COMPMID-2244: Extend CLFuseBatchNormalization to support DepthwiseConvolution weights

Change-Id: I7d1907f35cc4899379073759be2f7cce24e51e9d
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1327
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 91 insertions, 64 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index 51acd9f339..253da40077 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -298,7 +298,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "finalize", "optical_flow_pyramid_lk.cl" },
     { "flatten", "flatten.cl" },
     { "floor_layer", "floor.cl" },
-    { "fuse_batchnormalization_conv_layer", "batchnormalization_layer.cl" },
+    { "fuse_batchnormalization_layer", "batchnormalization_layer.cl" },
     { "gather", "gather.cl" },
     { "gaussian1x5_sub_x", "gaussian_pyramid.cl" },
     { "gaussian5x1_sub_y", "gaussian_pyramid.cl" },
diff --git a/src/core/CL/cl_kernels/batchnormalization_layer.cl b/src/core/CL/cl_kernels/batchnormalization_layer.cl
index a5321315d3..918caff212 100644
--- a/src/core/CL/cl_kernels/batchnormalization_layer.cl
+++ b/src/core/CL/cl_kernels/batchnormalization_layer.cl
@@ -259,12 +259,14 @@ __kernel void batchnormalization_layer_nhwc(TENSOR3D_DECLARATION(input),
 }
 #endif /* defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DATA_TYPE)*/
 
-#if defined(DIM2) && defined(DATA_TYPE) && defined(EPSILON)
-/** OpenCL kernel to fuse the weights of convolution layer with batch normalization when the data layout is either NCHW or NHWC
+#if defined(DATA_TYPE) && defined(EPSILON)
+/** OpenCL kernel to fuse the weights of convolution or depthwise convolution layer with batch normalization when the data layout is either NCHW or NHWC
  *
  * @note The input weights tensor is assumed 4D with the OFMs in the fourth dimension
  * @note Data type should be passed at compile time using the -DDATA_TYPE, e.g. -DDATA_TYPE=float
- * @note The third dimension of the input tensor should be passed at compile time using -DNUM_CHANNELS=size. e.g. -DNUM_CHANNELS=16
+ * @note The third dimension of the input tensor should be passed at compile time when weights belong to a convolution layer using -DDIM2=size. e.g. -DDIM2=16.
+ *       For depthwise convolution weight do not pass DIM2
+ * @note Data layout NHWC should be passed at compile time with -DNHWC. For data layout NCHW it is not required to pass any parameter
  * @note Batch normalization epsilon parameter should be passed at compile time using -DEPSILON=value. e.g. -DEPSILON=0.001f
  *
  * @param[in]  w_ptr                                 Pointer to the weights tensor. Supported data types: F16/F32
@@ -312,35 +314,45 @@ __kernel void batchnormalization_layer_nhwc(TENSOR3D_DECLARATION(input),
  * @param[in]  gamma_step_x                          (Optional) gamma_stride_x * number of elements along X processed per workitem(in bytes)
  * @param[in]  gamma_offset_first_element_in_bytes   (Optional) The offset of the first element in the gamma source tensor
  */
-__kernel void fuse_batchnormalization_conv_layer(TENSOR3D_DECLARATION(w),
+__kernel void fuse_batchnormalization_layer(TENSOR3D_DECLARATION(w),
 #if defined(BIAS)
-                                                 VECTOR_DECLARATION(b),
+                                            VECTOR_DECLARATION(b),
 #endif // defined(BIAS)
-                                                 VECTOR_DECLARATION(mean),
-                                                 VECTOR_DECLARATION(var)
+                                            VECTOR_DECLARATION(mean),
+                                            VECTOR_DECLARATION(var)
 #ifndef IN_PLACE_W
-                                                 ,
-                                                 TENSOR3D_DECLARATION(w_fused)
+                                            ,
+                                            TENSOR3D_DECLARATION(w_fused)
 #endif // ifndef IN_PLACE_W
 #ifndef IN_PLACE_B
-                                                 ,
-                                                 VECTOR_DECLARATION(b_fused)
+                                            ,
+                                            VECTOR_DECLARATION(b_fused)
 #endif // ifndef IN_PLACE_B
 #if defined(BETA)
-                                                 ,
-                                                 VECTOR_DECLARATION(beta)
+                                            ,
+                                            VECTOR_DECLARATION(beta)
 #endif // defined(BETA)
 #if defined(GAMMA)
-                                                 ,
-                                                 VECTOR_DECLARATION(gamma)
+                                            ,
+                                            VECTOR_DECLARATION(gamma)
 #endif // defined(GAMMA)
-                                                )
+                                           )
 {
-    int x  = get_global_id(0);
-    int y  = get_global_id(1);
-    int z  = get_global_id(2);
+    int x = get_global_id(0);
+    int y = get_global_id(1);
+    int z = get_global_id(2);
+
+#if defined(DIM2)
     int c0 = z % DIM2;
     int c1 = z / DIM2;
+#else // ! defined(DIM2)
+    int c0 = 0;
+#if defined(NHWC)
+    int c1 = x;
+#else  // defined(NHWC)
+    int c1 = z;
+#endif // defined(NHWC)
+#endif // defined(DIM2)
 
     int w_offset = x * sizeof(DATA_TYPE) + y * w_stride_y + z * w_stride_z;
     int v_offset = c1 * sizeof(DATA_TYPE);
@@ -368,11 +380,15 @@ __kernel void fuse_batchnormalization_conv_layer(TENSOR3D_DECLARATION(w),
 #if defined(IN_PLACE_W)
     *((__global DATA_TYPE *)(w_ptr + w_offset + w_offset_first_element_in_bytes)) = w_new;
 #else  // defined(IN_PLACE_W)
-    *((__global DATA_TYPE *)(w_fused_ptr + w_offset + w_fused_offset_first_element_in_bytes))     = w_new;
+    *((__global DATA_TYPE *)(w_fused_ptr + w_offset + w_fused_offset_first_element_in_bytes)) = w_new;
 #endif // defined(IN_PLACE_W)
 
     // Compute bias
+#if !defined(DIM2) && defined(NHWC)
+    if(z == 0 && y == 0)
+#else !defined(DIM2) && defined(NHWC)
     if(x == 0 && y == 0 && c0 == 0)
+#endif // !defined(DIM2) && defined(NHWC)
     {
 #if defined(BIAS)
         b_old = *((__global DATA_TYPE *)(b_ptr + v_offset + b_offset_first_element_in_bytes));
@@ -400,4 +416,4 @@ __kernel void fuse_batchnormalization_conv_layer(TENSOR3D_DECLARATION(w),
 #endif // defined(BIAS)
     }
 }
-#endif // defined(DIM2) && defined(DATA_TYPE) && defined(EPSILON)
\ No newline at end of file
+#endif // defined(DATA_TYPE) && defined(EPSILON)
\ No newline at end of file
diff --git a/src/core/CL/kernels/CLFuseBatchNormalizationKernel.cpp b/src/core/CL/kernels/CLFuseBatchNormalizationKernel.cpp
index 16ad7d970e..bf827bf2c2 100644
--- a/src/core/CL/kernels/CLFuseBatchNormalizationKernel.cpp
+++ b/src/core/CL/kernels/CLFuseBatchNormalizationKernel.cpp
@@ -38,50 +38,60 @@ namespace arm_compute
 {
 namespace
 {
-Status validate_arguments(const ITensorInfo *conv_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
+Status validate_arguments(const ITensorInfo *input_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
                           const ITensorInfo *fused_weights, const ITensorInfo *fused_bias,
-                          const ITensorInfo *conv_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
-                          float epsilon)
+                          const ITensorInfo *input_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
+                          float epsilon, FuseBatchNormalizationType fbn_type)
 {
     ARM_COMPUTE_UNUSED(epsilon);
-    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(conv_weights);
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(conv_weights, 1, DataType::F16, DataType::F32);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input_weights, bn_mean, bn_var);
+    ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input_weights);
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input_weights, 1, DataType::F16, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, bn_var);
-    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, bn_mean, bn_var);
-    ARM_COMPUTE_RETURN_ERROR_ON(conv_bias == nullptr && fused_bias == nullptr);
-    ARM_COMPUTE_RETURN_ERROR_ON(conv_weights->dimension(3) != bn_mean->dimension(0));
+    ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, bn_mean, bn_var);
+    ARM_COMPUTE_RETURN_ERROR_ON(input_bias == nullptr && fused_bias == nullptr);
     ARM_COMPUTE_RETURN_ERROR_ON(bn_mean->num_dimensions() > 1);
 
+    if(fbn_type == FuseBatchNormalizationType::CONVOLUTION)
+    {
+        ARM_COMPUTE_RETURN_ERROR_ON(input_weights->dimension(3) != bn_mean->dimension(0));
+    }
+    else
+    {
+        const size_t channel_idx = get_data_layout_dimension_index(input_weights->data_layout(), DataLayoutDimension::CHANNEL);
+        ARM_COMPUTE_RETURN_ERROR_ON(input_weights->dimension(channel_idx) != bn_mean->dimension(0));
+    }
+
     // Validate bias
-    if(conv_bias != nullptr)
+    if(input_bias != nullptr)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, conv_bias);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, conv_bias);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, input_bias);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, input_bias);
     }
     // Validate beta
     if(bn_beta != nullptr)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, bn_beta);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, bn_beta);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, bn_beta);
     }
     // Validate gamma
     if(bn_gamma != nullptr)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, bn_gamma);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, bn_gamma);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, bn_gamma);
     }
     // Validate output weights
     if(fused_weights != nullptr && fused_weights->total_size() != 0)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(conv_weights, fused_weights);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(conv_weights, fused_weights);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, fused_weights);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input_weights, fused_weights);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input_weights, fused_weights);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, fused_weights);
     }
     // Validate output bias
     if(fused_bias != nullptr && fused_bias->total_size() != 0)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(bn_mean, fused_bias);
-        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(conv_weights, fused_bias);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input_weights, fused_bias);
     }
 
     return Status{};
@@ -89,20 +99,20 @@ Status validate_arguments(const ITensorInfo *conv_weights, const ITensorInfo *bn
 } // namespace
 
 CLFuseBatchNormalizationKernel::CLFuseBatchNormalizationKernel()
-    : _conv_weights(nullptr), _conv_bias(nullptr), _bn_mean(nullptr), _bn_var(nullptr), _bn_gamma(nullptr), _bn_beta(nullptr), _fused_weights(nullptr), _fused_bias(nullptr), _epsilon(),
+    : _input_weights(nullptr), _input_bias(nullptr), _bn_mean(nullptr), _bn_var(nullptr), _bn_gamma(nullptr), _bn_beta(nullptr), _fused_weights(nullptr), _fused_bias(nullptr), _epsilon(),
       _run_in_place_weights(false), _run_in_place_bias(false)
 {
 }
 
-void CLFuseBatchNormalizationKernel::configure(const ICLTensor *conv_weights, const ICLTensor *bn_mean, const ICLTensor *bn_var,
+void CLFuseBatchNormalizationKernel::configure(const ICLTensor *input_weights, const ICLTensor *bn_mean, const ICLTensor *bn_var,
                                                ICLTensor *fused_weights, ICLTensor *fused_bias,
-                                               const ICLTensor *conv_bias, const ICLTensor *bn_beta, const ICLTensor *bn_gamma,
-                                               float epsilon)
+                                               const ICLTensor *input_bias, const ICLTensor *bn_beta, const ICLTensor *bn_gamma,
+                                               float epsilon, FuseBatchNormalizationType fbn_type)
 {
-    ARM_COMPUTE_ERROR_ON_NULLPTR(conv_weights, bn_mean, bn_var);
+    ARM_COMPUTE_ERROR_ON_NULLPTR(input_weights, bn_mean, bn_var);
 
-    _conv_weights  = conv_weights;
-    _conv_bias     = conv_bias;
+    _input_weights = input_weights;
+    _input_bias    = input_bias;
     _bn_mean       = bn_mean;
     _bn_var        = bn_var;
     _bn_beta       = bn_beta;
@@ -111,14 +121,14 @@ void CLFuseBatchNormalizationKernel::configure(const ICLTensor *conv_weights, co
     _fused_bias    = fused_bias;
     _epsilon       = epsilon;
 
-    _run_in_place_weights = (fused_weights == nullptr) || (fused_weights == conv_weights);
-    _run_in_place_bias    = (conv_bias != nullptr && fused_bias == nullptr) || (conv_bias != nullptr && fused_bias == conv_bias);
+    _run_in_place_weights = (fused_weights == nullptr) || (fused_weights == input_weights);
+    _run_in_place_bias    = (input_bias != nullptr && fused_bias == nullptr) || (input_bias != nullptr && fused_bias == input_bias);
 
     // Auto initialize outputs
     if(_fused_weights != nullptr)
     {
         // Output tensor auto initialization if not yet initialized
-        auto_init_if_empty(*_fused_weights->info(), *_conv_weights->info()->clone());
+        auto_init_if_empty(*_fused_weights->info(), *_input_weights->info()->clone());
     }
     if(_fused_bias != nullptr)
     {
@@ -127,39 +137,40 @@ void CLFuseBatchNormalizationKernel::configure(const ICLTensor *conv_weights, co
     }
 
     // Validate arguments
-    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(conv_weights->info(), bn_mean->info(), bn_var->info(),
+    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input_weights->info(), bn_mean->info(), bn_var->info(),
                                                   (fused_weights != nullptr) ? fused_weights->info() : nullptr,
                                                   (fused_bias != nullptr) ? fused_bias->info() : nullptr,
-                                                  (conv_bias != nullptr) ? conv_bias->info() : nullptr,
+                                                  (input_bias != nullptr) ? input_bias->info() : nullptr,
                                                   (bn_beta != nullptr) ? bn_beta->info() : nullptr,
                                                   (bn_gamma != nullptr) ? bn_gamma->info() : nullptr,
-                                                  epsilon));
+                                                  epsilon, fbn_type));
 
     // Configure kernel window
-    Window win = calculate_max_window(*conv_weights->info());
+    Window win = calculate_max_window(*input_weights->info());
     ICLKernel::configure_internal(win);
 
     // Set build options
     CLBuildOptions build_opts;
-    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(conv_weights->info()->data_type()));
-    build_opts.add_option("-DDIM2=" + support::cpp11::to_string(conv_weights->info()->dimension(2)));
+    build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input_weights->info()->data_type()));
+    build_opts.add_option_if(fbn_type == FuseBatchNormalizationType::CONVOLUTION, "-DDIM2=" + support::cpp11::to_string(input_weights->info()->dimension(2)));
     build_opts.add_option("-DEPSILON=" + float_to_string_with_full_precision(epsilon));
+    build_opts.add_option_if(_input_weights->info()->data_layout() == DataLayout::NHWC, "-DNHWC");
     build_opts.add_option_if(_run_in_place_weights, "-DIN_PLACE_W");
     build_opts.add_option_if(_run_in_place_bias, "-DIN_PLACE_B");
-    build_opts.add_option_if(conv_bias != nullptr, "-DBIAS");
+    build_opts.add_option_if(input_bias != nullptr, "-DBIAS");
     build_opts.add_option_if(bn_beta != nullptr, "-DBETA");
     build_opts.add_option_if(bn_gamma != nullptr, "-DGAMMA");
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("fuse_batchnormalization_conv_layer", build_opts.options()));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("fuse_batchnormalization_layer", build_opts.options()));
 }
 
-Status CLFuseBatchNormalizationKernel::validate(const ITensorInfo *conv_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
+Status CLFuseBatchNormalizationKernel::validate(const ITensorInfo *input_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var,
                                                 const ITensorInfo *fused_weights, const ITensorInfo *fused_bias,
-                                                const ITensorInfo *conv_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
-                                                float epsilon)
+                                                const ITensorInfo *input_bias, const ITensorInfo *bn_beta, const ITensorInfo *bn_gamma,
+                                                float epsilon, FuseBatchNormalizationType fbn_type)
 {
-    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(conv_weights, bn_mean, bn_var, fused_weights, fused_bias, conv_bias, bn_beta, bn_gamma, epsilon));
+    ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input_weights, bn_mean, bn_var, fused_weights, fused_bias, input_bias, bn_beta, bn_gamma, epsilon, fbn_type));
     return Status{};
 }
 
@@ -175,10 +186,10 @@ void CLFuseBatchNormalizationKernel::run(const arm_compute::Window &window, cl::
 
     // Add kernel arguments
     unsigned int idx = 0;
-    add_3D_tensor_argument(idx, _conv_weights, slice_3d);
-    if(_conv_bias != nullptr)
+    add_3D_tensor_argument(idx, _input_weights, slice_3d);
+    if(_input_bias != nullptr)
     {
-        add_1D_tensor_argument(idx, _conv_bias, slice_1d);
+        add_1D_tensor_argument(idx, _input_bias, slice_1d);
     }
     add_1D_tensor_argument(idx, _bn_mean, slice_1d);
     add_1D_tensor_argument(idx, _bn_var, slice_1d);