Add GEMM heuristic for Mali-G78

- Replace std::map with a basic container with std::array

Change-Id: I76f53ca61676ca0e5136ce61a3f3adb10e22b4c3
Signed-off-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5441
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 379 insertions, 120 deletions

diff --git a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyBifrost.cpp b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyBifrost.cpp
index 7e6ef72f34..3645a0e141 100644
--- a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyBifrost.cpp
+++ b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyBifrost.cpp
@@ -31,7 +31,6 @@
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "src/core/CL/gemm/CLGEMMHelpers.h"
 
-#include <map>
 #include <utility>
 
 namespace arm_compute
@@ -50,89 +49,42 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
     using ConfigurationFunctionExecutorPtr = std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> (CLGEMMDefaultConfigReshapedRHSOnlyBifrost::*)(unsigned int m, unsigned int n, unsigned int k,
                                              unsigned int b);
 
-    // Configurations for Mali-G51
-    static std::map<DataType, ConfigurationFunctionExecutorPtr> gemm_configs_G51 =
-    {
-        { DataType::F32, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G51_f32 },
-        { DataType::F16, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G51_f16 },
-        { DataType::QASYMM8, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G51_u8 },
-        { DataType::QSYMM8, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G51_u8 },
-        { DataType::QASYMM8_SIGNED, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G51_u8 },
-        { DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G51_u8 }
-    };
-
-    // Configurations for Mali-G52
-    static std::map<DataType, ConfigurationFunctionExecutorPtr> gemm_configs_G52 =
-    {
-        { DataType::F32, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G52_f32 },
-        { DataType::F16, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G52_f16 },
-        { DataType::QASYMM8, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_u8 },
-        { DataType::QSYMM8, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_u8 },
-        { DataType::QASYMM8_SIGNED, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_u8 },
-        { DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_u8 }
-    };
-
-    // Configurations for Mali-G76
-    static std::map<DataType, ConfigurationFunctionExecutorPtr> gemm_configs_G76 =
-    {
-        { DataType::F32, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G76_f32 },
-        { DataType::F16, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G76_f16 },
-        { DataType::QASYMM8, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G76_u8 },
-        { DataType::QSYMM8, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G76_u8 },
-        { DataType::QASYMM8_SIGNED, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G76_u8 },
-        { DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G76_u8 }
-    };
-
-    // Configurations for Mali-G7x
-    static std::map<DataType, ConfigurationFunctionExecutorPtr> gemm_configs_G7x =
-    {
-        { DataType::F32, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_f32 },
-        { DataType::F16, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_f16 },
-        { DataType::QASYMM8, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_u8 },
-        { DataType::QSYMM8, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_u8 },
-        { DataType::QASYMM8_SIGNED, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_u8 },
-        { DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_u8 }
-    };
+    CLGEMMConfigArray<ConfigurationFunctionExecutorPtr> configs_G51(&CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G51_f32,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G51_f16,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G51_u8);
+
+    CLGEMMConfigArray<ConfigurationFunctionExecutorPtr> configs_G52(&CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G52_f32,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G52_f16,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_u8);
+
+    CLGEMMConfigArray<ConfigurationFunctionExecutorPtr> configs_G76(&CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G76_f32,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G76_f16,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G76_u8);
+
+    CLGEMMConfigArray<ConfigurationFunctionExecutorPtr> configs_G7x(&CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_f32,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_f16,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_u8);
+
+    ConfigurationFunctionExecutorPtr func = nullptr;
 
     switch(_target)
     {
         case GPUTarget::G76:
-            if(gemm_configs_G76.find(data_type) != gemm_configs_G76.end())
-            {
-                return (this->*gemm_configs_G76[data_type])(m, n, k, b);
-            }
-            else
-            {
-                ARM_COMPUTE_ERROR("Not supported data type");
-            }
-        case GPUTarget::G52:
-            if(gemm_configs_G52.find(data_type) != gemm_configs_G52.end())
-            {
-                return (this->*gemm_configs_G52[data_type])(m, n, k, b);
-            }
-            else
-            {
-                ARM_COMPUTE_ERROR("Not supported data type");
-            }
+            func = configs_G76.get_function(data_type);
+            break;
         case GPUTarget::G51:
-            if(gemm_configs_G51.find(data_type) != gemm_configs_G51.end())
-            {
-                return (this->*gemm_configs_G51[data_type])(m, n, k, b);
-            }
-            else
-            {
-                ARM_COMPUTE_ERROR("Not supported data type");
-            }
+            func = configs_G51.get_function(data_type);
+            break;
+        case GPUTarget::G52:
+            func = configs_G52.get_function(data_type);
+            break;
         default:
-            if(gemm_configs_G7x.find(data_type) != gemm_configs_G7x.end())
-            {
-                return (this->*gemm_configs_G7x[data_type])(m, n, k, b);
-            }
-            else
-            {
-                ARM_COMPUTE_ERROR("Not supported data type");
-            }
+            func = configs_G7x.get_function(data_type);
+            break;
     }
+
+    ARM_COMPUTE_ERROR_ON_MSG(func == nullptr, "Data type not support for GEMM");
+    return (this->*func)(m, n, k, b);
 }
 
 std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOnlyBifrost::configure_G7x_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
diff --git a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyValhall.cpp b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyValhall.cpp
index a0d38e44ae..a3f0509eda 100644
--- a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyValhall.cpp
+++ b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyValhall.cpp
@@ -31,7 +31,6 @@
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "src/core/CL/gemm/CLGEMMHelpers.h"
 
-#include <map>
 #include <utility>
 
 namespace arm_compute
@@ -50,30 +49,29 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
     using ConfigurationFunctionExecutorPtr = std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> (CLGEMMDefaultConfigReshapedRHSOnlyValhall::*)(unsigned int m, unsigned int n, unsigned int k,
                                              unsigned int b);
 
-    // Configurations for Mali-G77
-    static std::map<DataType, ConfigurationFunctionExecutorPtr> gemm_configs_G77 =
-    {
-        { DataType::F32, &CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G77_f32 },
-        { DataType::F16, &CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G77_f16 },
-        { DataType::QASYMM8, &CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G77_u8 },
-        { DataType::QSYMM8, &CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G77_u8 },
-        { DataType::QASYMM8_SIGNED, &CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G77_u8 },
-        { DataType::QSYMM8_PER_CHANNEL, &CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G77_u8 }
-    };
+    CLGEMMConfigArray<ConfigurationFunctionExecutorPtr> configs_G77(&CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G77_f32,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G77_f16,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G77_u8);
+
+    CLGEMMConfigArray<ConfigurationFunctionExecutorPtr> configs_G78(&CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G78_f32,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G78_f16,
+                                                                    &CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G77_u8);
+
+    ConfigurationFunctionExecutorPtr func = nullptr;
 
     switch(_target)
     {
+        case GPUTarget::G78:
+            func = configs_G78.get_function(data_type);
+            break;
         case GPUTarget::G77:
         default:
-            if(gemm_configs_G77.find(data_type) != gemm_configs_G77.end())
-            {
-                return (this->*gemm_configs_G77[data_type])(m, n, k, b);
-            }
-            else
-            {
-                ARM_COMPUTE_ERROR("Not supported data type");
-            }
+            func = configs_G77.get_function(data_type);
+            break;
     }
+
+    ARM_COMPUTE_ERROR_ON_MSG(func == nullptr, "Data type not support for GEMM");
+    return (this->*func)(m, n, k, b);
 }
 
 std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G77_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
@@ -93,8 +91,8 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
                 GEMMRHSMatrixInfo rhs_info_img;
 
                 const unsigned int h0 = std::max(n / 4, 1U);
-                std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 1, 4, 8, 1, 16, false, true, false, false, true);
-                std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 1, 4, 4, 1, h0, false, true, false, true, false);
+                std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 1, 4, 8, 1, 16, 0, 1, 0, 0, 1);
+                std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 1, 4, 4, 1, h0, 0, 1, 0, 1, 0);
 
                 return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img),
                                            std::make_pair(lhs_info_buf, rhs_info_buf),
@@ -102,18 +100,18 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
             }
             else
             {
-                return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, 8, false, true, false, false, false);
+                return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, 8, 0, 1, 0, 0, 0);
             }
         }
         else
         {
             if(r_mk <= 0.020312500186264515)
             {
-                return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, 4, false, true, false, false, false);
+                return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, 4, 0, 1, 0, 0, 0);
             }
             else
             {
-                return configure_lhs_rhs_info(m, n, 1, 4, 16, 1, 16, false, true, false, true, false);
+                return configure_lhs_rhs_info(m, n, 1, 4, 16, 1, 16, 0, 1, 0, 1, 0);
             }
         }
     }
@@ -127,7 +125,7 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
         {
             if(workload <= 747.1999816894531)
             {
-                return configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, false, true, false, true, false);
+                return configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, 0, 1, 0, 1, 0);
             }
             else
             {
@@ -135,8 +133,8 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
                 GEMMRHSMatrixInfo rhs_info_buf;
                 GEMMLHSMatrixInfo lhs_info_img;
                 GEMMRHSMatrixInfo rhs_info_img;
-                std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 2, 4, 8, 1, 2, false, false, false, true, true);
-                std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, false, true, false, true, false);
+                std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 2, 4, 8, 1, 2, 0, 0, 0, 1, 1);
+                std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, 0, 1, 0, 1, 0);
 
                 return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img),
                                            std::make_pair(lhs_info_buf, rhs_info_buf),
@@ -149,7 +147,7 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
             {
                 if(r_mk <= 0.028125000186264515)
                 {
-                    return configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, false, true, false, true, false);
+                    return configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, 0, 1, 0, 1, 0);
                 }
                 else
                 {
@@ -157,8 +155,8 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
                     GEMMRHSMatrixInfo rhs_info_buf;
                     GEMMLHSMatrixInfo lhs_info_img;
                     GEMMRHSMatrixInfo rhs_info_img;
-                    std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 2, 4, 8, 1, 2, false, false, false, true, true);
-                    std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, false, true, false, true, false);
+                    std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 2, 4, 8, 1, 2, 0, 0, 0, 1, 1);
+                    std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 8, 0, 1, 0, 1, 0);
 
                     return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img),
                                                std::make_pair(lhs_info_buf, rhs_info_buf),
@@ -171,8 +169,8 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
                 GEMMRHSMatrixInfo rhs_info_buf;
                 GEMMLHSMatrixInfo lhs_info_img;
                 GEMMRHSMatrixInfo rhs_info_img;
-                std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 2, false, true, false, false, true);
-                std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 16, false, true, false, true, false);
+                std::tie(lhs_info_img, rhs_info_img) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 2, 0, 1, 0, 0, 1);
+                std::tie(lhs_info_buf, rhs_info_buf) = configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 16, 0, 1, 0, 1, 0);
 
                 return select_lhs_rhs_info(std::make_pair(lhs_info_img, rhs_info_img),
                                            std::make_pair(lhs_info_buf, rhs_info_buf),
@@ -192,11 +190,11 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
         const unsigned int h0 = std::max(n / 2, 1U);
         if(n <= 836.0)
         {
-            return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, h0, false, true, false, true, false);
+            return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, h0, 0, 1, 0, 1, 0);
         }
         else
         {
-            return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, false, true, false, true, false);
+            return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, h0, 0, 1, 0, 1, 0);
         }
     }
     else if(m < 128)
@@ -204,11 +202,11 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
         const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(256)), static_cast<int>(1));
         if(k >= 512)
         {
-            return configure_lhs_rhs_info(m, n, 2, 4, 16, 1, h0, false, true, false, false);
+            return configure_lhs_rhs_info(m, n, 2, 4, 16, 1, h0, 0, 1, 0, 0);
         }
         else
         {
-            return configure_lhs_rhs_info(m, n, 2, 4, 8, 1, h0, false, true, false, false);
+            return configure_lhs_rhs_info(m, n, 2, 4, 8, 1, h0, 0, 1, 0, 0);
         }
     }
     else
@@ -216,17 +214,17 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
         const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(256)), static_cast<int>(1));
         if(n >= 64)
         {
-            return configure_lhs_rhs_info(m, n, 4, 8, 4, 1, h0, false, true, false, false);
+            return configure_lhs_rhs_info(m, n, 4, 8, 4, 1, h0, 0, 1, 0, 0);
         }
         else
         {
             if(k >= 512)
             {
-                return configure_lhs_rhs_info(m, n, 2, 4, 16, 1, h0, false, true, false, false);
+                return configure_lhs_rhs_info(m, n, 2, 4, 16, 1, h0, 0, 1, 0, 0);
             }
             else
             {
-                return configure_lhs_rhs_info(m, n, 2, 4, 8, 1, h0, false, true, false, false);
+                return configure_lhs_rhs_info(m, n, 2, 4, 8, 1, h0, 0, 1, 0, 0);
             }
         }
     }
@@ -240,18 +238,325 @@ std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOn
     if(m == 1)
     {
         const unsigned int h0 = std::max(n / 2, 1U);
-        return configure_lhs_rhs_info(m, n, 1, 4, 16, 1, h0, false, true, false, true);
+        return configure_lhs_rhs_info(m, n, 1, 4, 16, 1, h0, 0, 1, 0, 1);
     }
     else
     {
         const int h0 = std::max(std::min(static_cast<int>(n / 4), static_cast<int>(256)), static_cast<int>(1));
         if(m >= 28)
         {
-            return configure_lhs_rhs_info(m, n, 4, 4, 16, 1, h0, false, true, false, true);
+            return configure_lhs_rhs_info(m, n, 4, 4, 16, 1, h0, 0, 1, 0, 1);
+        }
+        else
+        {
+            return configure_lhs_rhs_info(m, n, 2, 4, 16, 1, h0, 0, 1, 0, 1);
+        }
+    }
+}
+
+std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G78_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
+{
+    const float r_mn     = static_cast<float>(m) / static_cast<float>(n);
+    const float r_mk     = static_cast<float>(m) / static_cast<float>(k);
+    const float r_nk     = static_cast<float>(n) / static_cast<float>(k);
+    const float workload = (static_cast<float>(m) * static_cast<float>(n) * static_cast<float>(b)) / 20.0f;
+
+    if(m == 1)
+    {
+        if(workload <= 278.7000f)
+        {
+            if(workload <= 7.5000f)
+            {
+                return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, 2, 0, 1, 1, 0, 0);
+            }
+            else
+            {
+                if(r_mn <= 0.0031f)
+                {
+                    if(workload <= 256.6000f)
+                    {
+                        if(workload <= 16.7500f)
+                        {
+                            if(r_nk <= 1.6671f)
+                            {
+                                return configure_lhs_rhs_info(m, n, 1, 2, 2, 1, 32, 0, 0, 0, 1, 0);
+                            }
+                            else
+                            {
+                                return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, 2, 0, 1, 1, 0, 0);
+                            }
+                        }
+                        else
+                        {
+                            return configure_lhs_rhs_info(m, n, 1, 2, 2, 1, 32, 0, 0, 0, 1, 0);
+                        }
+                    }
+                    else
+                    {
+                        return configure_lhs_rhs_info(m, n, 1, 2, 2, 1, 32, 0, 0, 0, 1, 0);
+                    }
+                }
+                else
+                {
+                    if(r_mk <= 0.0027f)
+                    {
+                        if(r_mk <= 0.0014f)
+                        {
+                            return configure_lhs_rhs_info(m, n, 1, 2, 2, 1, 32, 0, 0, 0, 1, 0);
+                        }
+                        else
+                        {
+                            if(workload <= 8.9500f)
+                            {
+                                return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, 2, 0, 1, 1, 0, 0);
+                            }
+                            else
+                            {
+                                return configure_lhs_rhs_info(m, n, 1, 2, 2, 1, 32, 0, 0, 0, 1, 0);
+                            }
+                        }
+                    }
+                    else
+                    {
+                        if(workload <= 14.1500f)
+                        {
+                            return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, 2, 0, 1, 1, 0, 0);
+                        }
+                        else
+                        {
+                            if(r_mk <= 0.0041f)
+                            {
+                                return configure_lhs_rhs_info(m, n, 1, 2, 2, 1, 32, 0, 0, 0, 1, 0);
+                            }
+                            else
+                            {
+                                return configure_lhs_rhs_info(m, n, 1, 2, 8, 1, 2, 0, 1, 1, 0, 0);
+                            }
+                        }
+                    }
+                }
+            }
+        }
+        else
+        {
+            if(workload <= 363.7000f)
+            {
+                if(r_mk <= 0.0031f)
+                {
+                    return configure_lhs_rhs_info(m, n, 1, 4, 2, 1, 32, 0, 1, 0, 1, 0);
+                }
+                else
+                {
+                    return configure_lhs_rhs_info(m, n, 1, 4, 4, 1, 32, 0, 1, 0, 1, 0);
+                }
+            }
+            else
+            {
+                return configure_lhs_rhs_info(m, n, 1, 4, 2, 1, 32, 0, 1, 0, 1, 0);
+            }
+        }
+    }
+    else
+    {
+        if(workload <= 1384.8000f)
+        {
+            if(workload <= 704.0000f)
+            {
+                return configure_lhs_rhs_info(m, n, 2, 2, 4, 1, 32, 0, 1, 0, 1, 0);
+            }
+            else
+            {
+                return configure_lhs_rhs_info(m, n, 2, 4, 8, 1, 4, 0, 0, 0, 1, 1);
+            }
+        }
+        else
+        {
+            if(workload <= 16761.6006f)
+            {
+                if(r_mn <= 187.1250f)
+                {
+                    return configure_lhs_rhs_info(m, n, 4, 4, 4, 1, 16, 0, 0, 0, 1, 1);
+                }
+                else
+                {
+                    return configure_lhs_rhs_info(m, n, 2, 4, 8, 1, 4, 0, 0, 0, 1, 1);
+                }
+            }
+            else
+            {
+                if(r_mk <= 432.4630f)
+                {
+                    return configure_lhs_rhs_info(m, n, 5, 4, 4, 1, 16, 0, 0, 0, 1, 1);
+                }
+                else
+                {
+                    return configure_lhs_rhs_info(m, n, 2, 4, 4, 1, 16, 0, 1, 0, 1, 1);
+                }
+            }
+        }
+    }
+}
+
+std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> CLGEMMDefaultConfigReshapedRHSOnlyValhall::configure_G78_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b)
+{
+    const float r_mn     = static_cast<float>(m) / static_cast<float>(n);
+    const float r_mk     = static_cast<float>(m) / static_cast<float>(k);
+    const float r_nk     = static_cast<float>(n) / static_cast<float>(k);
+    const float workload = (static_cast<float>(m) * static_cast<float>(n) * static_cast<float>(b)) / 20.0f;
+
+    if(m == 1)
+    {
+        if(r_mn <= 0.0038f)
+        {
+            if(workload <= 353.9000f)
+            {
+                if(workload <= 278.7000f)
+                {
+                    return configure_lhs_rhs_info(m, n, 1, 2, 4, 1, 32, 0, 0, 1, 0, 0);
+                }
+                else
+                {
+                    if(r_mk <= 0.0004f)
+                    {
+                        return configure_lhs_rhs_info(m, n, 1, 2, 4, 1, 32, 0, 0, 1, 0, 0);
+                    }
+                    else
+                    {
+                        if(r_mk <= 0.0030f)
+                        {
+                            return configure_lhs_rhs_info(m, n, 1, 8, 4, 1, 8, 0, 1, 1, 0, 1);
+                        }
+                        else
+                        {
+                            return configure_lhs_rhs_info(m, n, 1, 2, 4, 1, 32, 0, 0, 1, 0, 0);
+                        }
+                    }
+                }
+            }
+            else
+            {
+                if(r_nk <= 1.9384f)
+                {
+                    return configure_lhs_rhs_info(m, n, 1, 2, 4, 1, 32, 0, 0, 1, 0, 0);
+                }
+                else
+                {
+                    return configure_lhs_rhs_info(m, n, 1, 8, 4, 1, 8, 0, 1, 1, 0, 1);
+                }
+            }
         }
         else
         {
-            return configure_lhs_rhs_info(m, n, 2, 4, 16, 1, h0, false, true, false, true);
+            if(r_nk <= 1.0368f)
+            {
+                return configure_lhs_rhs_info(m, n, 1, 2, 16, 1, 32, 0, 0, 1, 0, 0);
+            }
+            else
+            {
+                return configure_lhs_rhs_info(m, n, 1, 2, 4, 1, 32, 0, 0, 1, 0, 0);
+            }
+        }
+    }
+    else
+    {
+        if(workload <= 1422.4000f)
+        {
+            if(workload <= 704.0000f)
+            {
+                return configure_lhs_rhs_info(m, n, 2, 2, 8, 1, 32, 0, 0, 1, 0, 0);
+            }
+            else
+            {
+                if(workload <= 1197.6000f)
+                {
+                    return configure_lhs_rhs_info(m, n, 2, 4, 8, 1, 8, 0, 1, 1, 0, 1);
+                }
+                else
+                {
+                    if(workload <= 1241.6000f)
+                    {
+                        return configure_lhs_rhs_info(m, n, 2, 8, 8, 1, 16, 0, 1, 1, 0, 0);
+                    }
+                    else
+                    {
+                        return configure_lhs_rhs_info(m, n, 2, 4, 8, 1, 8, 0, 1, 1, 0, 1);
+                    }
+                }
+            }
+        }
+        else
+        {
+            if(workload <= 2769.6000f)
+            {
+                if(workload <= 1846.4000f)
+                {
+                    if(r_mn <= 2.4927f)
+                    {
+                        return configure_lhs_rhs_info(m, n, 2, 8, 8, 1, 16, 0, 1, 1, 0, 0);
+                    }
+                    else
+                    {
+                        return configure_lhs_rhs_info(m, n, 4, 4, 8, 1, 32, 0, 1, 1, 0, 0);
+                    }
+                }
+                else
+                {
+                    if(r_mn <= 0.6261f)
+                    {
+                        return configure_lhs_rhs_info(m, n, 4, 4, 8, 1, 32, 0, 1, 1, 0, 0);
+                    }
+                    else
+                    {
+                        if(r_mk <= 3.4453f)
+                        {
+                            if(r_mn <= 1.4135f)
+                            {
+                                return configure_lhs_rhs_info(m, n, 2, 8, 8, 1, 16, 0, 1, 1, 0, 0);
+                            }
+                            else
+                            {
+                                return configure_lhs_rhs_info(m, n, 4, 4, 8, 1, 32, 0, 1, 1, 0, 0);
+                            }
+                        }
+                        else
+                        {
+                            return configure_lhs_rhs_info(m, n, 2, 8, 8, 1, 16, 0, 1, 1, 0, 0);
+                        }
+                    }
+                }
+            }
+            else
+            {
+                if(r_nk <= 0.0302f)
+                {
+                    return configure_lhs_rhs_info(m, n, 2, 4, 8, 1, 8, 0, 1, 1, 0, 1);
+                }
+                else
+                {
+                    if(r_mk <= 181.3750f)
+                    {
+                        return configure_lhs_rhs_info(m, n, 4, 4, 8, 1, 32, 0, 1, 1, 0, 0);
+                    }
+                    else
+                    {
+                        if(workload <= 28035.2002f)
+                        {
+                            return configure_lhs_rhs_info(m, n, 2, 8, 8, 1, 16, 0, 1, 1, 0, 0);
+                        }
+                        else
+                        {
+                            if(r_mk <= 808.6667f)
+                            {
+                                return configure_lhs_rhs_info(m, n, 4, 4, 8, 1, 32, 0, 1, 1, 0, 0);
+                            }
+                            else
+                            {
+                                return configure_lhs_rhs_info(m, n, 2, 8, 8, 1, 16, 0, 1, 1, 0, 0);
+                            }
+                        }
+                    }
+                }
+            }
         }
     }
 }
diff --git a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyValhall.h b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyValhall.h
index 94c6a43c21..a3b556c441 100644
--- a/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyValhall.h
+++ b/src/core/CL/gemm/reshaped_only_rhs/CLGEMMDefaultConfigReshapedRHSOnlyValhall.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020 Arm Limited.
+ * Copyright (c) 2020-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -46,6 +46,8 @@ public:
 private:
     std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G77_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
     std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G77_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+    std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G78_f32(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
+    std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G78_f16(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
     std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> configure_G77_u8(unsigned int m, unsigned int n, unsigned int k, unsigned int b);
 };
 } // namespace cl_gemm