Fixed CTS failures CLInstanceNorm

* Resolves COMPMID-4400

Change-Id: I54c33a017c735194fbf4437d1c7df465208bc0ca
Signed-off-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5505
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>

4 files changed, 66 insertions, 51 deletions

diff --git a/src/core/CL/cl_kernels/instance_normalization.cl b/src/core/CL/cl_kernels/instance_normalization.cl
index d2507d94dd..adfbebd67d 100644
--- a/src/core/CL/cl_kernels/instance_normalization.cl
+++ b/src/core/CL/cl_kernels/instance_normalization.cl
@@ -23,7 +23,7 @@
  */
 #include "helpers.h"
 
-#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DIM_X) && defined(DIM_Y) && defined(DIM_Z)
+#if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(INTERNAL_DATA_TYPE) & defined(DIM_X) && defined(DIM_Y) && defined(DIM_Z)
 /** This function computes the mean and variance of each plane of the input tensor and provides it as output.
  *
  * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
@@ -57,32 +57,37 @@ __kernel void compute_mean_var(
     Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(output);
 
 #if defined(NHWC)
-    const int ch             = get_global_id(0); // Current channel
-    const int batch          = get_global_id(1); // Current batch
-    const int elements_plane = DIM_Y * DIM_Z;
-    float     part_sum       = 0.f;
-    float     part_sum_sq    = 0.f;
-    const int in_offset      = input_offset_first_element_in_bytes + batch * input_stride_w + ch * sizeof(DATA_TYPE);
-    for(int i = 0; i < (DIM_Y * DIM_Z); ++i)
+    const int          ch             = get_global_id(0); // Current channel
+    const int          batch          = get_global_id(1); // Current batch
+    const int          elements_plane = DIM_Y * DIM_Z;
+    INTERNAL_DATA_TYPE part_sum       = 0.f;
+    INTERNAL_DATA_TYPE part_sum_sq    = 0.f;
+    const int          in_offset      = input_offset_first_element_in_bytes + batch * input_stride_w + ch * sizeof(DATA_TYPE);
+
+    for(int i_w = 0; i_w < DIM_Y; ++i_w)
     {
-        const float data = *((__global DATA_TYPE *)(input_ptr + in_offset + i * input_stride_y));
-        part_sum += data;
-        part_sum_sq += data * data;
+        for(int i_h = 0; i_h < DIM_Z; ++i_h)
+        {
+            INTERNAL_DATA_TYPE data = (INTERNAL_DATA_TYPE) * ((__global DATA_TYPE *)tensor4D_offset(&in, ch, i_w, i_h, batch));
+            part_sum += data;
+            part_sum_sq += data * data;
+        }
     }
-    float    mean                       = (part_sum / elements_plane);
-    float    var                        = (part_sum_sq / elements_plane) - (mean * mean);
-    __global DATA_TYPE *output_address0 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 0, batch);
-    *output_address0                    = mean;
-    __global DATA_TYPE *output_address1 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 1, batch);
-    *output_address1                    = var;
+
+    INTERNAL_DATA_TYPE mean                      = (part_sum / elements_plane);
+    INTERNAL_DATA_TYPE var                       = (part_sum_sq / elements_plane) - (mean * mean);
+    __global INTERNAL_DATA_TYPE *output_address0 = (__global INTERNAL_DATA_TYPE *)tensor3D_offset(&out, ch, 0, batch);
+    *output_address0                             = mean;
+    __global INTERNAL_DATA_TYPE *output_address1 = (__global INTERNAL_DATA_TYPE *)tensor3D_offset(&out, ch, 1, batch);
+    *output_address1                             = var;
 #else // !defined(NHWC)
     const int ch             = get_global_id(2) % DIM_Z; // Current channel
     const int batch          = get_global_id(2) / DIM_Z; // Current batch
     const int elements_plane = DIM_X * DIM_Y;
 
-    VEC_DATA_TYPE(float, VEC_SIZE)
+    VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE)
     part_sum = 0.f;
-    VEC_DATA_TYPE(float, VEC_SIZE)
+    VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE)
     part_sum_sq = 0.f;
     // Calculate partial sum
     for(int y = 0; y < DIM_Y; ++y)
@@ -91,15 +96,15 @@ __kernel void compute_mean_var(
         for(; x <= (DIM_X - VEC_SIZE); x += VEC_SIZE)
         {
             // Load data
-            VEC_DATA_TYPE(float, VEC_SIZE)
-            data = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)), VEC_DATA_TYPE(float, VEC_SIZE));
+            VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE)
+            data = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)), VEC_DATA_TYPE(INTERNAL_DATA_TYPE, VEC_SIZE));
             part_sum += data;
             part_sum_sq += data * data;
         }
         // Left-overs loop
         for(; x < DIM_X; ++x)
         {
-            float data = (float)(*((__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)));
+            INTERNAL_DATA_TYPE data = (INTERNAL_DATA_TYPE)(*((__global DATA_TYPE *)tensor4D_offset(&in, x, y, ch, batch)));
             part_sum.s0 += data;
             part_sum_sq.s0 += data * data;
         }
@@ -120,16 +125,16 @@ __kernel void compute_mean_var(
     part_sum.s0 += part_sum.s1;
     part_sum_sq.s0 += part_sum_sq.s1;
 
-    float sum    = (float)part_sum.s0;
-    float sum_sq = (float)part_sum_sq.s0;
+    INTERNAL_DATA_TYPE sum    = (INTERNAL_DATA_TYPE)part_sum.s0;
+    INTERNAL_DATA_TYPE sum_sq = (INTERNAL_DATA_TYPE)part_sum_sq.s0;
 
-    const float mean = (sum / elements_plane);
-    const float var  = (sum_sq / elements_plane) - (mean * mean);
+    const INTERNAL_DATA_TYPE mean = (sum / elements_plane);
+    const INTERNAL_DATA_TYPE var  = (sum_sq / elements_plane) - (mean * mean);
 
-    __global DATA_TYPE *output_address0 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 0, batch);
-    *output_address0                    = mean;
-    __global DATA_TYPE *output_address1 = (__global DATA_TYPE *)tensor3D_offset(&out, ch, 1, batch);
-    *output_address1                    = var;
+    __global INTERNAL_DATA_TYPE *output_address0 = (__global INTERNAL_DATA_TYPE *)tensor3D_offset(&out, ch, 0, batch);
+    *output_address0                             = mean;
+    __global INTERNAL_DATA_TYPE *output_address1 = (__global INTERNAL_DATA_TYPE *)tensor3D_offset(&out, ch, 1, batch);
+    *output_address1                             = var;
 
 #endif // defined(NHWC)
 }
@@ -185,12 +190,12 @@ __kernel void instance_normalization(
     const int batch = get_global_id(2) / DIM_Z; // Current batch
 #endif                                  /* defined(NHWC) */
 
-    const __global DATA_TYPE *mean_ptr                   = (__global DATA_TYPE *)tensor3D_offset(&mean_var, ch, 0, batch);
-    const __global DATA_TYPE *var_ptr                    = (__global DATA_TYPE *)tensor3D_offset(&mean_var, ch, 1, batch);
-    const INTERNAL_DATA_TYPE                      mean   = (INTERNAL_DATA_TYPE) * mean_ptr;
-    const INTERNAL_DATA_TYPE                      var    = (INTERNAL_DATA_TYPE) * var_ptr;
-    const INTERNAL_DATA_TYPE                      multip = GAMMA / sqrt(var + EPSILON);
-    const INTERNAL_DATA_TYPE                      beta   = (INTERNAL_DATA_TYPE)BETA;
+    const __global INTERNAL_DATA_TYPE *mean_ptr                   = (__global INTERNAL_DATA_TYPE *)tensor3D_offset(&mean_var, ch, 0, batch);
+    const __global INTERNAL_DATA_TYPE *var_ptr                    = (__global INTERNAL_DATA_TYPE *)tensor3D_offset(&mean_var, ch, 1, batch);
+    const INTERNAL_DATA_TYPE                               mean   = (INTERNAL_DATA_TYPE) * mean_ptr;
+    const INTERNAL_DATA_TYPE                               var    = (INTERNAL_DATA_TYPE) * var_ptr;
+    const INTERNAL_DATA_TYPE                               multip = GAMMA / sqrt(var + EPSILON);
+    const INTERNAL_DATA_TYPE                               beta   = (INTERNAL_DATA_TYPE)BETA;
 
 #if defined(NHWC)
     const int in_offset = input_offset_first_element_in_bytes + batch * input_stride_w + ch * sizeof(DATA_TYPE);
@@ -198,17 +203,19 @@ __kernel void instance_normalization(
     const int out_offset = output_offset_first_element_in_bytes + batch * input_stride_w + ch * sizeof(DATA_TYPE);
 #endif /* IN_PLACE */
 
-    for(int i = 0; i < (DIM_Y * DIM_Z); ++i)
+    for(int i_w = 0; i_w < DIM_Y; ++i_w)
     {
-        __global DATA_TYPE *input_address = (__global DATA_TYPE *)(input_ptr + in_offset + i * input_stride_y);
+        for(int i_h = 0; i_h < DIM_Z; ++i_h)
+        {
+            __global DATA_TYPE *input_address = (__global DATA_TYPE *)tensor4D_offset(&in, ch, i_w, i_h, batch);
 #ifdef IN_PLACE
-        __global DATA_TYPE *output_address = input_address;
+            __global DATA_TYPE *output_address = input_address;
 #else  /* !IN_PLACE */
-        __global DATA_TYPE *output_address = (__global DATA_TYPE *)(output_ptr + out_offset + i * output_stride_y);
+            __global DATA_TYPE *output_address = (__global DATA_TYPE *)tensor4D_offset(&out, ch, i_w, i_h, batch);
 #endif /* IN_PLACE */
-        *(output_address) = (*(input_address) - mean) * multip + beta;
+            *(output_address) = (*(input_address) - mean) * multip + (INTERNAL_DATA_TYPE)BETA;
+        }
     }
-
 #else // !defined(NHWC)
     for(int y = 0; y < DIM_Y; ++y)
     {
diff --git a/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp b/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp
index 80a42cc3f5..dcde0850a7 100644
--- a/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.cpp
@@ -74,7 +74,7 @@ CLComputeMeanVariance::CLComputeMeanVariance()
 {
 }
 
-void CLComputeMeanVariance::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output)
+void CLComputeMeanVariance::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output, bool use_mixed_precision)
 {
     ARM_COMPUTE_ERROR_ON_NULLPTR(input);
     auto padding_info = get_padding_info({ input, output });
@@ -86,6 +86,7 @@ void CLComputeMeanVariance::configure(const CLCompileContext &compile_context, I
     const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
 
     CLBuildOptions build_opts;
+    build_opts.add_option("-DINTERNAL_DATA_TYPE=" + (use_mixed_precision ? "float" : get_cl_type_from_data_type(input->info()->data_type())));
     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
     build_opts.add_option("-DDIM_X=" + support::cpp11::to_string(input->info()->dimension(0)));
@@ -105,8 +106,14 @@ void CLComputeMeanVariance::configure(const CLCompileContext &compile_context, I
     const TensorShape  out_shape(input_channel, 2u, input_batches);
 
     // Output auto initialization if not yet initialized
-    auto_init_if_empty(*output->info(), out_shape, 1, input->info()->data_type());
-
+    if(use_mixed_precision)
+    {
+        auto_init_if_empty(*output->info(), out_shape, 1, DataType::F32);
+    }
+    else
+    {
+        auto_init_if_empty(*output->info(), out_shape, 1, input->info()->data_type());
+    }
     ICLKernel::configure_internal(win);
     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
 }
diff --git a/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.h b/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.h
index 33a3ff97c3..2f9014a651 100644
--- a/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.h
+++ b/src/core/CL/kernels/CLInstanceNormalizationLayerKernel.h
@@ -100,12 +100,13 @@ public:
 
     /** Set the input and output tensors.
      *
-     * @param[in]      compile_context The compile context to be used.
-     * @param[in, out] input           Source tensor. Data types supported: F16/F32. Data layout supported: NCHW, NHWC
-     *                                 In case of @p output tensor = nullptr this tensor will store the result of the normalization.
-     * @param[out]     output          Destination tensor. Data types and data layouts supported: same as @p input.
+     * @param[in]      compile_context     The compile context to be used.
+     * @param[in, out] input               Source tensor. Data types supported: F16/F32. Data layout supported: NCHW, NHWC
+     *                                     In case of @p output tensor = nullptr this tensor will store the result of the normalization.
+     * @param[out]     output              Destination tensor. Data types and data layouts supported: same as @p input.
+     * @param[in]      use_mixed_precision Use mixed precision in case of FP16 execution
      */
-    void configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output);
+    void configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output, bool use_mixed_precision);
 
     /** Static function to check if given info will lead to a valid configuration of @ref CLInstanceNormalizationLayer.
      *
diff --git a/src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp b/src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp
index f2406d68f4..4a0bda8255 100644
--- a/src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp
+++ b/src/runtime/CL/functions/CLInstanceNormalizationLayer.cpp
@@ -52,7 +52,7 @@ void CLInstanceNormalizationLayer::configure(ICLTensor *input, ICLTensor *output
 void CLInstanceNormalizationLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, ICLTensor *output, float gamma, float beta, float epsilon, bool use_mixed_precision)
 {
     auto w = std::make_unique<CLComputeMeanVariance>();
-    w->configure(compile_context, input, &_mean_var_tensor);
+    w->configure(compile_context, input, &_mean_var_tensor, use_mixed_precision);
     _mean_var_kernel = std::move(w);
     auto k           = std::make_unique<CLInstanceNormalizationLayerKernel>();
     k->configure(compile_context, input, &_mean_var_tensor, output, InstanceNormalizationLayerKernelInfo(gamma, beta, epsilon, use_mixed_precision));