COMPMID-811 Add NHWC data format support for CL depthwise convolution

Change-Id: I574f7945f0be009c638d860028bce8b52b4120fd
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/136484
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

1 files changed, 118 insertions, 53 deletions

diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
index d24ef0f496..1de08aa1a2 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
@@ -44,18 +44,27 @@ namespace
 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
                           const ActivationLayerInfo &act_info)
 {
-    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8);
-    ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.enabled()) && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
-                                    && (act_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
-                                    && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU),
-                                    "For QASYMM8 only relu, lower bounded relu and lower-upper bounded relu are supported");
+    ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::QASYMM8);
+    ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.enabled()) && (input->data_type() == DataType::F32 || ((act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
+                                                                                                     && (act_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
+                                                                                                     && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU))),
+                                    "For QASYMM8 only relu, lower bounded relu and lower-upper bounded relu are supported"); //COMPMID-1317 add fused activation for F32
     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
     ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1); // COMPMID-1071 Add depth multiplier support for NHWC
     ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(1) != 3 || weights->dimension(2) != 3);
 
+    const bool is_qasymm = is_data_type_quantized_asymmetric(input->data_type());
+
     if(biases != nullptr)
     {
-        ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+        if(is_qasymm)
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+        }
+        else
+        {
+            ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
+        }
         ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0));
         ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
     }
@@ -72,12 +81,23 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *output,
                                                         const PadStrideInfo &conv_info)
 {
-    const unsigned int num_rows_processed_per_iteration = 4;
-    const unsigned int num_elems_accessed_per_iteration = 4;
+    const bool is_qasymm   = is_data_type_quantized_asymmetric(input->data_type());
+    const bool is_stride_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
+
+    const unsigned int num_rows_processed_per_iteration = is_qasymm ? 4 : (is_stride_1 ? 2 : 1);
+    const unsigned int num_elems_accessed_per_iteration = is_qasymm ? 4 : 2;
     const unsigned int num_rows_read_per_iteration      = num_rows_processed_per_iteration + 2;
-    const unsigned int num_rows_written_per_iteration   = num_rows_processed_per_iteration / conv_info.stride().first;
+    const unsigned int num_rows_written_per_iteration   = std::ceil(num_rows_processed_per_iteration / static_cast<float>(conv_info.stride().first));
 
-    const BorderSize border_size(std::max(conv_info.pad_left(), conv_info.pad_top()), 0, std::max(conv_info.pad_right(), conv_info.pad_bottom()), 0);
+    BorderSize border_size;
+    if(is_qasymm)
+    {
+        border_size = BorderSize(std::max(conv_info.pad_left(), conv_info.pad_top()), 0, std::max(conv_info.pad_right(), conv_info.pad_bottom()), 0);
+    }
+    else
+    {
+        border_size = BorderSize(conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
+    }
 
     // Configure kernel window
     Window win = calculate_max_window(*output, Steps(num_elems_accessed_per_iteration, num_rows_written_per_iteration));
@@ -103,7 +123,7 @@ std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITen
 } // namespace
 
 CLDepthwiseConvolutionLayer3x3NHWCKernel::CLDepthwiseConvolutionLayer3x3NHWCKernel()
-    : _num_rows_processed_per_iteration(1)
+    : _num_rows_processed_per_iteration(1), _num_planes_processed_per_iteration(1)
 {
 }
 
@@ -135,66 +155,97 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input,
     ARM_COMPUTE_ERROR_ON(conv_stride_x < 1 || conv_stride_x > 2);
     ARM_COMPUTE_ERROR_ON(std::max(conv_info.pad_top(), conv_info.pad_bottom()) > 1);
 
-    _input                            = input;
-    _output                           = output;
-    _weights                          = weights;
-    _biases                           = biases;
-    _conv_stride_y                    = conv_info.stride().second;
-    _num_rows_processed_per_iteration = 4;
+    const bool is_qasymm   = is_data_type_quantized_asymmetric(input->info()->data_type());
+    const bool is_stride_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
 
-    const unsigned int num_elems_accessed_per_iteration = 4;
+    _input                              = input;
+    _output                             = output;
+    _weights                            = weights;
+    _biases                             = biases;
+    _conv_stride_y                      = conv_info.stride().second;
+    _num_rows_processed_per_iteration   = is_qasymm ? 4 : (is_stride_1 ? 2 : 1);
+    _num_planes_processed_per_iteration = (is_stride_1 && !is_qasymm) ? 2 : 1;
 
-    _border_size = BorderSize(std::max(conv_info.pad_left(), conv_info.pad_top()), 0, std::max(conv_info.pad_right(), conv_info.pad_bottom()), 0);
+    const unsigned int num_elems_accessed_per_iteration = is_qasymm ? 4 : 2;
 
-    float multiplier        = _input->info()->quantization_info().scale * _weights->info()->quantization_info().scale / _output->info()->quantization_info().scale;
-    int   output_multiplier = 0;
-    int   output_shift      = 0;
-    quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+    if(is_qasymm)
+    {
+        _border_size = BorderSize(std::max(conv_info.pad_left(), conv_info.pad_top()), 0, std::max(conv_info.pad_right(), conv_info.pad_bottom()), 0);
+    }
+    else
+    {
+        _border_size = BorderSize(conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
+    }
 
     CLBuildOptions build_opts;
     build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
-    build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-_input->info()->quantization_info().offset));
-    build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-_weights->info()->quantization_info().offset));
-    build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(_output->info()->quantization_info().offset));
-    build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(9 * input->info()->quantization_info().offset * weights->info()->quantization_info().offset));
-    build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
-    build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_accessed_per_iteration));
-    build_opts.add_option("-DSRC_DIM_1=" + support::cpp11::to_string(_input->info()->dimension(1)));
     build_opts.add_option("-DSRC_DIM_2=" + support::cpp11::to_string(_input->info()->dimension(2)));
     build_opts.add_option("-DCONV_PAD_TOP=" + support::cpp11::to_string(conv_info.pad_top()));
     build_opts.add_option("-DCONV_PAD_LEFT=" + support::cpp11::to_string(conv_info.pad_left()));
 
-    if(act_info.enabled())
+    if(is_qasymm)
     {
-        const int a_val = input->info()->quantization_info().quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
-        const int b_val = input->info()->quantization_info().quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
-        const int o1    = input->info()->quantization_info().offset;
-
-        build_opts.add_option("-DFUSED_ACTIVATION=" + lower_string(string_from_activation_func(act_info.activation())));
-        build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val));
-        build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val));
-        build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1));
-
-        if(output != nullptr)
+        float multiplier        = _input->info()->quantization_info().scale * _weights->info()->quantization_info().scale / _output->info()->quantization_info().scale;
+        int   output_multiplier = 0;
+        int   output_shift      = 0;
+        quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+        build_opts.add_option("-DSRC_DIM_1=" + support::cpp11::to_string(_input->info()->dimension(1)));
+        build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-_input->info()->quantization_info().offset));
+        build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-_weights->info()->quantization_info().offset));
+        build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(_output->info()->quantization_info().offset));
+        build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(9 * input->info()->quantization_info().offset * weights->info()->quantization_info().offset));
+        build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
+        build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
+
+        if(act_info.enabled())
         {
-            const float s1 = input->info()->quantization_info().scale;
-            const float s2 = output->info()->quantization_info().scale;
-            const int   o2 = output->info()->quantization_info().offset;
+            const int a_val = input->info()->quantization_info().quantize(act_info.a(), RoundingPolicy::TO_NEAREST_UP);
+            const int b_val = input->info()->quantization_info().quantize(act_info.b(), RoundingPolicy::TO_NEAREST_UP);
+            const int o1    = input->info()->quantization_info().offset;
+
+            build_opts.add_option("-DFUSED_ACTIVATION=" + lower_string(string_from_activation_func(act_info.activation())));
+            build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val));
+            build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val));
+            build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1));
 
-            if(o1 != o2 || s1 != s2)
+            if(output != nullptr)
             {
-                build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
-                build_opts.add_option("-DS2_VAL=" + float_to_string_with_full_precision(s2));
-                build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
-                build_opts.add_option("-DO2_VAL=" + support::cpp11::to_string(o2));
+                const float s1 = input->info()->quantization_info().scale;
+                const float s2 = output->info()->quantization_info().scale;
+                const int   o2 = output->info()->quantization_info().offset;
+
+                if(o1 != o2 || s1 != s2)
+                {
+                    build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
+                    build_opts.add_option("-DS2_VAL=" + float_to_string_with_full_precision(s2));
+                    build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
+                    build_opts.add_option("-DO2_VAL=" + support::cpp11::to_string(o2));
+                }
             }
         }
     }
+    else if(is_stride_1)
+    {
+        build_opts.add_option("-DNUM_ROWS_PROCESSED=" + support::cpp11::to_string(_num_rows_processed_per_iteration));
+        build_opts.add_option("-DNUM_PLANES_PROCESSED=" + support::cpp11::to_string(_num_planes_processed_per_iteration));
+        build_opts.add_option("-DDST_DIM_2=" + support::cpp11::to_string(_output->info()->dimension(2)));
+    }
+    else
+    {
+        build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(conv_stride_x));
+        build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(_conv_stride_y));
+    }
 
     // Create kernel
-    std::string kernel_name = std::string("depthwise_convolution_3x3_quantized_nhwc_stride") + support::cpp11::to_string(conv_stride_x);
-    _kernel                 = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
+    std::string kernel_name = std::string("depthwise_convolution_3x3") + (is_qasymm ? std::string("_quantized") : std::string()) + std::string("_nhwc");
+    if(is_qasymm || is_stride_1)
+    {
+        kernel_name += std::string("_stride") + support::cpp11::to_string(conv_stride_x);
+    }
+
+    _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
 
     // Configure kernel window
     auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info);
@@ -213,6 +264,8 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input,
     _config_id += support::cpp11::to_string(output->info()->dimension(0));
     _config_id += "_";
     _config_id += support::cpp11::to_string(output->info()->dimension(1));
+    _config_id += "_";
+    _config_id += string_from_data_type(input->info()->data_type());
 }
 
 Status CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
@@ -233,15 +286,18 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::run(const Window &window, cl::Com
     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
 
+    Window win = window;
+    win.set(Window::DimZ, Window::Dimension(0, std::ceil(_output->info()->dimension(2) / static_cast<float>(_num_planes_processed_per_iteration)), 1));
+
     // Create input window and adjust
-    Window win_in = window;
+    Window win_in = win;
     win_in.set_dimension_step(Window::DimY, _num_rows_processed_per_iteration);
     win_in.set_dimension_step(Window::DimZ, _conv_stride_y);
 
     ARM_COMPUTE_ERROR_ON((win_in.y().step() < window.y().step()) || (win_in.z().step() < window.z().step()));
 
     Window slice_in  = win_in.first_slice_window_3D();
-    Window slice_out = window.first_slice_window_3D();
+    Window slice_out = win.first_slice_window_3D();
 
     if(_biases != nullptr)
     {
@@ -252,6 +308,15 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::run(const Window &window, cl::Com
         add_1D_tensor_argument(idx, _biases, win_biases);
     }
 
+    if(!(is_data_type_quantized_asymmetric(_input->info()->data_type())))
+    {
+        unsigned int idx        = 3 * num_arguments_per_3D_tensor() + ((_biases != nullptr) ? num_arguments_per_1D_tensor() : 0);
+        const int    max_offset = _input->info()->strides_in_bytes().z() * _input->info()->dimension(2) - (_input->info()->padding().bottom + _input->info()->padding().top) *
+                                  _input->info()->strides_in_bytes().y();
+
+        _kernel.setArg(idx, max_offset);
+    }
+
     do
     {
         unsigned int idx = 0;