COMPMID-1710: Fix CL logistic activation for QASYMM8

Logistic activation in QASYMM8 is performed in floating point and then
converting back to QASYMM8. However, if input and output have different
quantization information, a double conversion is done leading to loss in
accuracy.

This patch creates a special case kernel for logistic activation.

Change-Id: Ia18ee0b2f84701674f88785bcd13753b50d64a08
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/696
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 99 insertions, 37 deletions

diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp
index ce846d1dc5..a7d371dabc 100644
--- a/src/core/CL/CLKernelLibrary.cpp
+++ b/src/core/CL/CLKernelLibrary.cpp
@@ -149,6 +149,7 @@ const std::map<std::string, std::string> CLKernelLibrary::_kernel_program_map =
     { "accumulate_weighted", "accumulate.cl" },
     { "activation_layer", "activation_layer.cl" },
     { "activation_layer_qa8", "activation_layer_qa8.cl" },
+    { "activation_layer_logistic_qa8", "activation_layer_qa8.cl" },
     { "batch_to_space_nchw", "batch_to_space.cl" },
     { "batch_to_space_static_nchw", "batch_to_space.cl" },
     { "batch_to_space_nhwc", "batch_to_space.cl" },
diff --git a/src/core/CL/cl_kernels/activation_layer_qa8.cl b/src/core/CL/cl_kernels/activation_layer_qa8.cl
index 8f6a807613..cfb61376ca 100644
--- a/src/core/CL/cl_kernels/activation_layer_qa8.cl
+++ b/src/core/CL/cl_kernels/activation_layer_qa8.cl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,16 +26,6 @@
 #define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
 #define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE)
 
-// Logistic Activation
-inline TYPE logistic_op(TYPE x)
-{
-    VEC_FLOAT x_flt = CONVERT(x, VEC_FLOAT);
-    x_flt           = round(x_flt - (float)O1_VAL) * ((float)S1_VAL);
-    x_flt           = 1.f / (1.f + exp(-x_flt));
-
-    const TYPE x_u8 = CONVERT_SAT(round(x_flt / ((float)S1_VAL)) + (float)O1_VAL, TYPE);
-    return x_u8;
-}
 // RELU Activation
 inline TYPE relu_op(TYPE x)
 {
@@ -95,14 +85,14 @@ inline TYPE lu_brelu_op(TYPE x)
  * @param[in]  input_stride_z                       Stride of the 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 source image
- * @param[out] output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
- * @param[in]  output_stride_x                      Stride of the destination image 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 image 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 source 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 image
+ * @param[out] output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]  output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
  */
 __kernel void activation_layer_qa8(
     TENSOR3D_DECLARATION(input)
@@ -131,3 +121,69 @@ __kernel void activation_layer_qa8(
 }
 
 #endif /* defined(ACT) */
+
+#if defined(O2_VAL) && defined(S2_VAL)
+#define OFFSET_OUT O2_VAL
+#define SCALE_OUT S2_VAL
+#else // defined(O2_VAL) && defined(S2_VAL)
+#define OFFSET_OUT O1_VAL
+#define SCALE_OUT S1_VAL
+#endif // defined(O2_VAL) && defined(S2_VAL)
+
+/** This performs a Logistic activation function on QASYMM8 inputs.
+ *
+ * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time
+ *
+ * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
+ * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
+ * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively.
+ * @note Quantization scales of the input/output tensors are passed in with -DS1_VAL= and -DS2_VAL= respectively.
+ * @note Quantization offsets of the input/output tensors are passed in with -DO1_VAL= and -DO2_VAL= respectively.
+ * @note Quantized value of constant zero should be given as a preprocessor argument using -DCONST_0=value. e.g. -DCONST_0=128.
+ *
+ * @param[in]  input_ptr                            Pointer to the source image. Supported data types: QASYMM8
+ * @param[in]  input_stride_x                       Stride of the source image 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 source image 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 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 source image
+ * @param[out] output_ptr                           (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr
+ * @param[in]  output_stride_x                      (Optional) Stride of the destination image in X dimension (in bytes)
+ * @param[in]  output_step_x                        (Optional) output_stride_x * number of elements along X processed per workitem(in bytes)
+ * @param[in]  output_stride_y                      (Optional) Stride of the destination image in Y dimension (in bytes)
+ * @param[in]  output_step_y                        (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes)
+ * @param[in]  output_stride_z                      (Optional) Stride of the source tensor in Z dimension (in bytes)
+ * @param[in]  output_step_z                        (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes)
+ * @param[in]  output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image
+ */
+__kernel void activation_layer_logistic_qa8(
+    TENSOR3D_DECLARATION(input)
+#ifndef IN_PLACE
+    ,
+    TENSOR3D_DECLARATION(output)
+#endif /* not IN_PLACE */
+)
+{
+    // Get pixels pointer
+    Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input);
+#ifdef IN_PLACE
+    Tensor3D output = input;
+#else  /* IN_PLACE */
+    Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output);
+#endif /* IN_PLACE */
+
+    // Load data
+    TYPE data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr);
+
+    VEC_FLOAT data_flt = CONVERT(data, VEC_FLOAT);
+    data_flt           = round(data_flt - (float)O1_VAL) * ((float)S1_VAL);
+    data_flt           = 1.f / (1.f + exp(-data_flt));
+
+    data = CONVERT_SAT(round(data_flt / ((float)SCALE_OUT)) + (float)OFFSET_OUT, TYPE);
+
+    // Store result
+    VSTORE(VEC_SIZE)
+    (data, 0, (__global DATA_TYPE *)output.ptr);
+}
diff --git a/src/core/CL/kernels/CLActivationLayerKernel.cpp b/src/core/CL/kernels/CLActivationLayerKernel.cpp
index 73a4d7d2c6..100184d2f3 100644
--- a/src/core/CL/kernels/CLActivationLayerKernel.cpp
+++ b/src/core/CL/kernels/CLActivationLayerKernel.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2018 ARM Limited.
+ * Copyright (c) 2016-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -129,24 +129,25 @@ void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, Act
         b_const_int = input->info()->quantization_info().quantize(b_const, RoundingPolicy::TO_NEAREST_UP);
     }
 
+    const bool is_logistic_activation_quantized = is_data_type_quantized_asymmetric(dt) && act_info.activation() == ActivationLayerInfo::ActivationFunction::LOGISTIC;
     // Set build options
-    std::set<std::string> build_opts;
-    build_opts.emplace(("-DACT=" + lower_string(string_from_activation_func(act_info.activation()))));
-    build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(dt)));
-    build_opts.emplace(("-DSELECT_DATA_TYPE=" + get_cl_select_type_from_data_type(dt)));
-    build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
+    CLBuildOptions build_opts;
+    build_opts.add_option_if(!is_logistic_activation_quantized, "-DACT=" + lower_string(string_from_activation_func(act_info.activation())));
+    build_opts.add_option(("-DDATA_TYPE=" + get_cl_type_from_data_type(dt)));
+    build_opts.add_option(("-DSELECT_DATA_TYPE=" + get_cl_select_type_from_data_type(dt)));
+    build_opts.add_option(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)));
 
     if(is_data_type_quantized(dt))
     {
-        build_opts.emplace(("-DA_VAL=" + support::cpp11::to_string(a_const_int)));
-        build_opts.emplace(("-DB_VAL=" + support::cpp11::to_string(b_const_int)));
+        build_opts.add_option(("-DA_VAL=" + support::cpp11::to_string(a_const_int)));
+        build_opts.add_option(("-DB_VAL=" + support::cpp11::to_string(b_const_int)));
 
         const int   o1 = input->info()->quantization_info().offset;
         const float s1 = input->info()->quantization_info().scale;
         // Quantized value of 0 corresponds to the offset o1
-        build_opts.emplace(("-DCONST_0=" + support::cpp11::to_string(o1)));
-        build_opts.emplace(("-DS1_VAL=" + float_to_string_with_full_precision(s1)));
-        build_opts.emplace(("-DO1_VAL=" + support::cpp11::to_string(o1)));
+        build_opts.add_option(("-DCONST_0=" + support::cpp11::to_string(o1)));
+        build_opts.add_option(("-DS1_VAL=" + float_to_string_with_full_precision(s1)));
+        build_opts.add_option(("-DO1_VAL=" + support::cpp11::to_string(o1)));
 
         // Set scale and offset of the input and output if they have different quantization info
         if(is_data_type_quantized_asymmetric(dt) && output != nullptr)
@@ -156,22 +157,26 @@ void CLActivationLayerKernel::configure(ICLTensor *input, ICLTensor *output, Act
 
             if(o1 != o2 || s1 != s2)
             {
-                build_opts.emplace(("-DS2_VAL=" + float_to_string_with_full_precision(s2)));
-                build_opts.emplace(("-DO2_VAL=" + support::cpp11::to_string(o2)));
+                build_opts.add_option(("-DS2_VAL=" + float_to_string_with_full_precision(s2)));
+                build_opts.add_option(("-DO2_VAL=" + support::cpp11::to_string(o2)));
             }
         }
     }
     else
     {
-        build_opts.emplace(("-DA_VAL=" + float_to_string_with_full_precision(a_const)));
-        build_opts.emplace(("-DB_VAL=" + float_to_string_with_full_precision(b_const)));
+        build_opts.add_option(("-DA_VAL=" + float_to_string_with_full_precision(a_const)));
+        build_opts.add_option(("-DB_VAL=" + float_to_string_with_full_precision(b_const)));
     }
 
-    build_opts.emplace((_run_in_place) ? "-DIN_PLACE" : "");
+    build_opts.add_option_if(_run_in_place, "-DIN_PLACE");
 
     // Create kernel
-    std::string kernel_name = is_data_type_quantized_asymmetric(dt) ? std::string("activation_layer_qa8") : std::string("activation_layer");
-    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts));
+    std::string kernel_name = std::string("activation_layer");
+    if(is_data_type_quantized_asymmetric(dt))
+    {
+        kernel_name += is_logistic_activation_quantized ? std::string("_logistic_qa8") : std::string("_qa8");
+    }
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Make sure _kernel is initialized before calling the parent's configure
     _input  = input;