COMPMID-803: Add NHWC data format support for CL batch normalisation

Change-Id: Ie37588f60b9cfc7b1d09b1e8628fcfb4b17e0717
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/123834
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

4 files changed, 136 insertions, 11 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index db0e51bb48..1c773bc42f 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -151,7 +151,8 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "activation_layer_qa8", "activation_layer_qa8.cl" },
     { "arithmetic_add", "arithmetic_op.cl" },
     { "arithmetic_sub", "arithmetic_op.cl" },
-    { "batchnormalization_layer", "batchnormalization_layer.cl" },
+    { "batchnormalization_layer_nchw", "batchnormalization_layer.cl" },
+    { "batchnormalization_layer_nhwc", "batchnormalization_layer.cl" },
     { "bitwise_or", "bitwise_op.cl" },
     { "bitwise_and", "bitwise_op.cl" },
     { "bitwise_xor", "bitwise_op.cl" },
diff --git a/src/core/CL/cl_kernels/batchnormalization_layer.cl b/src/core/CL/cl_kernels/batchnormalization_layer.cl
index 29b62d3d92..9c980da62a 100644
--- a/src/core/CL/cl_kernels/batchnormalization_layer.cl
+++ b/src/core/CL/cl_kernels/batchnormalization_layer.cl
@@ -87,19 +87,19 @@
  * @param[in]  gamma_offset_first_element_in_bytes  The offset of the first element in the gamma source tensor
  * @param[in]  epsilon                              Epsilon parameter in the batch normalization equation
  */
-__kernel void batchnormalization_layer(TENSOR3D_DECLARATION(input),
+__kernel void batchnormalization_layer_nchw(TENSOR3D_DECLARATION(input),
 #ifndef IN_PLACE
-                                       TENSOR3D_DECLARATION(output),
+                                            TENSOR3D_DECLARATION(output),
 #endif /* not IN_PLACE */
-                                       VECTOR_DECLARATION(mean),
-                                       VECTOR_DECLARATION(var),
+                                            VECTOR_DECLARATION(mean),
+                                            VECTOR_DECLARATION(var),
 #ifndef USE_DEFAULT_BETA
-                                       VECTOR_DECLARATION(beta),
+                                            VECTOR_DECLARATION(beta),
 #endif /* USE_DEFAULT_BETA */
 #ifndef USE_DEFAULT_GAMMA
-                                       VECTOR_DECLARATION(gamma),
+                                            VECTOR_DECLARATION(gamma),
 #endif /* USE_DEFAULT_GAMMA */
-                                       float epsilon)
+                                            float epsilon)
 {
     Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input);
 #ifdef IN_PLACE
@@ -145,7 +145,7 @@ __kernel void batchnormalization_layer(TENSOR3D_DECLARATION(input),
     res = MUL_OP(gamma_vec, x_bar);
 #else  /* USE_DEFAULT_GAMMA */
     // gamma is equal to 1, no need to perform multiplications
-    res = x_bar;
+    res          = x_bar;
 #endif /* USE_DEFAULT_GAMMA */
 
 #ifndef USE_DEFAULT_BETA
@@ -161,4 +161,113 @@ __kernel void batchnormalization_layer(TENSOR3D_DECLARATION(input),
     (res, 0, (__global DATA_TYPE *)out.ptr);
 }
 
+/** Apply batch normalization on tensors with NHWC format.
+ *
+ * @param[in]  input_ptr                            Pointer to the first source tensor. Supported data types: QS8/QS16/F16/F32
+ * @param[in]  input_stride_x                       Stride of the first source tensor in X dimension (in bytes)
+ * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  input_stride_y                       Stride of the first source tensor in Y dimension (in bytes)
+ * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  input_stride_z                       Stride of the first source tensor in Z dimension (in bytes)
+ * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the first source tensor
+ * @param[out] output_ptr                           Pointer to the destination tensor. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      Stride of the destination tensor in X dimension (in bytes)
+ * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
+ * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination tensor
+ * @param[in]  mean_ptr                             Pointer to the mean source tensor. Supported data types: same as @p input_ptr
+ * @param[in]  mean_stride_x                        Stride of the mean source tensor in X dimension (in bytes)
+ * @param[in]  mean_step_x                          mean_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  mean_offset_first_element_in_bytes   The offset of the first element in the mean source tensor
+ * @param[in]  var_ptr                              Pointer to the var tensor. Supported data types: same as @p input_ptr
+ * @param[in]  var_stride_x                         Stride of the var tensor in X dimension (in bytes)
+ * @param[in]  var_step_x                           var_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  var_offset_first_element_in_bytes    The offset of the first element in the var source tensor
+ * @param[in]  beta_ptr                             Pointer to the beta source tensor. Supported data types: same as @p input_ptr
+ * @param[in]  beta_stride_x                        Stride of the beta source tensor in X dimension (in bytes)
+ * @param[in]  beta_step_x                          beta_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  beta_offset_first_element_in_bytes   The offset of the first element in the beta source tensor
+ * @param[in]  gamma_ptr                            Pointer to the gamma source tensor. Supported data types: same as @p input_ptr
+ * @param[in]  gamma_stride_x                       Stride of the gamma source tensor in X dimension (in bytes)
+ * @param[in]  gamma_step_x                         gamma_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  gamma_offset_first_element_in_bytes  The offset of the first element in the gamma source tensor
+ * @param[in]  epsilon                              Epsilon parameter in the batch normalization equation
+ */
+__kernel void batchnormalization_layer_nhwc(TENSOR3D_DECLARATION(input),
+#ifndef IN_PLACE
+                                            TENSOR3D_DECLARATION(output),
+#endif /* not IN_PLACE */
+                                            VECTOR_DECLARATION(mean),
+                                            VECTOR_DECLARATION(var),
+#ifndef USE_DEFAULT_BETA
+                                            VECTOR_DECLARATION(beta),
+#endif /* USE_DEFAULT_BETA */
+#ifndef USE_DEFAULT_GAMMA
+                                            VECTOR_DECLARATION(gamma),
+#endif /* USE_DEFAULT_GAMMA */
+                                            float epsilon)
+{
+    Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input);
+#ifdef IN_PLACE
+    Tensor3D out = in;
+#else  /* IN_PLACE */
+    Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
+#endif /* IN_PLACE */
+    Vector mean = CONVERT_TO_VECTOR_STRUCT(mean);
+    Vector var  = CONVERT_TO_VECTOR_STRUCT(var);
+#ifndef USE_DEFAULT_BETA
+    Vector beta = CONVERT_TO_VECTOR_STRUCT(beta);
+#endif /* USE_DEFAULT_BETA */
+#ifndef USE_DEFAULT_GAMMA
+    Vector gamma = CONVERT_TO_VECTOR_STRUCT(gamma);
+#endif /* USE_DEFAULT_GAMMA */
+
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    data = 0;
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    denominator = 0;
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    numerator = 0;
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    x_bar = 0;
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    res = 0;
+
+    const int current_slice = get_global_id(0);
+
+    data        = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr);
+    denominator = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(var.ptr + current_slice * VEC_SIZE * var.stride_x));
+    denominator = INVSQRT_OP(ADD_OP(denominator, ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(epsilon))));
+
+    // Calculate x bar and store results
+    numerator = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(mean.ptr + current_slice * VEC_SIZE * mean.stride_x));
+    numerator = SUB_OP(data, numerator);
+    x_bar     = MUL_OP(numerator, denominator);
+
+#ifndef USE_DEFAULT_GAMMA
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    gamma_vec = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(gamma.ptr + current_slice * VEC_SIZE * gamma.stride_x));
+
+    res = MUL_OP(gamma_vec, x_bar);
+#else  /* USE_DEFAULT_GAMMA */
+    // gamma is equal to 1, no need to perform multiplications
+    res = x_bar;
+#endif /* USE_DEFAULT_GAMMA */
+
+#ifndef USE_DEFAULT_BETA
+    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
+    beta_vec = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(beta.ptr + current_slice * VEC_SIZE * beta.stride_x));
+    // beta is not zero, hence we need to perform the addition
+    res = ADD_OP(res, beta_vec);
+#endif /* USE_DEFAULT_BETA */
+
+    res = ACTIVATION_FUNC(res);
+
+    VSTORE(VEC_SIZE)
+    (res, 0, (__global DATA_TYPE *)out.ptr);
+}
 #endif /* defined(VEC_SIZE) && defined(DATA_TYPE) */
diff --git a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
index 62f21eed96..3a2211c522 100644
--- a/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLBatchNormalizationLayerKernel.cpp
@@ -49,6 +49,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output,
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, var);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, mean, var);
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, mean, var);
+    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::CHANNEL)) != mean->dimension(0));
     if(beta != nullptr)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mean, beta);
@@ -62,7 +63,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output,
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, gamma);
     }
 
-    ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) != mean->dimension(0));
     if(act_info.enabled())
     {
         ActivationLayerInfo::ActivationFunction act = act_info.activation();
@@ -75,6 +75,7 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output,
     if(output != nullptr && output->total_size() != 0)
     {
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
+        ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, output);
     }
@@ -152,7 +153,7 @@ void CLBatchNormalizationLayerKernel::configure(ICLTensor *input, ICLTensor *out
     build_opts.add_option_if(gamma == nullptr, "-DUSE_DEFAULT_GAMMA");
 
     // Create kernel
-    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("batchnormalization_layer", build_opts.options()));
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("batchnormalization_layer_" + lower_string(string_from_data_layout(input->info()->data_layout())), build_opts.options()));
 
     // Set kernel static arguments
     unsigned int include_output = (!_run_in_place) ? 1 : 0;
@@ -173,6 +174,8 @@ void CLBatchNormalizationLayerKernel::configure(ICLTensor *input, ICLTensor *out
     ICLKernel::configure(win_config.second);
 
     _config_id = "batch_normalization_layer_";
+    _config_id += string_from_data_layout(input->info()->data_layout());
+    _config_id += "_";
     _config_id += string_from_data_type(input->info()->data_type());
     _config_id += "_";
     _config_id += support::cpp11::to_string(input->info()->dimension(0));
diff --git a/src/core/Utils.cpp b/src/core/Utils.cpp
index 4a237f9daa..b5663e63e4 100644
--- a/src/core/Utils.cpp
+++ b/src/core/Utils.cpp
@@ -126,6 +126,18 @@ const std::string &arm_compute::string_from_channel(Channel channel)
     return channels_map[channel];
 }
 
+const std::string &arm_compute::string_from_data_layout(DataLayout dl)
+{
+    static std::map<DataLayout, const std::string> dl_map =
+    {
+        { DataLayout::UNKNOWN, "UNKNOWN" },
+        { DataLayout::NCHW, "NCHW" },
+        { DataLayout::NHWC, "NHWC" },
+    };
+
+    return dl_map[dl];
+}
+
 const std::string &arm_compute::string_from_data_type(DataType dt)
 {
     static std::map<DataType, const std::string> dt_map =