Implement Minimum Workload Size (MWS) in all CPPKernels used by small networks

 * create get_mws method in ICPPKernel class that retuns default value for all kernels
 * overwrite the default value for all the kernels used by small networks (according to banchmark case)

Resolves COMPMID-4648

Change-Id: I46d7cae61217213279d2ee740edc73f600b6d576
Signed-off-by: Dana Zlotnik <dana.zlotnik@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6412
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

16 files changed, 145 insertions, 1 deletions

diff --git a/arm_compute/core/CPP/ICPPKernel.h b/arm_compute/core/CPP/ICPPKernel.h
index ab369ffe1d..af4a896a6c 100644
--- a/arm_compute/core/CPP/ICPPKernel.h
+++ b/arm_compute/core/CPP/ICPPKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020 Arm Limited.
+ * Copyright (c) 2016-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -38,6 +38,9 @@ class ITensor;
 class ICPPKernel : public IKernel
 {
 public:
+    static constexpr size_t default_mws       = 128; /* Default minimum workload size value */
+    static constexpr size_t small_network_mws = 256; /* Default Minimum workload size value for small networks */
+
     /** Default destructor */
     virtual ~ICPPKernel() = default;
 
@@ -88,6 +91,20 @@ public:
         ARM_COMPUTE_UNUSED(tensors, window, info);
     }
 
+    /** Return minimum workload size of the relevant kernel
+     *
+     * @param[in] platform     The CPU platform used to create the context.
+     * @param[in] thread_count Number of threads in the execution.
+     *
+     * @return[out] mws         Minimum workload size for requsted configuration.
+     */
+    virtual size_t get_mws(const CPUInfo &platform, size_t thread_count) const
+    {
+        ARM_COMPUTE_UNUSED(platform, thread_count);
+
+        return default_mws;
+    }
+
     /** Name of the kernel
      *
      * @return Kernel name
diff --git a/src/core/NEON/kernels/NEPadLayerKernel.cpp b/src/core/NEON/kernels/NEPadLayerKernel.cpp
index 3e2c57a18c..60986812be 100644
--- a/src/core/NEON/kernels/NEPadLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEPadLayerKernel.cpp
@@ -258,4 +258,12 @@ void NEPadLayerKernel::run(const Window &window, const ThreadInfo &info)
         (this->*_func)(window);
     }
 }
+
+size_t NEPadLayerKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
+{
+    ARM_COMPUTE_UNUSED(platform, thread_count);
+
+    return ICPPKernel::small_network_mws;
+}
+
 } // namespace arm_compute
diff --git a/src/core/NEON/kernels/NEPadLayerKernel.h b/src/core/NEON/kernels/NEPadLayerKernel.h
index 00cda7dc22..b3b0725af8 100644
--- a/src/core/NEON/kernels/NEPadLayerKernel.h
+++ b/src/core/NEON/kernels/NEPadLayerKernel.h
@@ -79,6 +79,15 @@ public:
     // Inherited methods overridden:
     void run(const Window &window, const ThreadInfo &info) override;
 
+    /** Return minimum workload size of the relevant kernel
+     *
+     * @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.
+     */
+    size_t get_mws(const CPUInfo &platform, size_t thread_count) const override;
+
 private:
     /** Template function to run the padding function with constant padding
      *
diff --git a/src/cpu/kernels/CpuActivationKernel.cpp b/src/cpu/kernels/CpuActivationKernel.cpp
index 8fa7e9525e..70ab06fc8a 100644
--- a/src/cpu/kernels/CpuActivationKernel.cpp
+++ b/src/cpu/kernels/CpuActivationKernel.cpp
@@ -230,6 +230,13 @@ Status CpuActivationKernel::validate(const ITensorInfo *src, const ITensorInfo *
     return Status{};
 }
 
+size_t CpuActivationKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
+{
+    ARM_COMPUTE_UNUSED(platform, thread_count);
+
+    return ICPPKernel::small_network_mws;
+}
+
 void CpuActivationKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
 {
     // Early exit on disabled activation
diff --git a/src/cpu/kernels/CpuActivationKernel.h b/src/cpu/kernels/CpuActivationKernel.h
index 43c266529f..8e78d86016 100644
--- a/src/cpu/kernels/CpuActivationKernel.h
+++ b/src/cpu/kernels/CpuActivationKernel.h
@@ -57,6 +57,15 @@ public:
      */
     static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const ActivationLayerInfo &act_info);
 
+    /** Return minimum workload size of the relevant kernel
+     *
+     * @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.
+     */
+    size_t get_mws(const CPUInfo &platform, size_t thread_count) const override;
+
     // Inherited methods overridden:
     void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
     const char *name() const override;
diff --git a/src/cpu/kernels/CpuAddKernel.cpp b/src/cpu/kernels/CpuAddKernel.cpp
index edbab37301..77aa1ffb3f 100644
--- a/src/cpu/kernels/CpuAddKernel.cpp
+++ b/src/cpu/kernels/CpuAddKernel.cpp
@@ -291,6 +291,14 @@ const char *CpuAddKernel::name() const
 {
     return _name.c_str();
 }
+
+size_t CpuAddKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
+{
+    ARM_COMPUTE_UNUSED(platform, thread_count);
+
+    return ICPPKernel::small_network_mws;
+}
+
 } // namespace kernels
 } // namespace cpu
 } // namespace arm_compute
diff --git a/src/cpu/kernels/CpuAddKernel.h b/src/cpu/kernels/CpuAddKernel.h
index 11c0f67132..a0c7e497dd 100644
--- a/src/cpu/kernels/CpuAddKernel.h
+++ b/src/cpu/kernels/CpuAddKernel.h
@@ -70,6 +70,15 @@ public:
     void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
     const char *name() const override;
 
+    /** Return minimum workload size of the relevant kernel
+     *
+     * @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.
+     */
+    size_t get_mws(const CPUInfo &platform, size_t thread_count) const override;
+
 private:
     using AddKernelPtr = std::add_pointer<void(const ITensor *, const ITensor *, ITensor *, const ConvertPolicy &, const Window &)>::type;
 
diff --git a/src/cpu/kernels/CpuIm2ColKernel.cpp b/src/cpu/kernels/CpuIm2ColKernel.cpp
index 13764c49d1..5e3385d4ab 100644
--- a/src/cpu/kernels/CpuIm2ColKernel.cpp
+++ b/src/cpu/kernels/CpuIm2ColKernel.cpp
@@ -443,6 +443,13 @@ const char *CpuIm2ColKernel::name() const
 {
     return "CpuIm2ColKernel";
 }
+
+size_t CpuIm2ColKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
+{
+    ARM_COMPUTE_UNUSED(platform, thread_count);
+
+    return ICPPKernel::small_network_mws;
+}
 } // namespace kernels
 } // namespace cpu
 } // namespace arm_compute
\ No newline at end of file
diff --git a/src/cpu/kernels/CpuIm2ColKernel.h b/src/cpu/kernels/CpuIm2ColKernel.h
index fc8ae056bb..797d54c95c 100644
--- a/src/cpu/kernels/CpuIm2ColKernel.h
+++ b/src/cpu/kernels/CpuIm2ColKernel.h
@@ -91,6 +91,14 @@ public:
     // Inherited methods overridden:
     void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
     const char *name() const override;
+    /** Return minimum workload size of the relevant kernel
+     *
+     * @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.
+     */
+    size_t get_mws(const CPUInfo &platform, size_t thread_count) const override;
 
 private:
     /** Template function to run im2col
diff --git a/src/cpu/kernels/CpuReshapeKernel.cpp b/src/cpu/kernels/CpuReshapeKernel.cpp
index 3bbcc09cc5..91c549643f 100644
--- a/src/cpu/kernels/CpuReshapeKernel.cpp
+++ b/src/cpu/kernels/CpuReshapeKernel.cpp
@@ -134,6 +134,14 @@ const char *CpuReshapeKernel::name() const
 {
     return "CpuReshapeKernel";
 }
+
+size_t CpuReshapeKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
+{
+    ARM_COMPUTE_UNUSED(platform, thread_count);
+
+    return ICPPKernel::small_network_mws;
+}
+
 } // namespace kernels
 } // namespace cpu
 } // namespace arm_compute
diff --git a/src/cpu/kernels/CpuReshapeKernel.h b/src/cpu/kernels/CpuReshapeKernel.h
index 9fe4350445..d4e2b44b54 100644
--- a/src/cpu/kernels/CpuReshapeKernel.h
+++ b/src/cpu/kernels/CpuReshapeKernel.h
@@ -57,6 +57,15 @@ public:
     // Inherited methods overridden:
     void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override;
     const char *name() const override;
+
+    /** Return minimum workload size of the relevant kernel
+     *
+     * @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.
+     */
+    size_t get_mws(const CPUInfo &platform, size_t thread_count) const override;
 };
 } // namespace kernels
 } // namespace cpu
diff --git a/src/cpu/kernels/assembly/CpuGemmAssemblyWrapperKernel.h b/src/cpu/kernels/assembly/CpuGemmAssemblyWrapperKernel.h
index 3b9a6b4760..ff8b0b143f 100644
--- a/src/cpu/kernels/assembly/CpuGemmAssemblyWrapperKernel.h
+++ b/src/cpu/kernels/assembly/CpuGemmAssemblyWrapperKernel.h
@@ -115,6 +115,19 @@ public:
             _name += "/" + kernel_name_tag;
         }
     }
+    /** Return minimum workload size of the relevant kernel
+     *
+     * @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.
+     */
+    size_t get_mws(const CPUInfo &platform, size_t thread_count) const override
+    {
+        ARM_COMPUTE_UNUSED(platform, thread_count);
+
+        return ICPPKernel::small_network_mws;
+    }
 
 private:
     arm_gemm::GemmCommon<TypeInput, TypeOutput> *_kernel;
diff --git a/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp
index eed4bb9dd5..a71864c10c 100644
--- a/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp
+++ b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.cpp
@@ -354,6 +354,13 @@ const char *CpuDepthwiseConv2dAssemblyWrapperKernel::name() const
 {
     return "CpuDepthwiseConv2dAssemblyWrapperKernel";
 }
+
+size_t CpuDepthwiseConv2dAssemblyWrapperKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
+{
+    ARM_COMPUTE_UNUSED(platform, thread_count);
+
+    return ICPPKernel::small_network_mws;
+}
 } // namespace kernels
 } // namespace cpu
 } // namespace arm_compute
diff --git a/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.h b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.h
index 8ee24a6613..8980922945 100644
--- a/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.h
+++ b/src/cpu/kernels/internal/CpuDepthwiseConv2dAssemblyWrapperKernel.h
@@ -108,6 +108,15 @@ public:
      */
     bool is_configured() const;
 
+    /** Return minimum workload size of the relevant kernel
+     *
+     * @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.
+     */
+    size_t get_mws(const CPUInfo &platform, size_t thread_count) const override;
+
 private:
     std::unique_ptr<arm_conv::depthwise::IDepthwiseCommon> _kernel_asm;
     std::vector<int32_t>                                   _multipliers{};
diff --git a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp
index 958c04b677..f9c11fd4bd 100644
--- a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp
+++ b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.cpp
@@ -274,6 +274,13 @@ void CpuPool2dAssemblyWrapperKernel::create_arm_pooling_requant(const ITensorInf
 
     _kernel_asm = std::move(pooling_kernel_asm);
 }
+
+size_t CpuPool2dAssemblyWrapperKernel::get_mws(const CPUInfo &platform, size_t thread_count) const
+{
+    ARM_COMPUTE_UNUSED(platform, thread_count);
+
+    return ICPPKernel::small_network_mws;
+}
 } // namespace kernels
 } // namespace cpu
 } // namespace arm_compute
diff --git a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.h b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.h
index ab3ed25b1e..8625fd96b4 100644
--- a/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.h
+++ b/src/cpu/kernels/internal/CpuPool2dAssemblyWrapperKernel.h
@@ -112,6 +112,15 @@ private:
     void create_arm_pooling_requant(const ITensorInfo *src, ITensorInfo *dst, const PoolingLayerInfo &info, const CPUInfo &cpu_info);
 
     std::unique_ptr<arm_conv::pooling::IPoolingCommon> _kernel_asm{ nullptr };
+
+    /** Return minimum workload size of the relevant kernel
+     *
+     * @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.
+     */
+    size_t get_mws(const CPUInfo &platform, size_t thread_count) const override;
 };
 } // namespace kernels
 } // namespace cpu