Implement 1D Adaptive Workload Splitting in CPPScheduler

Resolves COMPMID-4649
Change-Id: I941d2f8a40737ff05c49f6695a42884731ef2dc9
Signed-off-by: Dana Zlotnik <dana.zlotnik@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6656
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

7 files changed, 71 insertions, 22 deletions

diff --git a/src/cpu/kernels/CpuActivationKernel.cpp b/src/cpu/kernels/CpuActivationKernel.cpp
index 70ab06fc8a..4a6468d022 100644
--- a/src/cpu/kernels/CpuActivationKernel.cpp
+++ b/src/cpu/kernels/CpuActivationKernel.cpp
@@ -232,9 +232,16 @@ Status CpuActivationKernel::validate(const ITensorInfo *src, const ITensorInfo *
 
 size_t CpuActivationKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
 {
-    ARM_COMPUTE_UNUSED(platform, thread_count);
-
-    return ICPPKernel::small_network_mws;
+    ARM_COMPUTE_UNUSED(thread_count);
+    // Tuning results that gave optimized results in performance investigation 
+    if (platform.get_cpu_model() == CPUModel::A73 ) 
+    {
+        return 10240;
+    }
+    else 
+    {
+        return 9216;
+    }
 }
 
 void CpuActivationKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
diff --git a/src/cpu/kernels/CpuAddKernel.cpp b/src/cpu/kernels/CpuAddKernel.cpp
index 77aa1ffb3f..73c1fda711 100644
--- a/src/cpu/kernels/CpuAddKernel.cpp
+++ b/src/cpu/kernels/CpuAddKernel.cpp
@@ -294,9 +294,16 @@ const char *CpuAddKernel::name() const
 
 size_t CpuAddKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
 {
-    ARM_COMPUTE_UNUSED(platform, thread_count);
-
-    return ICPPKernel::small_network_mws;
+    ARM_COMPUTE_UNUSED(thread_count);
+    // Tuning results that gave optimized results in performance investigation 
+    if (platform.get_cpu_model() == CPUModel::A73 ) 
+    {
+        return 10240;
+    }
+    else 
+    {
+        return 9216;
+    }
 }
 
 } // namespace kernels
diff --git a/src/cpu/kernels/CpuIm2ColKernel.cpp b/src/cpu/kernels/CpuIm2ColKernel.cpp
index 5e3385d4ab..ecd1748a44 100644
--- a/src/cpu/kernels/CpuIm2ColKernel.cpp
+++ b/src/cpu/kernels/CpuIm2ColKernel.cpp
@@ -446,9 +446,16 @@ const char *CpuIm2ColKernel::name() const
 
 size_t CpuIm2ColKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
 {
-    ARM_COMPUTE_UNUSED(platform, thread_count);
-
-    return ICPPKernel::small_network_mws;
+    ARM_COMPUTE_UNUSED(thread_count);
+    // Tuning results that gave optimized results in performance investigation 
+    if (platform.get_cpu_model() == CPUModel::A73 ) 
+    {
+        return 10240;
+    }
+    else 
+    {
+        return 9216;
+    }
 }
 } // namespace kernels
 } // namespace cpu
diff --git a/src/cpu/kernels/CpuReshapeKernel.cpp b/src/cpu/kernels/CpuReshapeKernel.cpp
index 91c549643f..e19707dd0c 100644
--- a/src/cpu/kernels/CpuReshapeKernel.cpp
+++ b/src/cpu/kernels/CpuReshapeKernel.cpp
@@ -137,9 +137,16 @@ const char *CpuReshapeKernel::name() const
 
 size_t CpuReshapeKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
 {
-    ARM_COMPUTE_UNUSED(platform, thread_count);
-
-    return ICPPKernel::small_network_mws;
+    ARM_COMPUTE_UNUSED(thread_count);
+    // Tuning results that gave optimized results in performance investigation 
+    if (platform.get_cpu_model() == CPUModel::A73 ) 
+    {
+        return 10240;
+    }
+    else 
+    {
+        return 9216;
+    }
 }
 
 } // namespace kernels
diff --git a/src/cpu/kernels/assembly/CpuGemmAssemblyWrapperKernel.h b/src/cpu/kernels/assembly/CpuGemmAssemblyWrapperKernel.h
index ff8b0b143f..47548b2538 100644
--- a/src/cpu/kernels/assembly/CpuGemmAssemblyWrapperKernel.h
+++ b/src/cpu/kernels/assembly/CpuGemmAssemblyWrapperKernel.h
@@ -120,13 +120,20 @@ public:
      * @param[in] platform     The CPU platform used to create the context.
      * @param[in] thread_count Number of threads in the execution.
      *
-     * @return[out] small_network_mws         Minimum workload size for requsted configuration.
+     * @return[out] small_network_mws         Minimum workload size for requested configuration.
      */
     size_t get_mws(const CPUInfo &platform, size_t thread_count) const override
     {
-        ARM_COMPUTE_UNUSED(platform, thread_count);
-
-        return ICPPKernel::small_network_mws;
+        ARM_COMPUTE_UNUSED(thread_count);
+        // Tuning results that gave optimized results in performance investigation 
+        if (platform.get_cpu_model() == CPUModel::A73 ) 
+        {
+            return 3072;
+        }
+        else 
+        {
+            return 4096;
+        }
     }
 
 private:
diff --git a/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp
index a71864c10c..934e38b054 100644
--- a/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp
+++ b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp
@@ -357,9 +357,16 @@ const char *CpuDepthwiseConv2dAssemblyWrapperKernel::name() const
 
 size_t CpuDepthwiseConv2dAssemblyWrapperKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
 {
-    ARM_COMPUTE_UNUSED(platform, thread_count);
-
-    return ICPPKernel::small_network_mws;
+    ARM_COMPUTE_UNUSED(thread_count);
+    // Tuning results that gave optimized results in performance investigation 
+    if (platform.get_cpu_model() == CPUModel::A73 ) 
+    {
+        return 10240;
+    }
+    else 
+    {
+        return 9216;
+    }
 }
 } // namespace kernels
 } // namespace cpu
diff --git a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp
index f9c11fd4bd..78ac134604 100644
--- a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp
+++ b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp
@@ -277,9 +277,16 @@ void CpuPool2dAssemblyWrapperKernel::create_arm_pooling_requant(const ITensorInf
 
 size_t CpuPool2dAssemblyWrapperKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
 {
-    ARM_COMPUTE_UNUSED(platform, thread_count);
-
-    return ICPPKernel::small_network_mws;
+    ARM_COMPUTE_UNUSED(thread_count);
+    // Tuning results that gave optimized results in performance investigation 
+    if (platform.get_cpu_model() == CPUModel::A73 ) 
+    {
+        return 10240;
+    }
+    else 
+    {
+        return 9216;
+    }
 }
 } // namespace kernels
 } // namespace cpu