32_dot_8x4, int8_t, int32_t>(args); }
 },
 {
     GemmMethod::GEMM_HYBRID,
     "a64_smallK_hybrid_s8s32_dot_6x4",
     [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_s8s32_dot_6x4, int8_t, int32_t>(args); }
 },
 {
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp
index aa62815438..b2882b34ba 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_qint8.cpp
@@ -92,7 +92,7 @@ GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID_QUANTIZED,
     "sve_smallK_hybrid_s8s32_dot_8x1VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve() && args._Ksize<=64 && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args, const Requantize32 &) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<cls_sve_smallK_hybrid_s8s32_dot_8x1VL, int8_t, int8_t>(args, qp); }
 },
 GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
@@ -156,14 +156,14 @@ GemmImplementation<int8_t, int8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID_QUANTIZED,
     "a64_smallK_hybrid_s8s32_dot_8x4",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args, const Requantize32 &) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<cls_a64_smallK_hybrid_s8s32_dot_8x4, int8_t, int8_t>(args, qp); }
 },
 {
     GemmMethod::GEMM_HYBRID_QUANTIZED,
     "a64_smallK_hybrid_s8s32_dot_6x4",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args, const Requantize32 &) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<cls_a64_smallK_hybrid_s8s32_dot_6x4, int8_t, int8_t>(args, qp); }
 },
 {
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
index abd2799583..b277e56000 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_quint8.cpp
@@ -80,7 +80,7 @@ GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID_QUANTIZED,
     "sve_smallK_hybrid_u8u32_dot_8x1VL",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_sve() && args._Ksize<=64 && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args, const Requantize32 &) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<cls_sve_smallK_hybrid_u8u32_dot_8x1VL, uint8_t, uint8_t>(args, qp); }
 },
 GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
@@ -130,14 +130,14 @@ GemmImplementation<uint8_t, uint8_t, Requantize32>::with_estimate(
     GemmMethod::GEMM_HYBRID_QUANTIZED,
     "a64_smallK_hybrid_u8u32_dot_8x4",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args, const Requantize32 &) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<cls_a64_smallK_hybrid_u8u32_dot_8x4, uint8_t, uint8_t>(args, qp); }
 },
 {
     GemmMethod::GEMM_HYBRID_QUANTIZED,
     "a64_smallK_hybrid_u8u32_dot_6x4",
     [](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args, const Requantize32 &) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<cls_a64_smallK_hybrid_u8u32_dot_6x4, uint8_t, uint8_t>(args, qp); }
 },
 {
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
index 75d6b362cc..4eb2a6f624 100644
--- a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
+++ b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp
@@ -67,7 +67,7 @@ GemmImplementation<uint8_t, uint32_t>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "sve_smallK_hybrid_u8u32_dot_8x1VL",
     [](const GemmArgs &args) { return args._ci->has_sve() && args._Ksize<=64 && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args) { return new GemmHybrid<cls_sve_smallK_hybrid_u8u32_dot_8x1VL, uint8_t, uint32_t>(args); }
 },
 GemmImplementation<uint8_t, uint32_t>::with_estimate(
@@ -103,14 +103,14 @@ GemmImplementation<uint8_t, uint32_t>::with_estimate(
     GemmMethod::GEMM_HYBRID,
     "a64_smallK_hybrid_u8u32_dot_8x4",
     [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_u8u32_dot_8x4, uint8_t, uint32_t>(args); }
 },
 {
     GemmMethod::GEMM_HYBRID,
     "a64_smallK_hybrid_u8u32_dot_6x4",
     [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input; },
-    nullptr,
+    [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); },
     [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_u8u32_dot_6x4, uint8_t, uint32_t>(args); }
 },
 {